To New Beginnings: A Survey of Unified Perception in Autonomous Vehicle Software

TL;DR

This survey reviews unified perception in autonomous vehicles, which integrates detection, tracking, and prediction tasks into a shared architecture to improve robustness, efficiency, and interpretability.

Contribution

It introduces a comprehensive taxonomy and framework for unified perception, categorizing methods and guiding future research in the field.

Findings

Defines three paradigms: Early, Late, and Full Unified Perception.

Provides systematic review of existing methods, architectures, and datasets.

Highlights open challenges and future research directions.

Abstract

Autonomous vehicle perception typically relies on modular pipelines that decompose the task into detection, tracking, and prediction. While interpretable, these pipelines suffer from error accumulation and limited inter-task synergy. Unified perception has emerged as a promising paradigm that integrates these sub-tasks within a shared architecture, potentially improving robustness, contextual reasoning, and efficiency while retaining interpretable outputs. In this survey, we provide a comprehensive overview of unified perception, introducing a holistic and systemic taxonomy that categorizes methods along task integration, tracking formulation, and representation flow. We define three paradigms -Early, Late, and Full Unified Perception- and systematically review existing methods, their architectures, training strategies, datasets used, and open-source availability, while highlighting future research directions. This work establishes the first comprehensive framework for understanding and advancing unified perception, consolidates fragmented efforts, and guides future research toward more robust, generalizable, and interpretable perception.

Tables

Table 1. TABLE I: Survey overview of modular, end-to-end, and unified perception approaches

Type	Perception
	Detection	Tracking	Prediction
Modular AV Software	[8], [6], [9], [10], [11], [12], [13]	[19], [14], [15], [18], [16], [17]	[20], [7], [21], [23]
Early-Unified Perception	Ours
Late-Unified Perception		Ours
Full-Unified Perception	[33], Ours
End-to-End AV Stacks	[29], [28], [30], [31], [32]

Table 2. TABLE II: Early-Unified Perception Approaches: Overview

Publ.	Year	Modality11footnotemark: 1	Input22footnotemark: 2	Paradigm33footnotemark: 3	Output44footnotemark: 4	Architecture	Learning Strategy55footnotemark: 5	Dataset66footnotemark: 6	Code
CenterTrack[35]	2020	I	2	JDTP - PNA	2/3D BB	CenterNet-backbone[36] + Offset Prediction	Emp. W.L.	kitti, nuScenes, MOT17	✓
TubeTK[37]	2020	I	8	JDTP - INA	2D BB	3D-ResNet[38] + 3D-FPN + CNN Heads	Unb. W.L	MOT15/16/17	✓
PermaTrack[39]	2021	I	1+M	JDTP - INA	2D BB	CenterTrack[35] + Spatio-temporal Recurrent Memory Module	Emp. W.L.	kitti, MOT17	✓
P3AFormer[40]	2022	I	2	JDTP - INA	2D BB	ResNet-50[41] + Swin-Transformer E. + DETR[42] D. + FlowNet[43]	Emp. W.L.	kitti, MOT17/20	✓
QD3DT[44]	2023	I	1+M	JDTP - PNA	3D BB	Faster-RCNN-backbone[45] + Quasi-dense Similarity Learning	Emp. W.L.	kitti, nuScenes, WOD	✓
STAR-track[46]	2024	MVI	1+M.	JDTP/Q - INA	3D BB	DETR3D[47] + Latent Motion Model + Learnable Track Embedding	-	nuScenes	✗
CenterPoint[48]	2021	L	1+M	JDTP - PNA	3D BB	VoxelNet[49]/ PointPillars-backbone[50] + CNN + Association	-	nuScenes, WOD	✓
SimTrack[51]	2021	L	2	JDTP - INA	3D BB	VoxelNet[49]/ PointPillars-backbone[50] + Hybrid-Time Center/ Movement Branch	Emp. W.L.	nuScenes, WOD	✓
CenterTube[52]	2023	L	5	JDTP - INA	3D BB	VoxelNet[49]/ PointPillars-backbone[50] + Center/ Regression/ Movement Branch	Emp. W.L.	nuScenes, kitti	✗
VoxelNeXt[53]	2023	L	1+M	JDTP - INA	3D BB	3D Sparse CNN Backbone + Heads + Query Voxel Association	-	nuScenes, WOD, Argoverse2	✓
MMF-JDT[54]	2024	I+L	1+M	JDTP - INA	3D BB	2/3D RPN + 2/3D Trajectory Regression and Refinement + Fusion	-	kitti, WOD	✓
JDE[55]	2020	I	1+M	JDTE - PNA	2D BB	DarkNet-53[56] + Prediction Heads	Unc. W.L.	MOT16	✓
RetinaTrack[57]	2020	I	1+M	JDTE - PNA	2D BB	FPN + Task-shared Post-FPN Layers	Unb. W.L.	WOD, MOT17	✓
Siamese Track-RCNN[58]	2020	I	1+M	JDTE - INA	2D BB	ResNet-101[41] + FPN + Deformable Convolution	Unb. W.L.	MOT16/17, JTA	✗
CTracker[59]	2020	I	2	JDTE - INA	2D BB	ResNet-50[41] + FPN + Joint Attention	Emp. W.L.	MOT16/17	✓
DEFT[60]	2021	I	1+M	JDTE - PNA	3D BB	CenterNet-backbone[36] + Embedding Extractor	Unc. W.L.	kitti, nuScenes, MOT17	✓
FairMOT[61]	2021	I	1+M	JDTE - PNA	2D BB	CenterNet-backbone[36] + Re-ID Branch	Unc. W.L.	MOT15/16/17/20	✓
GSDT[62]	2021	I	2	JDTE - PNA	2D BB	CenterNet-backbone[36] + GNN	Unc. W.L.	MOT15/16/17/20	✓
CorrTracker[63]	2021	I	4	JDTE - PNA	2D BB	CenterNet-backbone[36] + Correlation Learning in Pyramid	Unb. W.L	MOT15/16/17/20	✗
AlignPS[64]	2021	I	1+M	JDTE - PNA	2D BB	FCOS[65] + Feature Aggregation + Re-ID Embedding	Unb. W.L.	CUHK-SYSU, PRW	✓
Time3D[66]	2022	I	1+M	JDTE - PNA	2/3D BB	KM3D[67] + Attention-based Re-ID Embedding	Unb. W.L	nuScenes	✗
TraDeS[68]	2021	I	3	JDTE/P - INA	2/3D BB	CenterNet-backbone[36]+ Re-ID Embedding + Motion Propagation	Unb. W.L.	nuScenes, MOT16/17, MOTS, Youtube-VIS	✓
Minkowski Tracker[69]	2022	L	3	JDTE/P - PNA	3D BB	SECOND-based[70] +patio-temporal R-CNN	Emp. W.L.	nuScenes	✗
JMODT[71]	2021	I+L	1+M	JDTE/P - PNA	3D BB	RPN + Correlation Backbone + Association	Emp. W.L.	kitti	✓
AlphaTrack[72]	2021	MVI+L	1+M	JDTE/P - PNA	3D BB	CenterPoint-based [48] Branch + Embedding + Affinity	Emp. W.L.	nuScenes	✗
3D DetecTrack[73]	2022	I+L	2	JDTE - PNA	3D BB	3D-CVF-backbone[74] + Spatio-temporal Gated GNN	Unb. W.L.	kitti, nuScenes	✗
TransTrack[75]	2020	I	1+M	JDTQ - PNA	2D BB	DETR-backbone[42] + Transformer E.-D. + Matching	Emp. W.L.	MOT17/20	✓
TrackFormer[76]	2022	I	1+M	JDTQ - INA	2D BB	DETR-backbone[42] + Transformer E.-D.	Emp. W.L.	MOT17, MOTS20	✓
TransCenter[77]	2022	I	2	JDTQ - INA	2D BB	DETR[78]/ PVT[79] E. + Query Learning Network + Transformer D.	Emp. W.L.	kitti, MOT17/20	✓
MOTR[80]	2022	I	1+M	JDTQ - INA	2D BB	DETR-backbone[42] + Transformer E.-D.	Emp. W.L.	MOT17/20	✓
MeMOT[34]	2022	I	1+M	JDTQ/E - INA	2D BB	DETR-backbone[42] + Transformer Memory Bank	Emp. W.L.	MOT16/17/20	✗
MOTRv2[81]	2023	I	1+M	JDTQ - INA	2D BB	YOLOX[82] proposals + MOTR[80]	Emp. W.L.	MOT17/20, DanceTrack	✓
MOTRv3[83]	2023	I	1+M	JDTQ - INA	2D BB	YOLOX[82] proposals + MOTR[80] + Release-Fetch-Supervision	Emp. W.L.	MOT17, DanceTrack	✗
Co-MOT[84]	2023	I	1+M	JDTQ - INA	2D BB	ResNet-50[41] + Deformable Transformer E. + Stacked Deformable D.	Unb. W.L	MOT17, DanceTrack, BDD100K	✓
MeMOTR[85]	2023	I	1+M	JDTQ - INA	2D BB	ResNet-50[41] + Transformer E.-D. + Temporal Interaction Module	Emp. W.L.	MOT17, DanceTrack, BDD100K	✓
LAID[86]	2024	I	1+M	JDTQ - INA	2D BB	YOLOX-backbone[82] + Learnable Associator	Emp. W.L.	DanceTrack, SportsMOT	✗
SambaMOTR[87]	2024	I	1+M	JDTQ - INA	2D BB	MOTR[80] E.-D. + Set-of-sequences Model	Emp. W.L.	MOT17, DanceTrack, SportsMOT, BFT	✓
MUTR3D[88]	2022	MVI	1+M	JDTQ - INA	3D BB	DETR-inspired[78] Multi-camera Transformer E.-D.	Emp. W.L.	nuScenes	✓
DQTrack[89]	2023	MVI	1+M	JDTQ - INA	3D BB	DETR3D[47] E. + Transformer-based Learnable Association	Emp. W.L.	nuScenes	✓
Ada-track[90]	2024	MVI	3	JDTQ - INA	3D BB	DETR3D-backbone[47] + Transformer D. + Learnable Association	Emp. W.L.	nuScenes	✓
OneTrack[91]	2024	MVI	1+M	JDTQ - INA	3D BB	StreamPetr-E.[92] + Transformer D. + Masked Attention	Emp. W.L.	nuScenes	✗
MCTR[93]	2025	MVI	1+M	JDTQ - INA	2D BB	DETR[42] Embeddings + Transformer D. + Association Module	Unb. W.L.	MMPTrack	✗
JDT3D[94]	2024	L	1+M	JDTQ/P - INA	3D BB	VoxelNet[49] + Transformer D. + Past/Future Reasoning	Emp. W.L.	nuScenes	✓
Mask4Former[95]	2024	L	2	JDTQ - INA	3D Seg.	DETR-inspired[78] + Mask Module + Query Refinement	Emp. W.L.	kitti(semanticKITTI)	✓
MotionTrack[96]	2023	MVI+L	1+M	JDTQ - INA	3D BB	TransFusion[97] + Transformer-based Data Association	-	nuScenes	✗
6 Dataset: (kitti:[25], nuScenes:[26], WOD:[27], Argoverse2:[98], MOT15:[99], MOT16:[100], MOT17:[100], MOT20:[101], MOTS:[102], CUHK-SYSU:[103], PRW:[104], DanceTrack:[105], JTA:[106], BDD100K:[107], SportsMOT:[108], BFT:[109], Youtube-VIS:[110], MMPTrack:[111])
5 Learning Strategy: (Emp. W.L.: Empirical Weighted Learning, Unb. W.L.: Unbalanced Weighted Learning, Unc. W.L.: Uncertainty Weighted Learning [112], -: No Information)
4 Output: Dimension (2D/ 3D), BB: Bounding Box, Seg.: Segmentation
3 Paradigm: ([JDTP: Joint Detection and Tracking by Propagation, JDTE: Joint Detection and Tracking by Embedding, JDTQ: Joint Detection and Tracking by Query], [PNA: Post-Network Association, INA: In-Netowrk Association])
2 Input: ([Number of Previous Frames Input, M: Use of Memory])
1 Modality: (I: Image, MVI: Multi-View Image, L: LiDAR)

Table 3. TABLE III: Late-Unified Perception Approaches: Overview

Publ.	Year	Modality11footnotemark: 1	Input22footnotemark: 2	Paradigm33footnotemark: 3	Output44footnotemark: 4	Architecture	Learning Strategy55footnotemark: 5	Dataset66footnotemark: 6	Code
PTP [116]	2021	A	BB, (1+M, -)	M, Imp. T	T(3, V,P,C)	GNN + Diversity Sampling + CVAE [117]	Emp. W.L.	kitti, nuScenes	✗
MTP[118]	2022	A	BB, (4, -)	M, Imp. T	T(6, V,P,C)	Multi-Hypothesis Association + Predictor + Sampling	-	kitti, nuScenes	✗
Affinipred [114]	2022	A	BB, (4, Map)	M, Imp. T	T(6, V,P,C)	Affinity-based Transformer + CVAE	Emp. W.L.	kitti, nuScenes	✗
TTFD [119]	2023	A	BB, (4, -)	M, Imp. T	T(3, V)	GRU-based Affinity-Motion E.-D.	Emp. W.L.	nuScenes, Argoverse	✗
PF-Track[120]	2023	MVI	LR, (1+M, -)	A + M, Exp. T	T(4, V,P,C)	Attention-based Past + Future Reasoning	Emp. W.L.	nuScenes	✓
StreamMOTP [121]	2024	A	BB, (4, -)	M, Exp. T	T(6, V,P,C)	Spatio-temporal E. + Interaction + Dual-stream Predictor	Emp. W.L.	nuScenes	✗
Uhlemann et al. [122]	2025	A	BB, (20, Map)	M, Imp. T	O(6, P)	Transformer E.[123]-D.[124]	Emp. W.L.	Argoverse2	✗
6 Dataset: (kitti:[25], nuScenes:[26], Argoverse: [125] , Argoverse2:[98])
5 Learning Strategy: (Emp. W.L.: Empirical Weighted Learning, Unb. W.L.: Unbalanced Weighted Learning, Unc. W.L.: Uncertainty Weighted Learning [112], -: No Information)
4 Output: T: Trajectory, O: Occupancy ([Prediction Horizon in s], [Class Types Predicted (V: Vehicle, P: Pedestrian, C: Cyclist)])
3 Paradigm: ([M: Motion, A: Appearance], [Imp. T: Implicit Tracking, Exp. T: Explicit Tracking])
2 Input: BB: Bounding Boxes, LR: Latent Representations (Number of Frames Input, M: Use of Memory, Map: Use of Map Information)
1 Modality: (A: Agnostic, MVI: Multi-View Image)

Table 4. TABLE IV: Full-Unified Perception Approaches: Overview

Publ.	Year	Modality11footnotemark: 1	Input22footnotemark: 2	Paradigm33footnotemark: 3	Output44footnotemark: 4	Architecture	Learning Strategy55footnotemark: 5	Dataset66footnotemark: 6	Code
PIP [138]	2022	MVI	(4, -)	M, Imp. T	T(6, V)	ResNet-50[41] + Transformer E.-D. + Motion Interactor	Emp. W.L.	nuScenes	✗
JLA [139]	2022	I	(10, -)	A, Exp. T	T(2, P)	FairMOT[61] (JDTE, Table II) + RNN D.	Unc. W.L.	MOT15/16/17/20	✗
S2F2 [140]	2022	I	(1+M, -)	A, Exp. T	T(1, P)	FairMOT[61] (JDTE, Table II) + GRU + Flow D.	Unc. W.L.	MOT17/20	✗
ViP3D [141]	2023	MVI	(1+M, Map)	A, Exp. T	T(6, V,P)	MOTR[80] (JDTQ, Table II) + Query-based Predictor	Unb. W.L.	nuScenes	✓
ODTP [142]	2023	MVI	(1+M, Map)	M, Exp. T	T(4, V,P,C)	QD-3DT[44] (JDTQ, Table II) + DCENet++[143]	-	nuScenes	✗
FaF [144]	2018	L	(5, -)	M, Imp. T	T(1, V)	One-stage CNN	Emp. W.L.	ATG4D	✗
IntentNet [145]	2018	L	(10, Map)	M, Imp. T	T(3, V)	One-stage CNN	Emp. W.L.	ATG4D	✗
SPF2 [146]	2020	L(RV)	(2, -)	M, Imp. T	T(6, A)	Unsupervised CNN E. + LSTM + CNN D.	Emp. W.L.	kitti, nuScenes	✗
LaserFlow [147]	2020	L(RV)	(5, -)	A + M, Imp. T	T(3, V,P,C)	CNN Backbone + Probabilistic Trajectory Head	Emp. W.L.	nuScenes, ATG4D	✗
SpAGNN [148]	2020	L	(10, Map)	A + M, Imp. T	T(3, V)	PIXOR-backbone[149] + Spatial-aware GNN	-	nuScenes, ATG4D	✗
PnPNet [150]	2020	L	(5, Map)	A + M, Exp. T	T(3, V,P)	CNN + Discrete-continuous Tracker + Predictor	Unb. W.L.	nuScenes, ATG4D	✗
STINet [151]	2020	L	(6, Map)	A + M, Imp. T	T(3, P)	CNN Temporal RPN	Emp. W.L.	WOD, Lyft	✗
ILVM [152]	2020	L	(1+M, Map)	A + M, Imp. T	T(5, V)	CNN + GNN + Trajectory D.	Emp. W.L.	nuScenes, ATG4D	✗
Casas et al.[153]	2020	L	(1, Map)	A + M, Imp. T	T(5, V)	SpAGNN-inspired[148] + MTP [154] Sampler	Emp. W.L.	nuScenes, ATG4D	✗
SDP-Net [155]	2020	L	(5, -)	M, Exp. T	T(0.5, V)	BEV Scene Flow Estimator + CNN Backbone	Emp. W.L.	kitti	✗
RV-FuseNet [156]	2021	L(RV)	(5, -)	A + M, Imp. T	T(3, V)	CNN Backbone + Per-point Prediction	Unb. W.L.	nuScenes	✗
Phillips et al. [157]	2021	L	(1+M, Map)	A + M, Imp. T	T(5, V,P)	ILVM[152] + Localization	Emp. W.L.	ATG4D	✗
MultiXNet [158]	2021	L	(10, Map)	M, Imp. T	T(3, V,P,C)	IntentNet[145] + Prediction Refinement Stage	Emp. W.L.	nuScenes, ATG4D	✗
SDAPNet [159]	2021	L	(5, -)	A + M, Imp. T	T(2.5, V)	EfficientNet-backbone[160] + Multi-task Head	Emp. W.L.	nuScenes	✗
FutureDet [161]	2022	L	(20, -)	M, Exp. T	T(3, V,P)	CenterPoint-backbone[48] + Backcasting	-	nuScenes	✓
FS-GRU [162]	2022	L	(1+M, Map)	A + M, Imp. T	T(3, V)	FPN-backbone[163] + ConvGRU + D.	Emp. W.L.	Argoverse	✗
DeTra [164]	2024	L	(5, Map)	A + M, Imp. T	T(5, V)	PointNet-backbone[165] + Refinement Transformer	Emp. W.L.	Argoverse2, WOD	✗
Interaction Transformer [166]	2020	I+L	(10, Map)	A + M, Imp. T	T(3, V)	CNN E. + Interactive Forecasting Transformer	Emp. W.L.	nuScenes, ATG4D	✗
LiRaNet [167]	2020	L+R	(5, Map)	M, Imp. T	T(3, V,P,C)	Spatio-temporal RaDAR Feature Learning + PnPNet[150]	Emp. W.L.	nuScenes, X17k	✗
MVFuseNet [168]	2021	L+L(RV)	(5, Map)	A + M, Imp. T	T(3, V,P,C)	Multi-view Asymmetric U-Net[169] + CNN Heads	Unb. W.L.	nuScenes	✗
Fadadu et al. [170]	2022	I+L+L(RV)	(10, Map)	A + M, Imp. T	T(3, V,P,C)	Multi-view Fusion + MultiXNet[158]	Emp. W.L	nuScenes, ATG4D	✗
FIERY [171]	2021	MVI	(3, -)	A + M, Exp. T	O(2, V)	LSS[172] + CNN-GRU	Unb. W.L.	nuScenes, Lyft	✓
BEVerse [173]	2022	MVI	(3, Map)	A + M, Imp. T	O(2, V,P,C)	Transformer E. + Parallel D. Heads	Emp. W.L.	nuScenes	✓
StretchBEV [174]	2022	MVI	(3, -)	A + M, Imp. T	O(6, V)	Stochastic Residual State-space Temporal Model	-	nuScenes	✓
PowerBEV [175]	2023	MVI	(3, -)	A + M, Exp. T	O(2, V)	LSS[172] + Multi-scale U-Net-inspired[169] E.-D.	Unc. W.L.	nuScenes	✓
TBP-Former [176]	2023	MVI	(3, Map)	A + M, Imp. T	O(2, V,P)	Spatio-temporal Pyramid Transformer	-	nuScenes	✓
SA-GNN [177]	2021	L	(5, Map)	A + M, Imp. T	O(7, P)	CNN + Scene-Actor GNN	Emp. W.L.	nuScenes, ATG4D	✗
ImplicitO [178]	2023	L	(5, Map)	A + M, Imp. T	O(5, V)	PIXOR[149] E. + Implicit D.	Emp. W.L.	Argoverse2, HighwaySim	✗
Khurana et al.[179]	2023	L	(6, -)	A + M, Imp. T	O(3, V,P,C)	Voxelization + Differentiable volumetric Render	-	nuScenes	✓
Occ4Cast [180]	2023	L	(10, -)	A + M, Imp. T	O(1, C)	CNN E. + LSTM + D.	-	Lyft, Argoverse, ApolloScape	✓
FISHING Net [181]	2020	MVI+L+R	(5, -)	A + M, Imp. T	O(2, V,P,C)	Per-modality CNN E.-D. + Aggregation	Emp. W.L.	nuScenes, Lyft	✗
6 Dataset: (kitti:[25], nuScenes:[26], WOD:[27], Lyft: [182], Argoverse: [125], Argoverse2:[98], ATG4D: [144], X17k: [167], HighwaySim: [178], ApolloScape: [183], MOT15:[99], MOT16:[100], MOT17:[100], MOT20:[101])
5 Learning Strategy: (Emp. W.L.: Empirical Weighted Learning, Unb. W.L.: Unbalanced Weighted Learning, Unc. W.L.: Uncertainty Weighted Learning [112], -: No Information)
4 Output: T: Trajectory, O: Occupancy ([Prediction Horizon in s], [Class Types Predicted (V: Vehicle, P: Pedestrian, C: Cyclist, A: Agnostic to Class)])
3 Paradigm: ([M: Motion, A: Appearance], [Imp. T: Implicit Tracking, Exp. T: Explicit Tracking])
2 Input: (Number of Frames Input, M: Use of Memory, Map: Use of Map Information)
1 Modality: (I: Image, MVI: Multi-View Image, L: LiDAR, L(RV): LiDAR Range View, R: RADAR)

Full text

I Introduction

Autonomous driving software is typically organized into four modules: sensing, perception, planning, and control. Sensing collects information about the environment, mostly through cameras, LiDAR, or RADAR. Perception then interprets this information, providing the basis for planning, which generates safe and efficient trajectories, while control executes these on the vehicle.

In this construct, perception plays a central role, as it enables the vehicle to interpret its surroundings, understand the behavior of road users, and anticipate their future motion, thereby making sense of the environment. Perception can further be divided into two key sub-tasks:

• •

Localization – Where am I?

• •

Scene Understanding – Where is what, going where?

While localization ensures precise local and global ego positioning, scene understanding covers detection, tracking, and prediction of surrounding objects, providing a comprehensive view of the ego vehicle’s environment. This survey focuses on the latter, primarily defining perception as scene understanding.

Perception traditionally follows modular pipelines, where detection, tracking, and prediction are handled in separate sub-modules, as in widely used platforms such as Autoware [1] and Apollo [2]. Modular pipelines are interpretable but suffer from error accumulation and limited information flow between sub-tasks.

End-to-end autonomous driving stacks, in contrast, map raw sensor data directly to planning and sometimes control outputs [3, 4, 5]. This tightly integrates sensing, perception, planning, and control, enabling holistic reasoning across tasks. However, explicit perception outputs are lost, reducing explainability and complicating validation in safety-critical settings. In addition, learning the entire driving task at once is highly complex, making such systems difficult to train and data inefficient. Despite these challenges, end-to-end approaches are increasingly adopted both in research, such as UniAD [3], and in industry, with Wayve [4] and NVIDIA [5].

This contrast highlights a key gap: current perception systems either prioritize modular interpretability or global reasoning through interconnectivity, as illustrated in Fig.  1. Unified perception, as an extension of modular pipelines, addresses this gap, integrating scene understanding into a unified module while retaining interpretable perception outputs. Several methods have been proposed in this direction, but vary in scope and design, and no systematic overview exists.

This survey consolidates existing unified perception approaches into a holistic and systemic taxonomy based on input, intermediate and output representation as well as tracking formulation. From this taxonomy, three levels of task integration are deduced: Early Unified Perception (EUP), unifying detection and tracking; Late Unified Perception (LUP), unifying tracking and prediction; and Full Unified Perception (FUP), unifying detection, tracking, and prediction. We further provide an overview of such architectures, training strategies, datasets used, and open-source availability, highlighting the current state of unified perception.

The remainder of this survey is organized as follows: Sec. II introduces the state of the art of modular perception and highlights the unified perception survey gap. Sec. III introduces the proposed taxonomy. Secs. IV–VI survey the three paradigms of EUP, LUP, and FUP. Sec. VII discusses insights and open challenges, and Sec. VIII concludes.

II State of the Art

Perception is a critical task for autonomous driving. Today, it is mainly realized through modular perception approaches, which solve the overall task as a feedforward cascade of clearly defined and self-contained sub-modules.

While modular perception approaches perform well on benchmarks, their real-world use exhibits limitations in robustness and generalizability [6, 7]. These limitations led to the development of unified perception, which extends core ideas of modular perception. Therefore, the modular perception state of the art is first introduced through its current survey landscape. Secondly, unified perception is introduced, and the existing survey gap is highlighted.

Modular perception systems represent an integral part of modular AV software (Fig. 1). They describe a set of clearly defined sub-modules with dedicated objectives, interfaces, and module-specific evaluation metrics. This design maximizes interpretability, enables parallel development, resulting in a well-established state of the art. In detail, modular perception feeds raw sensor data into a detection module to identify objects, followed by a tracking module to associate objects over time with unique IDs, and a prediction module to estimate future motion. Importantly, the intermediate interfaces are often simple, human-interpretable bounding box representations, which facilitate system integration and benchmarking. Having been extensively surveyed, the state-of-the-art of 3D object detection [6, 8, 9, 10, 11, 12, 13], multi-object tracking [14, 15, 16, 17, 18, 19], and motion prediction [19, 20, 7, 21, 22, 23, 24] modules is further introduced in the following.

Object Detection is the task of localizing and classifying objects within sensor inputs into 2D/3D bounding boxes. A holistic overview of 3D object detection is thereby proposed by Mao et al. [8], which categorize methods from the perspective of sensor input (LiDAR, cameras, multi-modal) and working principle (network architecture) into a widely accepted taxonomy. The survey explicitly focuses on network architectures and introduces challenges, advantages, and limitations for each method type. Furthermore, a quantitative comparison of these methods is performed, reporting the performance on three representative public datasets: Kitti [25], nuScenes [26], and WOD [27]. Addressing the current limitations of modular perception, Song et al.[6] provide a survey of 3D object detection, with a focus on robustness, accuracy, and latency. While adapting the general categorization principles established in [8], the authors propose a slightly modified taxonomy w.r.t working principle, still focusing on camera, LiDAR, and multi-modal approaches. Their analysis emphasizes the resilience of each method type to environmental variations, sensor noise, and calibration errors, and underscores the superior robustness of multi-modal systems. The refined categorization aims to better capture the robustness aspects of contemporary detection methods and guide future developments. Further, modality-specific surveys include Ma et al. [9] for cameras, Aung et al.[13] for LiDAR, and Liu et al.[10] for RADAR 3D object detection. Multi-modal object detection, including camera, LiDAR, and RADAR, is additionally surveyed by Wang et al. [11], and a detailed camera-RADAR fusion overview is proposed by Yao et al.[12].

Multi-Object Tracking (MOT) focuses on reconstructing the motion of detected objects up to the current timestep and assigning consistent identifiers to each object instance. In modular perception systems, MOT is typically addressed using the tracking-by-detection paradigm, where object detections are fed into either model-based or learning-based tracking modules. A central challenge in this paradigm is solving the data association problem in real-time, dynamic, and multi-object environments.

Model-based tracking is comprehensively surveyed by Luo et al. [14], who categorize methods based on initialization type, processing mode, and output type (stochastic or deterministic). These approaches are typically structured into three main components: feature extraction (e.g., appearance, motion, interaction modeling), data association, and track management. A core characteristic of model-based tracking is its reliance on features derived from object motion. Kalman filters are widely used to extrapolate object trajectories through motion models and to refine detections over time using a prediction-update mechanism. In this context, Rakai et al. [18] provide a dedicated survey on data association strategies, evaluating their performance with respect to accuracy, robustness, computational cost, and complexity. Key limitations are identified in areas such as occlusion handling, efficiency, and generalization.

In contrast, learning-based tracking focuses on learning similarity metrics between object instances, often based on appearance features or learned motion cues, rather than explicitly modeling object dynamics. Leon et al. [19] differentiate model-based and learning-based tracking methods for camera inputs, subdividing learning-based approaches into convolutional and recurrent neural network models. Building on this, Bashar et al. [15] propose a broader taxonomy of learning-based tracking for camera data, highlighting their ability to exploit rich semantic information. While their taxonomy does not aim to be exhaustive, it captures the main architecture clusters used in practice. Most recently, Agrawal et al. [16] refine the classification of learning-based MOT into nine families based on network architecture. In parallel, Hassan et al. [17] provide a broader historical survey covering both traditional and deep learning-based tracking methods, with continued emphasis on tracking-by-detection.

Motion Prediction in modular perception refers to forecasting the future trajectories of objects. Inputs are typically tracked objects, contextual cues, such as interactions and scene semantics given by the road network, which play a crucial role in prediction accuracy [24]. Similar to tracking, motion prediction methods are broadly divided into physics-based and learning-based approaches, the latter prevailing in the current state-of-the-art.

Physics-based prediction relies on handcrafted motion models grounded in physical assumptions. Karle et al. [7] and Huang et al. [20] classify such approaches alongside learning-based alternatives. Importantly, while these approaches are interpretable and require no training data, the authors highlight the need for incorporating interactions and complex non-linear behavior, two key limitations of purely model-based systems.

Learning-based methods, by contrast, anticipate motion through supervised training and dominate modern prediction research. Huang et al. [20] provide a comprehensive and comparative survey of trajectory prediction methods, categorizing them into classic machine-, deep-, and reinforcement learning models. These are further grouped by their underlying technologies and evaluated both qualitatively and quantitatively. Karle et al. [7] recast learning-based prediction into pattern- and planning-based methods. Pattern-based approaches refer to deep learning models that learn motion patterns from data, while planning-based methods align with reinforcement learning strategies that optimize predicted behaviors through interaction with the environment. A complementary perspective is offered by Ding et al. [21], who survey interaction-aware learning-based prediction. They distinguish between knowledge-free and knowledge-based approaches, depending on whether domain knowledge (traffic rules or common driving behavior) is explicitly incorporated. Their review highlights how structured priors can enhance generalization and safety in data-driven models. To capture recent architectural trends, Huang et al. [23] propose a taxonomy of deep learning-based prediction systems into three categories: scene input representation, context refinement, and prediction rationality improvement. This structure reflects how modern models encode semantic context, agent interactions, and ensure physically plausible and consistent predictions. An emerging direction within learning-based methods is occupancy prediction, where systems forecast future space-time occupancy distributions instead of discrete trajectories. This approach, also covered in [23], has gained traction due to increased computational capacity and enables dense, map-like predictions.

While modular perception provides well-defined interfaces and interpretability, its strict task separation limits the system’s ability to reason globally across the scene and leverage inter-task synergy. For instance, detection errors can propagate directly through tracking and prediction, and useful contextual information is often lost at module boundaries.

In contrast, end-to-end AV stacks leverages the concept of learning the driving task holistically, mapping raw sensor inputs directly to actions upon the environment. Such approaches are on the rise [28], as powerful hardware and data is readily accessible. These systems promise holistic reasoning and generalization across scenarios by learning from large-scale data without intermediate representations or modules. Thus, perception is performed as an integrated part of the end-to-end system, and learned holistically within. However, end-to-end surveys [29, 28, 30, 31, 32] convey such approaches are highly complex to train -defining clear learning objectives is non-trivial-, challenging to verify, and lack interpretability, therefore limiting safety-critical deployment. These limitations motivate the exploration of more balanced solutions to perception.

Unified perception, an extension to modular perception, has emerged as a promising design paradigm that restructures sub-modules into larger, integrated components. This approach leverages the foundational principles of modular systems while explicitly enabling global reasoning and inter-task synergy. Rather than addressing perception sub-tasks in isolation, unified perception integrates them within a shared architecture, promoting deep contextual understanding.

The key idea behind unified perception is that perception sub-tasks, primarily detection, tracking, and prediction, thus the scene understanding definition of perception are inherently interdependent. Inspired by the holistic learning of perception in end-to-end software systems, unified perception aims to leverage shared representations and joint optimization across sub-tasks. This integration not only enhances the flow of information between components but also supports robustness (reduced latency, temporal consistency), generalizability, and interpretability.

To formalize this design space, we categorize unified perception into three types, as illustrated in Fig. 2, based on the degree and stage of task integration:

• •

Early-Unified Perception (EUP): Integrates object detection and tracking into a single module, enabling joint detection and tracking. Unlike modular tracking-by-detection paradigms, EUP can explicitly retain feature information typically lost between separate detection and tracking stages.

• •

Late-Unified Perception (LUP): Integrates tracking and prediction into a single module, enabling joint reasoning over an object’s current state and its future trajectory. By leveraging shared temporal representations, LUP facilitates more coherent and interaction-aware future state estimation.

• •

Full-Unified Perception (FUP): Primarily integrates scene understanding (detection, tracking, prediction) within a single module, representing the highest level of task integration. However, FUP approaches can also incorporate localization. This paradigm solves perception as a holistic, general task, maximizing inter-task synergy and global scene understanding.

As summarized in Tab. I, the current survey landscape reflects no holistic coverage of unified perception —a gap this survey is closing. Early-Unified Perception (EUP), or joint detection and tracking, is often adopted but has not been explicitly surveyed as a distinct paradigm, despite being acknowledged in recent works such as [16, 17]. Late-Unified Perception (LUP) remains unaddressed in literature. Full-Unified Perception (FUP) has recently been surveyed by Dal’Col et al. [33] under the term joint detection and prediction, which we refine. Thus, this survey provides a holistic and structured overview of unified perception. The taxonomy introduced in the following chapter is generally applicable across all unified perception approaches, offering a complete framework that clarifies existing methods and highlights paradigms within this design space.

III Unified Perception Taxonomy

Unified perception, comprising the three paradigms EUP, LUP, and FUP, can be systematically described and analyzed using the holistic taxonomy introduced in Figure 3. This taxonomy offers a comprehensive framework for understanding how detection, tracking, and prediction tasks are unified from a system perspective. It characterizes unified perception systems based on information flow, system boundaries (input/output), intermediate representations, and integration strategies, without constraining architectural choices.

Any component originally developed under the modular paradigm (Sec.II) can, in principle, be integrated into unified perception, provided it supports joint training. These architectural details are decoupled from the taxonomy itself and discussed in Sec.VII. Importantly, the central design element across all unified perception paradigms is tracking, which serves as the intermediate stage connecting object detection and motion prediction. Consequently, categorization is centered around tracking in our taxonomy.

We categorize unified perception systems along two dimensions: (1) the nature of cues used and (2) the tracking formulation, distinguishing between explicit and implicit tracking. Thereby, tracking cues (1) fall into two categories. Appearance-based cues describe features often derived from visual characteristics of objects, such as shape, texture, or learned embeddings. Motion cues, in contrast, rely on spatio-temporal properties like velocity, direction, or kinematic constraints. Systems may combine appearance and motion cues to improve robustness.

The second dimension (2) distinguishes between explicit and implicit tracking. Explicit tracking assigns persistent identities to objects across frames, creating temporally consistent tracks via hard, one-to-one associations. Thereby, two dominant mechanisms exist within explicit tracking: post-network association (PNA) and in-network association (INA), as introduced by Cai et al. [34].

In PNA, identity assignment is decoupled from the model’s inference process. The model produces intermediate outputs such as object features, embeddings, or motion cues, but the actual linking of identities across time is performed in a separate post-processing step. This association typically relies on external heuristics, optimization algorithms, or rule-based matching strategies. Although the model may be trained with temporal supervision, identity persistence is not handled within the network. In contrast, INA, integrates identity tracking directly into the model architecture and training process. The system jointly reasons over temporally extended inputs and assigns consistent identities as part of a single inference pass. Identity persistence is treated as a learnable function of the model, embedded into its representation and output space, eliminating the need for external association logic.

Implicit tracking, on the other hand, avoids persistent identity assignment altogether. Instead, it maintains temporal continuity within the model’s latent representations, enabling soft associations across time, making it a form of INA. This allows the system to maintain temporal context and reason about object dynamics, rather than committing to deterministic identity assignments. While this increases flexibility and supports reasoning over multiple hypotheses, it comes at the cost of losing explicit, interpretable tracks.

A key aspect of our taxonomy is the flexible combination of input types, intermediate representations, and output formats within unified perception. Inputs to EUP and FUP consist of raw sensor data such as radar, LiDAR, or camera images. Between unified tasks, intermediate representations —bounding boxes, occupancy grids, or latent features— are produced and can be freely combined or propagated. For LUP, the system operates on intermediate representations generated after detection as input. This flexibility allows any combination of inputs, architecture, intermediate formats, and outputs depending on system design. Outputs are thereby possible in the form of trajectory, occupancy, or flow representations. Critically, the representation type plays a crucial role in information sharing across sub-tasks, directly affecting explainability, interconnectivity (Fig. 1), and likely system performance.

In summary, this taxonomy presents a general and systemic framework for describing unified perception systems. It defines a structured design space based on task integration, tracking formulation, and representation flow. It supports both retrospective classification of existing methods and prospective design of new approaches. Following, we survey EUP, LUP, and FUP methods in detail.

IV Early-Unified Perception

Joint detection and tracking methods, summarized into EUP, integrate the task of detection and multi-object tracking into a unified module, as explained in Sec. II. For the survey of EUP methods, we follow our taxonomy, introduced in Sec. III.

Specifically, we identify three sub-paradigms for EUP methods from our taxonomy, surveying these along their input modality and the tracking formulation used: Joint Detection and Tracking by Propagation (JDTP), which uses motion cues to predict object displacement and associate instances over time; Joint Detection and Tracking by Embedding (JDTE), which augments detection with appearance embeddings and associates objects by measuring embedding similarity; and Joint Detection and Tracking by Query (JDTQ), which treats detection and tracking as a single set prediction problem, using a fixed set of learnable queries to represent and update object instances and their identities. A visual summary of how these paradigms align with our taxonomy is presented in Fig. 4. Combinations of appearance and motion cues are thereby also possible.

EUP relies exclusively on explicit tracking due to the common use of bounding box outputs, which require bipartite assignments to produce one-to-one matches. Specifically, both PNA and INA methods are employed in this context. While the taxonomy allows for broader combinations, implicit tracking and occupancy representations have not yet been proposed in the EUP context.

An overview of all EUP methods surveyed is summarized in Tab. II. On top of the categorization used below, additional key attributes such as number of input timesteps, network architecture, learning strategy, datasets, and code availability are presented. This structure enables readers to understand both the conceptual differences and practical implementations of EUP approaches across the literature.

IV-A Joint Detection and Tracking by Propagation

JDTP unifies detection and tracking by leveraging motion cues to estimate the physical displacement of detected objects up to the current frame, making it the unified counterpart to modular model-based tracking-by-detection. It primarily relies on intermediate representations such as bounding boxes or ”bounding tubes”, especially in PNA-based methods. In contrast, INA-based methods apply latent representations for temporal association. JDTP is particularly well-suited to sensors with high spatial accuracy, such as LiDAR, due to its explicit use of spatial relationships.

PNA-based JDTP approaches are optimized toward learning temporal cues within the network but rely on external, rule-based data association to output identity assignments. Such methods include CenterTrack [35], which operates on single-image data and extends a keypoint-based object detector CenterNet [36] with additional offset prediction heads that estimate per-object motion between two consecutive image frames. The predicted 2D center displacements, essentially sparse optical-flow-like displacements, then allow detections to be greedily matched to prior instances without any external optimization, making the object detection and displacement learning process differentiable and end-to-end. An extension to 3D image-based JDTP is proposed with QD3DT [44], which adopts a hybrid approach that integrates 3D detection, based on a Faster-RCNN detector [45] with a motion-aware LSTM and depth-ordering heuristics. While it incorporates temporal modeling and instance-level embedding within the network, identity association is performed post hoc using a handcrafted affinity matrix and greedy assignment strategy. Lastly, CenterPoint [48], a LiDAR-based JDTP approach, unifies detection and tracking simply by regressing object velocities directly from temporal pointcloud data through an additional head. Identity matching is then performed post-network, by greedily associating predicted object centers across frames based on spatial proximity.

INA-based JDTP, embeds identity association within the network implicitly, typically through motion-aware heads, memory modules, or the use of recurrent or Transformer-based query propagation. For image-based JDTP, TubeTK [37], reformulates tracking as tube prediction: a 3D convolutional network regresses short temporal sequences of bounding boxes (tubes) in a single forward pass. By capturing intra-tube temporal consistency, TubeTK performs joint detection and short-term association without requiring external matching during inference. However, longer-term identity preservation is not modeled within the network, and thus tube-level non-maximum suppression in a PNA manner remains necessary. Subsequent models introduced internal memory mechanisms to extend tracking beyond short windows. PermaTrack [39] augments CenterTrack [35] with a spatio-temporal recurrent memory via ConvGRU modules, enabling the network to maintain persistent object representations across time. This design allows the model to recover from occlusions and re-identify objects after temporary disappearances—all within the same end-to-end trainable framework. Recently, Transformer-based architectures have enabled rich temporal modeling allowing learnable motion queries to be propagated across frames. P3AFormer [40] utilizes a DETR-like architecture with a Swin-Transformer encoder and feature propagation via optical flow. Here, object queries act as persistent identity carriers, updated at each timestep through attention-driven cross-frame interactions. STAR-Track [46] extends this concept by incorporating a Latent Motion Model (LMM) that encodes dynamic state information into each object query. This allows the attention mechanism to account for both spatial displacement and appearance variation, fully integrating identity management into the Transformer pipeline without any post hoc logic.

For LiDAR-based JDTP approaches, SimTrack [51] employed hybrid-time centerness maps and motion offsets to propagate identities, embedding temporal logic directly in the detection heads. The concept of temporal regression from TubeTK [37] is further extended by CenterTube [52], which generalizes 2D/3D tubelets to 3D/4D space, linking LiDAR detections over time via spatio-temporal regression branches. Furthermore, VoxelNeXt [53] extends CenterPoint [48] by introducing sparse voxel feature propagation across frames. The network encodes voxel-level temporal continuity, enabling identity association to emerge as a property of persistent spatial features, with no reliance on explicit matching modules.

Finally, multi-modal models such as MMF-JDT [54] combine image and LiDAR inputs to exploit complementary spatial and semantic cues. The network integrates features from past tracklets directly into the fusion-based detection heads, enabling joint reasoning over appearance and geometry. Temporal associations are resolved via shared attention layers and NMS modules operating within the network, maintaining consistent identities across sensing modalities.

IV-B Joint Detection and Tracking by Embedding

JDTE focuses on unifying object detection and tracking by using appearance embeddings derived within the system. The learning process centers on teaching the network to distinguish between similar and dissimilar objects by optimizing embedding similarity through contrastive or association-based losses, without requiring explicit labels for the embeddings themselves. Therefore, latent intermediate representations are propagated through the system. During inference, PNA-based JDTE methods perform association outside the network using the learned embeddings and standard data association techniques, such as those surveyed by Rakai et al. [18], while INA-based methods integrate the association directly within the network’s inference process.

PNA methods include JDE [55] and RetinaTrack [57], which jointly predict detection outputs and appearance embeddings in a single network pass, but rely on external association mechanisms. In JDE, object detections and embeddings are extracted via a Feature Pyramid Network (FPN) backbone, and identity assignment is performed post hoc using a cost matrix composed of cosine appearance similarities and Mahalanobis motion distances, solved by the Hungarian algorithm [113]. RetinaTrack employs a RetinaNet-based detector with additional regression and embedding heads, and performs association via greedy bipartite matching using the learned embeddings, without incorporating association into the forward pass. FairMOT [61] similarly produces detections and embeddings from a shared backbone but applies hierarchical association in two steps: Kalman filtering for motion prediction, followed by Hungarian matching. DEFT [60] integrates detection with an appearance embedding extractor and a matching head that outputs affinity matrices; during inference, the Hungarian algorithm is applied to the computed affinities, incorporating motion constraints via an LSTM-based forecasting module. CorrTracker [63] augments the detection pipeline with a correlation-learning module but retains external Hungarian matching for identity assignment. AlignPS [64] focuses on learning robust embeddings from an anchor-free detection architecture, using external similarity-based association during inference. GSDT [62] and 3D DetecTrack [73] embed graph-based feature aggregation into the network to refine appearance and motion cues, but still compute affinities that are matched post-network via combinatorial algorithms. Time3D [66], JMODT [71], AlphaTrack [72], and Minkowski Tracker [69] extend this approach to multi-modal data by learning spatial-temporal embeddings or affinity scores internally, yet perform identity association as a discrete post-processing step during inference, using strategies such as greedy-, Hungarian matching [113], or mixed-integer optimization.

INA methods embed the association mechanism directly within the network inference pipeline, enabling simultaneous prediction of detections and identity associations without reliance on external solvers. Siamese Track-RCNN [58] modifies Faster R-CNN [41] by incorporating dedicated branches for detection, tracking, and re-identification. The model outputs detection results and identity associations within a single forward pass by leveraging the tracking and re-ID branches to propagate tracklets, thereby eliminating the need for post hoc matching algorithms such as the Hungarian method. CTracker [59] utilizes a dual-branch backbone with shared weights to extract features from two consecutive frames. The frame-wise features are concatenated and fed to a paired bounding box regression module, which is conditioned on attention maps from object classification and identity verification modules. This mechanism performs implicit association between frames by directly regressing object pairings, making the model end-to-end. TraDeS [68] introduces a cost volume-based association module and a motion-guided feature warper. The cost volume computes similarity scores between re-identification embeddings across frames, which are then used to estimate tracking offsets. These offsets guide the motion-aware propagation of features across time, allowing the network to jointly infer detection and association results without a separate data association stage.

IV-C Joint Detection and Tracking by Query

Lastly, JDTQ unifies detection and tracking by formulating the latter as a set prediction problem, where object instances are represented and maintained through learnable query embeddings. These latent intermediate representations are propagated across frames using attention-based mechanisms in Transformer-based architectures, enabling both detection and identity association within the same framework. As the association is performed internally, JDTQ methods are inherently INA. The core idea is the use of specialized appearance-based queries —typically divided into detection queries for identifying new objects and track queries for propagating known identities. These queries interact with encoded input features via cross-attention to retrieve, localize, and maintain object representations over time.

Early image-based JDTQ methods include TransTrack [75], which introduces a hybrid query mechanism by fusing detection and tracking outputs. TransTrack applies learned detect queries for new object discovery and injects past-frame features as track queries, relying on concatenation and late fusion to jointly perform detection and association. However, this design partially relies on handcrafted components and post-decoding IoU matching, which makes it PNA. TrackFormer [76] advanced the field by propagating learned track queries across frames via decoder memory, enabling consistent identity tracking through learned attention without external association heuristics. TransCenter [77] challenged the reliance on sparse, content-independent queries by adopting dense detection queries derived from image features. It updates both dense and sparse queries via Transformer attention mechanisms, improving robustness to occlusions and small objects. MOTR [80] marked a significant milestone by extending DETR to video, wherein track queries are updated frame-by-frame through decoder interactions. By carrying object-specific embeddings across time, MOTR achieves fully integrated detection and association. To enhance long-term tracking, MeMOT [34] introduced spatio-temporal memory modules that store historical features. This decouples association from strict query propagation, instead relying on similarity-based memory retrieval. Building on MOTR, MOTRv2 [81] and MOTRv3 [83] refine training strategies—introducing external detection priors and label assignment mechanisms—to improve the balance between detection and tracking supervision, while maintaining an end-to-end association pipeline. Subsequent models further enhance query dynamics. Co-MOT [84] incorporates a cooperative label assignment framework, promoting mutual learning between detect and track queries. MeMOTR [85] augments query refinement with a long-term memory that stabilizes tracking through extended temporal contexts. Recent developments, such as LAID [86] and SambaMOTR [87], push the frontier toward practical deployment: LAID integrates pretrained detectors and novel alignment modules to balance accuracy and training efficiency, while SambaMOTR enhances robustness via synchronized memory units that propagate queries through occlusions.

In the multi-camera setup, JDTQ methods extend the query paradigm to handle cross-view consistency. MUTR3D [88] is one of the earliest to propose multi-view Transformer-based tracking by introducing view-aware deformable attention to process multiple images simultaneously. DQTrack [89] builds on this by decoupling detection and tracking queries and optimizing temporal query propagation with deformable attention layers and visibility priors. Ada-track [90] introduces adaptive temporal query selection based on object visibility and confidence, helping reduce redundant computation and false positives in cluttered scenes. OneTrack [91] proposes a unified query interaction design where object queries are fused across views and time, enabling effective spatio-temporal reasoning without post-processing. MCTR [93] pushes this further by incorporating contrastive query refinement and explicit cross-view temporal correlation, enabling robust identity preservation in crowded and occluded settings.

For LiDAR-based tracking, JDTQ methods are emerging. JDT3D [94] extends the JDTQ framework to point cloud data by adapting detection and track queries to operate on voxelized 3D features. It leverages temporal deformable attention to propagate object-specific queries across time, allowing consistent identity modeling in 3D space. Mask4Former [95] brings a mask-based query formulation to the LiDAR domain, learning spatio-temporal segmentation masks through Transformer decoders that directly correspond to tracked objects. This formulation offers both precise localization and robust temporal association, marking a shift toward dense tracking representations in 3D scenes.

Combining image and LiDAR, MotionTrack [96] exemplifies multi-modal JDTQ by fusing vision and point cloud features within a unified Transformer framework. It employs joint queries that attend to both modalities via modality-specific cross-attention, then aggregates the outputs for detection and tracking. By synchronizing feature maps and aligning temporal information across modalities, MotionTrack enables accurate object localization and tracking even under challenging lighting or occlusion conditions. This fusion strategy demonstrates the versatility of JDTQ when extended to heterogeneous data sources and sets the stage for more sophisticated multi-modal tracking solutions.

V Late-Unified Perception

Joint tracking and prediction methods, summarized into LUP, integrate tracking and prediction into a unified module and are motivated by two key considerations. First, jointly addressing tracking and prediction can reduce identity switches and track fragmentations [114], which directly influence prediction performance [115]. Second, shared spatio-temporal information about the ego vehicle’s environment, can be directly leveraged enabling coherent reasoning over object dynamics in the past and future.

Following our taxonomy (Sec. III), LUP can take detection results in the form of bounding box, occupancy or latent representations as input, and output trajectory, occupancy or flow prediction. Tracking can then be performed based on motion and/or appearance cues in combination with explicit or implicit tracking. Specifically, proposed LUP and FUP diverge from EUP in the tracking formulation, as tracking results are not the system output. Therefore, both explicit (hard associations) and implicit tracking formulations (soft) are feasible, as illustrated in Fig. 5. As tracking is addressed in union with prediction, PNA is not feasible, as post hoc association would break the joint reasoning process and prevent shared optimization between tasks.

The LUP methods surveyed are summarized in Tab. III, which provides complementary information about inputs, network architecture, learning strategy, datasets, and code availability in accordance with the survey scope. As most surveyed LUP methods use motion cues, the following discussion of LUP methods is centered around their tracking formulation, working principles and output representations. Intermediate representation types are also pointed out.

Implicit LUP approaches with trajectory prediction output, include PTP, proposed by Weng et al.[116], who introduce a parallel framework for joint tracking and prediction. They propose a Graph Neural Network (GNN) taking fed back trajectory information and bounding box representations and encoding those to interactions among agents and providing shared features for two parallel heads: a 3D MOT head for explicit association and a parallel forecasting head using a Conditional Variational Autoencoder[117]. The prediction module relies only on the GNN features, passed as latent representations, not on association results, allowing it to model temporal consistency and interactions implicitly, where trajectory prediction operates without enforcing explicit object identities. Building on this idea, Weng et al. [118] propose MTP, introducing multi-hypothesis tracking and prediction to further reduce the impact of association errors. Instead of committing to a single assignment, MTP generates multiple plausible tracking hypotheses using Murty’s H-best assignment [126]. All bounding box tracklet sets are passed to the prediction module, which reasons over them jointly. This renders tracking partially implicit, as prediction no longer depends on a single association. However, the tracking hypotheses themselves are generated via a deterministic, non-differentiable process external to the network, a form of PNA, meaning MTP does not fully realize end-to-end implicit tracking but rather proposes a motion prediction algorithm capable of coping with multi-hypothesis inputs. Further advancing this line of work, Weng et al. [114] introduce AffiniPred, which pushes implicit tracking to a stricter formulation by entirely removing tracklet generation from the pipeline. Instead of relying on association hypotheses, AffiniPred uses affinity matrices across frames to encode “soft” identity cues between detections. These affinity matrices serve as the sole input to a Transformer-based architecture that models interactions and predicts future trajectories. Temporal consistency is thus learned directly from affinity and bounding box information. This strictly implicit formulation with latent intermediate representations avoids potential limitations of association-based hypotheses while preserving rich information about object similarity across time. In a similar spirit, Zhang et al. [119] present TTFD, predicting future motion directly from raw detections without forming explicit trajectories. The method uses an affinity-aware mechanism to estimate short- and long-term similarities between detections across time, capturing both appearance and motion cues, sharing latent representations within. These affinity features guide a recurrent state update process that encodes temporal information and uncertainty without enforcing hard associations. The resulting motion representation is then decoded into future trajectories using a social interaction modeling function. By reasoning over affinity cues rather than fixed tracks, TTFD maintains temporal consistency while preserving uncertainty over object identities.

Uhlemann et al. [122] propose an implicit tracking, occupancy prediction approach that eliminates the need for object association altogether. Instead of learning object-wise distributions from scratch, detection clustering using DBSCAN is applied [127] for training, representing each cluster as a Gaussian distribution that captures its spatial extent and uncertainty. The model is trained to predict Gaussian occupancies directly from raw bounding box detections, using a snapshot-inspired [123] encoder and a decoder inspired by HiVT [124]. Thus the network propagates latent intermediate representations. Crucially, no clustering or association is performed during inference — the model reasons over the scene holistically and outputs spatial occupancy distributions, that implicitly encode temporal and spatial consistency, without relying on persistent object identities.

Explicit tracking LUP approaches with trajectory output benefit from directly propagating instances with uniquely assigned encoded past motion. Thus, this type of LUP is directly developed towards long-term resistance to occlusions, interpretability, and stable reasoning in a unified way. To this end, Pang et al. [120] propose PF-Track, a LUP method built to input latent detection representations as input. Specifically, these inputs are retrieved from any JDTQ-style (EUP) detector which leverage query representations for detection. PF-Track primarily utilizes a Transformer framework, developed from SceneTransformer [128], but is technically detector-agnostic. The LUP method focuses on spatio-temporal coherence, introducing past and future reasoning modules. The past reasoning module refines object features and bounding boxes using historical query states, allowing for explicit and persistent object identities, while the future reasoning module predicts trajectories that stabilize query propagation under occlusions or low-confidence detections. A track extension mechanism uses predicted trajectories to maintain object positions when detections are missing, enabling re-identification upon reappearance. Building on query-based LUP, Zhuang et al. [121] introduce StreamMOTP, a streaming framework advancing explicit LUP with a memory bank that stores long-term latent features for each object. This memory maintains persistent identity tracking across frames and enhances association and prediction. A spatio-temporal encoder extracts context features from current detections and historical tracklets, while unique identities from the memory are explicitly associated using optimal transport with a differentiable log-Sinkhorn algorithm [129]. Trajectory prediction is performed via a dual-stream Transformer decoder: the primary branch predicts trajectories for current detections, while the auxiliary branch leverages historical features for identity-consistent, intention-aware prediction. Through its memory-based design, StreamMOTP tightly integrates long-term identity preservation into tracking and prediction, improving temporal consistency and robustness to occlusions.

VI Full-Unified Perception

Joint detection, tracking and prediction methods, summarized into FUP, integrate scene understanding as one unified module, enabling holistic reasoning about the perception task. FUP approaches are increasingly being proposed, and have, to some extent, been surveyed by Dal’Col et al. [33] in their review of ”joint detection and prediction” methods. Critically however, the authors include methods extending to planning [130, 131, 132, 133, 134, 135, 3, 136, 137], which we strictly separate from FUP (Fig.  2). Furthermore, their analysis is limited to input–output formulations primarily, without addressing systemic integration options, whereas we propose an overview within our taxonomy (Sec. III).

From that, FUP can be designed with camera, LiDAR, and RADAR combination inputs and output motion in the form of trajectories, occupancy, or flow. Our survey focuses on trajectory and occupancy, illustrated in Fig.  6 due to their common adoption in real-world driving systems. For FUP with flow output, we refer to Dal’Col et al. [33] for deeper insights, noting that our taxonomy holds for these methods. Importantly, methods that incorporate localization in addition to scene understanding tasks are considered as valid FUP methods, but are highlighted as such. Intermediate representations propagated in FUP can be of bounding box, occupancy, and latent nature, and tracking can be performed implicitly and explicitly (as pointed out in Fig.  5).

A summary of the surveyed FUP methods is provided in Tab. IV, which provides complementary information to the following method discussion. Because FUP takes raw sensor information, tracking cues are mostly appearance-motion combinations. We discuss methods based on input and output, intermediate representation, as well as their tracking formulation and working principles.

VI-A Trajectory Output and Implicit Tracking

Trajectory-based FUP methods leveraging implicit tracking are proposed for camera, LiDAR, and multi-modal inputs, the order of which we keep in the following. These methods mostly propagate latent intermediate representations.

Firstly, for multi-view camera inputs, PIP [138] uses a Transformer-based framework that jointly performs online mapping, object detection, and forecasting. Perception and motion queries interact with agent-wise map features to produce temporally consistent trajectory outputs, with implicit tracking achieved through a unified scene representation and latent intermediate representations.

Secondly, for LiDAR inputs, FaF [144] stacks multiple sweeps into a 4D tensor, with a single-stage detector predicting bounding boxes across time and aggregating overlaps to implicitly handle tracking and short-term forecasting. Building on this, IntentNet [145] incorporates HD maps via temporally corrected sweeps. A two-stream backbone processes LiDAR and map features to jointly produce 3D detections, intention classes, and trajectories, with tracking handled implicitly through fused temporal features -an early integration of environmental context into FUP models. To better capture interactions, SpAGNN [148] applies graph neural networks to refine actor states from LiDAR (PIXOR-backbone [149]) and map inputs. Actor features are iteratively updated via message passing, enabling coherent trajectory prediction without explicit identity association and underscoring the growing role of relational reasoning in the pipeline.

Further evolution comes with STINet [151], which explicitly constructs spatio-temporal features using pillar-based LiDAR encodings. A Temporal Region Proposal Network links bounding box representations over time, while an interactive feature extractor learns movement and interaction patterns. Tracking remains implicit, even though the proposal association introduces a more structured temporal linking. Advancing scene understanding, ILVM [152] models interactions through a latent variable graph, marking a shift towards probabilistic and generative approaches. Actor features are refined via message passing, and future trajectories are sampled from the latent scene representation, ensuring temporal consistency and implicit tracking. Recognizing the importance of traffic context, Casas et al. [153] extend SpAGNN [148] by incorporating traffic rule priors as soft constraints. A probabilistic trajectory head based on MTP [154] predicts multi-modal futures, while maintaining implicit tracking via latent space continuity. This reflects a growing trend towards embedding structured prior knowledge into the modeling process.

SPF2 [146] offers a distinct perspective by inverting the perception pipeline. Instead of directly detecting objects, it predicts future point clouds with an auto-regressive unsupervised model, followed by standard detection on these forecasts. Object continuity and tracking arise naturally from consistent predicted sensor data, a form of latent representation. Meanwhile, methods like LaserFlow [147] and RV-FuseNet [156] explore temporal fusion in LiDAR range image representations with shared latent spaces. LaserFlow aggregates consecutive range images via curriculum learning to predict trajectories relative to LiDAR points, while RV-FuseNet warps and aligns features across frames before instance segmentation and trajectory averaging. Both achieve implicit tracking through temporal alignment and segmentation, reinforcing the role of raw sensor space fusion.

Further developments to FUP are proposed by Phillips et al. [157], who embed localization into an ILVM-based framework. By combining LiDAR and semantic maps into a unified latent representation and refining actor states via a GNN, they jointly decode future trajectories while maintaining implicit tracking through latent continuity. Building on this line of enhancing prediction quality, MultiXNet [158] extends IntentNet with uncertainty-aware forecasting, decomposing positional uncertainties and adding a second-stage refinement network to improve forecast accuracy, while retaining implicit tracking via feature association and latent refinement.

SDAPNet [159] processes multi-frame BEV (Bird’s Eye View) LiDAR representations using temporal fusion and a multi-task detection head. Temporal consistency is maintained through feature aggregation, eliminating the need for explicit identity tracking, while reinforcing the trend of BEV-based temporal modeling. Addressing long-term temporal reasoning, FS-GRU [162] integrates a ConvGRU that shares features across frames and latent representations through the model, coupled with HD map information and temporal alignment of hidden states. The most recent advancement, DeTra [164], reformulates FUP as an iterative trajectory refinement task. By processing LiDAR and map features into BEV representations and leveraging Transformer-based pose updates, object queries (latent representations) are refined across multiple iterations. This results in multi-modal, interaction reasoning, trajectory forecasts with implicit tracking naturally emerging from the temporally consistent query refinement.

Thirdly, for multi-modal FUP methods, Interaction Transformer [166] fuses LiDAR, cameras, and HD maps in a common BEV detection network. A subsequent recurrent interactive Transformer-based module captures spatial-temporal dependencies between agents, while a per-time-step refinement mechanism incorporates recent behavioral cues, enabling consistent trajectory forecasting without explicit identity association. LiRaNet [167] extends this idea by incorporating RADAR alongside LiDAR and camera inputs. Modality-specific BEV feature volumes are generated and fused within a shared backbone, followed by a PnPNet [150] inspired prediction backbone. Temporal feature association across consecutive sweeps ensures robust trajectory forecasting with implicit tracking. Building on multi-view fusion, MVFuseNet [168] jointly processes range-view and BEV representations within a temporal fusion network. Data from consecutive sweeps are projected across frames, while complementary sensor views and map information are fused for joint spatio-temporal feature learning. This enables accurate detection and trajectory forecasting, with implicit tracking emerging from temporal feature aggregation. Finally, Fadadu et al. [170] propose a multi-view fusion framework that integrates cameras and LiDAR (point cloud and RV-based encodings) into a MultiXNet-style architecture [158]. A range view branch processes point clouds and images before fusion into BEV space, while map and multi-sweep LiDAR features are incorporated jointly. This design enables rich spatial and semantic representation, with temporally consistent trajectory forecasts achieved through cross-modal feature association.

VI-B Trajectory Output and Explicit Tracking

FUP with trajectory outputs and explicit tracking formulations have been proposed with camera and LiDAR inputs. Notably, several methods build on EUP (Tab.  II), which naturally employ explicit tracking, extending these to the FUP setting.

For one, JLA [139] extends FairMOT [61] (JDTE) by adding a trajectory forecast head: RNNs encode past bounding boxes and velocities to latent intermediate representations, while decoding predicts future states that replace Kalman Filter estimates during online association. A three-step data association ensures robust identity tracking, with lost tracks removed after a fixed number of missed frames. A different approach, S2F2 [140] also builds upon FairMOT [61] (JDTE) by introducing future flow prediction into an anchor-free, single-stage architecture propagating latent representations. The system uses the FairMOT backbone for context feature extraction, followed by a GRU-based encoder-decoder that estimates residual future flows as updates to an initialized flow field, propagating latent representations. These predicted flows guide online association by adaptively lowering confidence thresholds when future positions align with existing tracks. ViP3D [141] builds on MOTR [80] (JDTQ) with an explicit trajectory-level association based on agent queries. A dynamic memory bank preserves identity, while a DETR3D-based [47] extractor updates the queries, which are decoded into future trajectories using a map encoder and trajectory decoder.

Furthermore, ODTP [142] uses QD3DT [44] (JDTP) and combines it with DCENet++[143], an explicitly linked trajectory prediction module. Persistent agent queries in QD‑3DT enable identity association across frames, with their trajectories and poses feeding into dynamic BEV maps that capture motion, shape, and interaction cues. DCENet++ applies self-attention and CVAE to forecast diverse future trajectories, trained directly on noisy tracker outputs rather than curated ground truth, ensuring joint learning and explicit identity preservation. ODTP [142] integrates QD3DT [44] (JDTP) with the trajectory prediction module DCENet++ [143]. Thereby, ODTP uses bounding box intermediate representations between QD3DT and DCENet++, which then applies self-attention and a CVAE to forecast diverse future trajectories.

In the LiDAR domain, PnPNet [150] presents a differentiable, end-to-end framework that outputs object tracks and future trajectories at each timestep, propagating bounding box representations. It uses a discrete-continuous tracking module combining learned affinities with Hungarian matching for joint data association and trajectory estimation, aided by single-object tracking for occlusions. Trajectory-level representations encode temporal dynamics for prediction. Importantly, gradient flow through the association step is not addressed. A lightweight approach is proposed with SDP-Net [155], which extracts BEV occupancy maps from LiDAR point clouds and applies 2D convolutions for feature encoding, propagating latent representations. Motion cues are obtained from BEV flow maps, estimating both object and ego-motion. Past features are warped based on this estimated motion, enabling temporal aggregation. Tracking is achieved through spatial association with predicted positions, while motion prediction operates on the aggregated features. Finally, FutureDet [161] directly predicts future object locations from LiDAR backcasting these to infer trajectories through bounding box intermediate representations, based on CenterPoint [48] (JDTP). This explicit reasoning over past, present, and future provides trajectory-level identity association while supporting multi-modal trajectory forecasting.

VI-C Occupancy Output and Implicit Tracking

Occupancy-based approaches differ from trajectory-based methods in their output representation, still jointly reasoning about detection, tracking and prediction as introduced in our taxonomy. Methods leveraging implicit tracking have been proposed with camera or LiDAR inputs.

In the former, image domain, future occupancy is increasingly predicted via BEV-based approaches that maintain temporal consistency without explicit tracking. For example, BEVerse [173] lifts multi-view, multi-timestamp image features into BEV space, aligns them with ego-motion, and propagates latent representations through a spatio-temporal encoder, decoding them into occupancy outputs. Tackling the uncertainty inherent in long-horizon forecasting, StretchBEV [174] models temporal dynamics within a stochastic latent space. Encoded BEV states are updated via residual transitions inferred by a ConvGRU-based recurrent module, with stochastic latent variables injecting diversity into future predictions. Thereby, latent representations are propagated and decoded into occupancy grids as output level. Lastly, TBP-Former [176] employs a Transformer-based design that ensures spatial-temporal alignment despite arbitrary camera poses, sharing latent representations. A pose-synchronized BEV encoder maps image sequences into a unified BEV space, followed by a spatial-temporal pyramid Transformer that extracts multi-scale features for comprehensive future state prediction. Scene priors derived from the latest BEV representation guide the process.

The latter, LiDAR-based methods also use implicit mechanisms for occupancy and motion prediction. For one, SA-GNN [177] constructs a Scene-Actor Graph, where detected pedestrians are nodes linked based on spatial proximity, and the entire scene is represented as a supernode, in latent representation. Voxelized LiDAR and map features are processed using graph neural networks with 2D convolution-based message passing to preserve spatial relations and model interactions. This approach jointly predicts individual trajectories and a shared occupancy map. Lastly, ImplicitO [178] employs implicit continuous occupancy representations from raw LiDAR and map inputs. Its PIXOR-inspired [149] BEV encoder with ResNet and Feature Pyramid Networks produces multi-scale features, while the decoder predicts occupancy and flow by learning spatial offsets from query points, allowing dense spatio-temporal forecasting without explicit object association.

An alternative field generates occupancy outputs without relying on object detection, implicit tracking, and prediction as separate goals. Instead, these methods condense the scene understanding architectures to propagate latent spatio-temporal features to model scene evolution directly. LiDAR and multi-modal approaches have been proposed.

For LiDAR-based inputs, Khurana et al. [179] propose to directly estimate future 4D space-time occupancy along rays defined by timestamp, origin, and direction. This ray-based formulation abstracts away sensor specifics and focuses purely on predicting when occupied space would be encountered, based on past point clouds and ego-pose information. Furthermore, Occ4Cast [180] formulates future occupancy prediction as dense binary estimation over spatial-temporal voxel grids. Using a shared 2D or optionally 3D convolutional encoder with recurrent LSTM modules, the model captures temporal dynamics and predicts future occupancy volumes.

Lastly, the multi-modal FISHING Net [181] introduces a top-down, sensor-agnostic semantic grid for perception and prediction. Independent encoder-decoder networks process LiDAR, RADAR, and camera inputs, producing short-term semantic grids which are fused into a unified, ego-centric scene representation.

VI-D Occupancy Output and Explicit Tracking

Occupancy-based prediction output in combination with explicit tracking approaches have been proposed for image-based inputs. FIERY [171] predicts future instance segmentation and motion from multi-view monocular images by lifting features into latent intermediate BEV representations via Lift-Splat-Shoot [172] and aligning them temporally with ego-motion. A spatio-temporal model with gated recurrent units decodes instance centers, offsets, and occupancy, enabling explicit tracking by warping past locations and matching across frames. PowerBEV [175] similarly encodes multi-view images into BEV and predicts segmentation and backward centripetal flow, explicitly propagating instance IDs for consistent tracking without multi-task complexity, resulting in an occupancy prediction output.

VII Discussion

Having introduced our taxonomy (Sec. III) and the three types of unified perception EUP, LUP and FUP within, interesting observations about these can be deduced. Specifically, a discussion about the intermediate representation types propagated in unified perception and training strategies is proposed. Lastly, key future research directions are pointed out for unified perception.

VII-A Intermediate Representations

As described in our taxonomy, representations shared between unified detection-tracking and tracking-prediction structures can be of bounding box, latent or occupancy nature, whereby the latter is not proposed so far. Therefore, the two former intermediate representation types are discussed, as illustrated in Fig. 7.

Perception unification with intermediate bounding box representations (Fig. 7, top) directly leverages modular perception know-how as base components, unifying through joint model training. To achieve this, gradients must be backpropagated through all modules, updating their weights together and preserving explainability through human-interpretable bounding box outputs. However, this approach is subject to two key drawbacks: differentiability and information bottlenecks.

Differentiability is essential for unified learning, as gradient computation must be supported by all operations. In modular pipelines, however, steps such as argmax in matching or NMS in detection are typically non-differentiable and block gradient flow, restricting joint optimization. EUP mitigates this with the use of NMS-free detectors and post-network association, effectively excluding these operations from unified training, such as CenterPoint [48]. In LUP and FUP, the use of intermediate bounding box representations, by definition, requires bounding-box tracking outputs. Thus, tracking must be explicit. However, while explicit tracking is a necessary feature, it is not exclusive to it; explicit tracking can also exist within approaches that follow the latent representation paradigm. As a result, differentiable matching alternatives such as the Sinkhorn algorithm [184] can be used, as with StreamMOTP [121]. In FUP, FutureDet [161] follows a CenterPoint[48] based approach, essentially excluding the matching step from learnable structures.

Information bottlenecks, in turn, arise because interfaces constrained to bounding box representations force the network to compress rich feature information into a few geometric parameters. While these representations are interpretable, they discard much of the spatial and semantic detail captured in earlier layers. As a result, the model’s capacity to learn complex inter-task dependencies may be reduced, affecting overall performance.

In contrast, unification with latent intermediate representations (Fig. 7, bottom) relaxes module boundaries by propagating high-dimensional embeddings, such as learned queries. This preserves rich spatial and semantic information, avoids non-differentiable operations, and allows flexible network designs. Tracking can be implicit, since identity and motion cues are encoded in the latent space, but explicit association remains possible. Here, unification is achieved primarily through the structural combination of modules rather than the training process. Thus, with latent intermediate representations, not only information bottlenecks but cascading networks can be challenged, and tailored approaches, such as DeTra [164] be proposed.

VII-B Training Strategies

Critically, training unified perception is challenging, as learning must balance multiple tasks (detection, tracking, prediction) and objectives simultaneously. This complexity affects all approaches within our taxonomy and motivates the need for specialized training strategies. The overall performance thereby depends on balanced optimization across tasks, yet two core issues arise: tasks often progress at different learning speeds, requiring careful balancing, and task losses may be conflicting in their inherent goals, making loss function design and weighting crucial for overall performance [91]. These difficulties grow with the number of tasks to optimize [30, 31].

To address this, EUP, LUP, and FUP approaches adopt loss balancing strategies (see Tab.  II,  III,  IV). Simpler schemes sum task losses directly, but fail to resolve imbalance or conflict. More effective is empirical weighted learning, a standard in modular perception, where losses are combined with manually tuned weights. Extensions include sigmoid scheduling [119], which dynamically shifts task weights during training, and uncertainty-weighted balancing [112], which adapts weights based on task uncertainty to stabilize optimization. Beyond loss weighting, structural strategies can also improve training: DeTra [164], for example, employs selective gradient stopping between refinement stages, reducing task interference and unexpectedly boosting performance. Such controlled gradient flow illustrates how multi-task learning can benefit not only from careful weighting but also from tailored architectural choices.

VII-C Future Research Directions

Unified perception offers a promising extension to modular perception. However, for adoption in real-world applied systems, improvements in performance, robustness, and verifiability must be made. This section therefore outlines key research directions for unified perception in three perspectives: Evaluation and benchmarking, architecture and system design, and training strategies.

For one, research should focus on evaluation and benchmarking, to allow for comprehensive system performance statements:

**Unified vs. modular evaluation: ** A clear and systematic comparison between unified (EUP, LUP, FUP) and modular perception across different sensor modalities is still missing. Existing studies are limited in scope and do not provide a comprehensive overview of unified system performance [115]. Benchmarking is therefore required to quantitatively understand the trade-offs of unified perception.

**Closed-loop evaluation: ** Current evaluations rely on task-specific metrics that capture individual performance but not the overall utility of a perception system. Future research should focus on closed-loop evaluation, integrating the perception system with vehicle decision-making and control to reveal how modular, unified, or end-to-end approaches affect actual driving behavior.

Furthermore, the performance of unified perception should be robustified through architecture and system design development:

Cross-modality: Unified perception systems show modality biases—EUP is mostly image-based, while FUP relies primarily on LiDAR. Future work should develop multi-modal architectures that combine complementary sensor strengths for greater robustness and generalization. This includes advancing sensor fusion strategies and enabling flexible designs across different input modalities.

Implicit vs. explicit tracking: A key question is whether implicit tracking can match or surpass explicit methods under long occlusions and limited history [150]. Advancing implicit models with stronger temporal reasoning and memory is essential, particularly for LUP and FUP. This also raises broader questions about the role of object permanence in unified models.

Modularity through latent representations: A central challenge is whether modularity can be standardized within unified systems that rely on model-specific latent representations. While such representations are learned automatically, exploring standardized, object- or agent-centric formats could allow interchangeable submodules without breaking unified training. Approaches such as PF-Track [120], JLA [139], and ViP3D [141] suggest that modular and unified benefits may coexist if common latent interfaces can be established.

Map integration: While many unified perception methods leverage map information, it is often underutilized [115]. Future research should investigate deeper map integration (on-/offline) including dynamic maps, not only to improve prediction in LUP and FUP, but also to refine detection.

Object type generalization: Current unified models focus mainly on vehicles, pedestrians, and cyclists. For real-world deployment, they must also handle rare and diverse categories, including vulnerable users and temporary obstacles. This could be approached with broader formulations—such as movable vs. static, open-set, class-free, or even instance-free perception. Such approaches could share intermediate occupancy representations.

Open source and reproducibility: Progress is slowed by limited code release, especially for LUP and FUP. Open benchmarks and models are needed to enable fair comparison and cumulative progress. Based on the methods surveyed in this paper, closed-source contributions account for 31.1% in EUP, 85.7% in LUP, and 75.0% in FUP—underscoring the need for broader open-sourcing across unified approaches.

Safety, explainability, and verification: Unified perception complicates safety and verification due to its integrated design. Future work should develop tools and methods—both online (runtime monitoring) and offline (formal verification)—to ensure safe and reliable operation.

Lastly, the performance of unified perception should also be robustified through specific training strategies:

Multi-task learning: As discussed in Sec. VII-B, training unified perception across multiple tasks is challenging. Future research should develop scalable and efficient strategies to balance losses, optimize shared representations, and reduce negative task interference.

Learning paradigm: Due to the high cost of joint-task annotation, unified perception could benefit from semi- or unsupervised learning. Leveraging unlabeled sequences and self-supervised objectives can reduce reliance on large-scale curated datasets.

VIII Conclusion

With this survey, we provide a deep dive into unified perception, introducing a holistic taxonomy and formalizing three paradigms: Early, Late, and Full Unified Perception. Unified perception extends modular scene understanding pipelines by jointly addressing detection, tracking, and prediction within integrated system designs.

Our taxonomy categorizes methods along input, output, and intermediate representations, and distinguishes between different tracking formulations and cue types. In addition, we survey existing approaches, their architectures, training strategies, datasets used, and open-source availability, and highlight open challenges.

In summary, this work contributes the first holistic taxonomy of unified perception, positions existing methods within this design space, and provides a critical analysis of their system-level properties. By consolidating fragmented efforts, we clarify the principles underlying EUP, LUP, and FUP, and highlight the role of tracking, representation flow, and task integration. We hope this survey equips the community with the tools and perspective to advance the promising field of unified perception, foster progress on underexplored paradigms, and guide the development of architectures, training strategies, and benchmarks toward more robust, generalizable while interpretable perception systems.