# LAT-BirdDrone: A dedicated dataset for high-precision classification of low-altitude small target trajectories enhanced by hybrid neural networks

## TL;DR

The paper introduces LAT-BirdDrone, a new dataset for classifying trajectories of small flying targets like birds and drones, using advanced detection systems and hybrid neural networks.

## Contribution

The novel contribution is the creation of a dedicated dataset for low-altitude micro-target trajectory classification, addressing a gap in existing datasets.

## Key findings

- LAT-BirdDrone contains 33,262 image samples and 665 trajectory data entries covering birds and drones.
- The dataset supports evaluation of CNN, Transformer, and hybrid models for trajectory classification.
- It provides resources for optimizing optoelectronic sensors in low-altitude monitoring.

## Abstract

Currently, most existing datasets predominantly focus on object classification based on appearance, while datasets specifically designed for trajectory classification of low-altitude flying micro-targets (11pt × 11pt or smaller) remain scarce. There is also an urgent need for a dedicated dataset to provide comprehensive data support for verifying the impact of the CNN module on trajectory classification models. To address these gaps, this research utilized the Three-band refrigeration photoelectric turntable Z50IV-CTVC690110–21,100—an advanced long-distance optoelectronic detection system independently developed by Hepu Vision—to collect data. The data collection covered multiple videos, including visible light and infrared footage of birds and drones. For data processing, the YOLOv11n algorithm was applied for object detection and trajectory extraction, and the ByteTrack algorithm was further employed to enhance the accuracy and reliability of target trajectory extraction. This process resulted in the establishment of LAT-BirdDrone, a dedicated dataset for low-altitude micro-target trajectory classification. The dataset contains 33,262 image samples and 665 pieces of detailed trajectory data, comprehensively covering the dynamic motion patterns of low-altitude flying micro-targets (11pt × 11pt or smaller) under various lighting, weather, and complex background conditions. In addition, the dataset comprises 65 videos (15 drone videos and 50 bird videos) with an average trajectory length of 48 frames. It is important to note that due to the complexity and cost of data collection, the dataset is relatively limited in scale, which may restrict the training and evaluation of certain complex models. Nevertheless, LAT-BirdDrone provides valuable data support for research on low-altitude micro-target trajectory classification and lays the foundation for the construction of larger datasets in the future. In terms of reuse potential, LAT-BirdDrone fills the scarcity of datasets for trajectory classification of low-altitude flying micro-targets (11pt × 11pt or smaller). It can provide essential data support for verifying the impact of the CNN module on trajectory classification models, and be used to evaluate the trajectory classification performance of basic models such as CNN, Transformer, iTransformer, and LSTM, as well as support research on multi-modal hybrid architectures (e.g., Transformer+LSTM, CNN+BiLSTM) that fuse local spatial features with global temporal dependencies. Additionally, the dataset offers resources for studies on micro-targets in low-altitude security applications and references for optimizing the application of optoelectronic sensors in low-altitude micro-target monitoring scenarios.