PCP-MV

Official project website for the ICRA 2025 paper "Learning better representations for crowded pedestrians in offboard LiDAR-camera 3D tracking-by-detection", which studies the representation learning problem for 3D multiple object tracking using point clouds and images as inputs, especially for environments with crowded pedestrians.

[Poster]

Environment

Before you start, please refer to ENV.md to build this project.

Data preparation

Please follow DATASET.md to download and prepare the nuScenes dataset.

Usage: inference demo

Refer to INFERENCE.md to perform LiDAR-camera tracking-by-detection and quantitative evaluation.

Usage: training experiments

Refer to TRAIN.md to perform training with various configurations.

License

A MIT license is used for this repository. Third-party datasets like nuScenes are subject to their own licenses and the user should obey them strictly.

Acknowledgement

This repository is developed based on BEVFusion. Thank the authors for their contributions.

Citation

Please star this repository and cite the following paper in your publications if it helps your research:

@INPROCEEDINGS{11128508,
  author={Li, Shichao and Li, Peiliang and Lian, Qing and Yun, Peng and Chen, Xiaozhi},
  booktitle={2025 IEEE International Conference on Robotics and Automation (ICRA)}, 
  title={Learning Better Representations for Crowded Pedestrians in Offboard LiDAR-Camera 3D Tracking-by-detection}, 
  year={2025},
  volume={},
  number={},
  pages={2740-2747},
  keywords={Point cloud compression;Training;Pedestrians;Three-dimensional displays;Urban areas;Benchmark testing;Semisupervised learning;Trajectory;Robotics and automation;System analysis and design},
  doi={10.1109/ICRA55743.2025.11128508}
  }

Link to the paper