# Dynamic Feature Fusion for Sparse Radar Detection: Motion-Centric BEV Learning with Adaptive Task Balancing

## TL;DR

This paper introduces a new radar detection framework for autonomous driving that improves object detection by focusing on motion patterns and efficiently balancing tasks.

## Contribution

The novel framework introduces motion-aware fusion, adaptive task balancing, and progressive training for sparse radar detection.

## Key findings

- The method achieves 33.25% mean Average Precision (mAP)3D with minimal parameter overhead.
- It shows a 4.16% improvement in pedestrian detection while maintaining real-time performance.
- Ablation studies confirm the effectiveness of each component in motion pattern learning.

## Abstract

This paper proposes a novel motion-aware framework to address key challenges in 4D millimeter-wave radar detection for autonomous driving. While existing methods struggle with sparse point clouds and dynamic object characterization, our approach introduces three key innovations: (1) A Bird’s Eye View (BEV) fusion network incorporating velocity vector decomposition and dynamic gating mechanisms, effectively encoding motion patterns through independent XY-component convolutions; (2) a gradient-aware multi-task balancing scheme using learnable uncertainty parameters and dynamic weight normalization, resolving optimization conflicts between classification and regression tasks; and (3) a two-phase progressive training strategy combining multi-frame pre-training with sparse single-frame refinement. Evaluated on the TJ4D benchmark, our method achieves 33.25% mean Average Precision (mAP)3D with minimal parameter overhead (1.73 M), showing particular superiority in pedestrian detection (+4.16% AP) while maintaining real-time performance (24.4 FPS on embedded platforms). Comprehensive ablation studies validate each component’s contribution, with thermal map visualization demonstrating effective motion pattern learning. This work advances robust perception under challenging conditions through principled motion modeling and efficient architecture design.