Pretraining on Synthetic Driving Data for Trajectory Prediction

Yiheng Li, Seth Z. Zhao, Chenfeng Xu, Chen Tang, Chenran Li, Mingyu Ding, Masayoshi Tomizuka, Wei Zhan

Official Repo for IROS 2024 paper Pre-training on Synthetic Driving Data for Trajectory Prediction (https://arxiv.org/pdf/2309.10121).

Repo Structure

In this repo, we present the generated dataset we built for pre-training, as well as our pre-training and fine-tuning pipeline with combined reconstruction setting. Note that the trajectory reconstruction and map reconstruction can be realized easily by minimal modification to the presented version. The pipeline is illustrated below:

• Pretrain_Data_360k_Aug_Orig contains the pre-training data we generated on both original and augmented maps. Those motion data files on augmented map are started with augmap.
• src contains the source code for pre-training and fine-tuning.
• pretrain contains the pre-trained models, visualizations and source code we used when doing our experiments.
• finetune contains the fine-tuned models, source codes, and evaluation logs we used when doing our experiments.

Usage

• 1. Please first refer to Quick Start part in denseTNT repo (https://github.com/Tsinghua-MARS-Lab/DenseTNT) for preparing environment.
• 2. Please refer to bash where we provide bash commands for using the code.

Performances

In our experiments, we see the benefits of our pipeline as followed:

Method	$MR_6$(%)	$minFDE_6$	$minADE_6$
Baseline	9.73	1.0673	0.8052
Map Reconstruction	9.20 (-5.45%)	1.0343 (-3.09%)	0.7571 (-5.97%)
Trajectory Reconstruction	9.24 (-5.04%)	1.0263 (-3.84%)	0.7384 (-8.30%)
Combined Reconstruction	9.27 (-4.73%)	1.0349 (-3.04%)	0.7284 (-9.54%)

Where the checkpoints of combined reconstruction pre-training has been provided in this repo.

Acknowledgement

This implementation is based on DenseTNT (https://github.com/Tsinghua-MARS-Lab/DenseTNT). This work was supported by Berkeley DeepDrive.

Citation

@INPROCEEDINGS{li2024pretrain,
  author={Li, Yiheng and Zhao, Seth Z. and Xu, Chenfeng and Tang, Chen and Li, Chenran and Ding, Mingyu and Tomizuka, Masayoshi and Zhan, Wei},
  booktitle={2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}, 
  title={Pre-training on Synthetic Driving Data for Trajectory Prediction}, 
  year={2024},
  volume={},
  number={},
  pages={5910-5917},
  keywords={Codes;Pipelines;Predictive models;Data collection;Data models;Vectors;Trajectory;Forecasting;Intelligent robots;Synthetic data},
  doi={10.1109/IROS58592.2024.10802492}
}