FollowNet: A Comprehensive Benchmark for Car-Following Behavior Modeling

Source code for the following paper:

Chen, Xianda, et al. "FollowNet: A Comprehensive Benchmark for Car-Following Behavior Modeling." https://www.nature.com/articles/s41597-023-02718-7

📝Description

This notebook demonstrates how to achieve the car following models from traditonal models to data driven models. Motivation: given extracted car following events from five open datasets with the same data formate and train the car follow models. Author: Chen Xianda.

The extracted car following events are avaliable for download. Provide a tutorial of the data format and how to run the traditional models and the data-driven models.

🚕 Data

Extracted car-following events are stored in data/ folder. The colab tutorial takes the highD data for experiments first.

The datasets are HighD, SPMD(DAS1, DAS2), Waymo, Lyft, NGSIM. Each has its own training, validation and test part.

🛠 Quick Start

Run the colab notebook directly! Details are in the notebook below.

📚 Pretrained Models

Pretrained models are stored in trained_model/ folder.

📈 Dataset distribution

Below is the average time gap during car following (s). For more results stored in results/ folder.

📊 Evaluation Metrics

Collsion rate

MSE of spacing

📭Contact

meixin@ust.hk

xchen595@connect.hkust-gz.edu.cn

📎 References

If you use extracted car following data / FollowNet in your own work, please cite:

Chen, X., Zhu, M., Chen, K. et al. FollowNet: A Comprehensive Benchmark for Car-Following Behavior Modeling. Sci Data 10, 828 (2023). https://doi.org/10.1038/s41597-023-02718-7