Learning Collision Risk Proactively from Naturalistic Driving at Scale

This study is being submitted and under review. A preprint is provided at arXiv. Questions, suggestions, comments, and collaborations are welcome. Please feel free to reach out.

Discussions open to everyone

We enable GitHub Discussions for this repository, where you are welcome to ask questions, share insights, and discuss the content of the paper or future research. We encourage everyone to participate in the discussions, as it helps foster a collaborative environment for learning and improvement.

Access to trajectory data of crashes and near-crashes

Collaborated with Virginia Tech Transportation Institute (VITTI), we have made the trajectory reconstruction dataset of naturalistic crashes and near-crashes in SHRP2 NDS accessible. You are welcome to refer to BirdsEyeTrajectoryReconstructionSHRP2NDS for more information and guidelines to use.

Highlights

• Collision risk is learnt from naturalistic interactions without crash or near-crash labels.
• Context-conditioned distributions of multi-directional spacing characterise interactions.
• Deviations from typical safe spacing towards closer extremes are quantified as risk.
• Outperformance over existing methods is validated on 2,591 real-world (near-)crashes.
• Environmental and historical kinematic features provide performance enhancement.
• Spacing direction, road-surface condition, and past kinematics are main risk factors.

This work enables context-aware, scalable, and generalisable learning of collision risk from everyday interactions. I believe it is opening a window to foundational models for proactive risk quantification of potential collisions. This hopefully will facilitate research in safe autonomous driving and traffic safety analytics.

In order to repeat the experiments

Below we offer a step-by-step workflow to repeat the experiments in the paper. On the top of each script, we provide a brief description of its purpose. These scripts are designed to be run in sequence, and each script outputs necessary files for the next one.

Quick Navigation:

• 1 Dependencies
• 2 Data
• 3 Bird's eye trajectory reconstruction
• 4 Training data preparation
• 5 Posterior inference
• 6 Test data preparation and first-stage evaluation
• 7 Second-stage evaluation and result analysis

1 Dependencies

pandas, pytables, tqdm, numpy, matplotlib, torch, torchvision, scikit-learn, scipy, see more detailed dependencies in requirements.txt.

2 Data

Three datasets are used in this study.

• SHRP2 NDS:
  Two options are available depending on whether you are interested in trajectory reconstruction from the original event data.

  • If you would like to skip trajectory reconstruction:
    Download the dataset following the guidlines at https://github.com/Yiru-Jiao/BirdsEyeTrajectoryReconstructionSHRP2NDS. Put the files in ReconstructedTrajectories.zip under the directory ./ProcessedData/SHRP2/. Put the files in SafetyCriticalTestSet.zip under the directory ./ResultData/EventData/.
  • If you would like to reconstruct the trajectories:
    Apply for the original datasets
    • E. Sears, M. A. Perez, K. Dan, T. Shimamiya, T. Hashimoto, M. Kimura, S. Yamada, and T. Seo. A Study on the Factors That Affect the Occurrence of Crashes and Near-Crashes. Version V2. 2019. URL: https://doi.org/10.15787/VTT1/FQLUWZ
    • C. K. Layman, M. A. Perez, T. Sugino, and J. Eggert. Research of Driver Assistant System. Version V3. 2019. URL: https://doi.org/10.15787/VTT1/DEDACT
      and put them under the directory ./RawData/SHRP2/.

• highD:
  Download the dataset from https://www.highd-dataset.com/ to the directory ./RawData/highD/.

• ArgoverseHV:
  Download the dataset following the guidlines at https://github.com/RomainLITUD/conflict_resolution_dataset. Put the files under the directory ./RawData/Argoverse2/.

Resulting data such as trained models, loss logs, and evaluation results are compressed as two zip files of the ./PreparedData/ and ./ResultData/ folders. They can be downloaded from https://doi.org/10.4121/9caa1e6c-9abd-4e36-ae28-c9ea4542d940.

3 Bird's eye trajectory reconstruction

Run the following scripts in order to reconstruct the trajectories of the events in SHRP2 NDS.

• Transform .xlsx to .csv: ./src_trajectory_reconstruction/transform_files.py
• Summarise meta data: ./src_trajectory_reconstruction/organise_metadata.py
• Search for EKF parameters: ./src_trajectory_reconstruction/search_ekf_parameter.py
• Reconstruct trajectories: ./src_trajectory_reconstruction/reconstruct_birdseye.py
• Visualise reconstructed trajectories: ./src_trajectory_reconstruction/event_visualiser.ipynb (This is optional in case you would like to check the reconstructed events.)

4 Training data preparation

Run the following scripts in order to prepare the training data. The SHRP2 SafeBaseline data are processed in the previous step, here we prepare the highD and ArgoverseHV data. Then we segment the data into samples, and split each dataset into a train set (80%) and a val set (20%) for GSSM training.

• Extract lane-changes in highD: ./src_data_preparation/prepare_highD.py
• Filter HV-HV crossings and turnings: ./src_data_preparation/prepare_argoverse.py
• Segment samples: ./src_data_preparation/segment_datasets.py

5 Posterior inference

Run the following script in order to train the GSSMs. The trained models and loss logs will be saved for later use. You can run this script for multiple times. It will skip trained models and continue training until all models are trained.

• GSSM training: ./src_posterior_inference/pi_train_eval.py

6 Test data preparation and first-stage evaluation

Run the following scripts in order to prepare the test set and implement the first-stage evaluation.

• Prepare test set: ./src_safety_evaluation/organise_events.py (This can be skipped if you have downloaded the reconstructed trajectories and put SafetyCriticalTestSet.zip under the directory ./ResultData/EventData/.)
• Apply GSSMs to 4,875 events in the test set: ./src_safety_evaluation/evaluate_safety.py
• Initially evaluate warning at different thresholds: ./src_safety_evaluation/analyse_events.py

7 Second-stage evaluation and result analysis

Run the following scripts in order to implement the second-stage evaluation and analyse the results.

• Vote for conflicting objects: ./src_safety_evaluation/vote_conflicting_target.py
• Re-evaluate warnings for 2,591 events that have conflicting objects determined: ./src_safety_evaluation/risk_evaluation.py
• Attribute risk to contextual representations: ./src_safety_evaluation/attribute_intensity.py

8 Visualise results

Use the following notebook to reproduce the figures and tables in the paper.

• Visualise results: ./src_visualisation/figure_table_to_reproduce.ipynb

Copyright

Citation

@article{jiao2025gssm,
    title = {Learning Collision Risk from Naturalistic Driving with Generalised Surrogate Safety Measures},
    author = {Yiru Jiao and Simeon C. Calvert and Sander {van Cranenburgh} and Hans {van Lint}},
    year = {2025},
    journal = {arXiv preprint},
    pages = {arXiv:2505.13556}
}

License

This repository is licensed under the MIT License. You are free to use, modify, and distribute the code, but please retain the original copyright notice and license in any copies or substantial portions of the software.

Repo references

Thanks to GitHub for offering the open environment, from which this work reuses/learns/adapts the following repositories to different extents:

• Multidimensional Kalman-Filter
  • https://github.com/balzer82/Kalman/blob/master/Extended-Kalman-Filter-CTRA.ipynb
  • https://github.com/balzer82/Kalman/blob/master/Extended-Kalman-Filter-CHCV.ipynb
• Two-dimensional traffic safety evaluation
  • SSMsOnPlane https://github.com/Yiru-Jiao/SSMsOnPlane
  • EmergencyIndex https://github.com/AutoChengh/EmergencyIndex

We are grateful for the authors' contributions to open science and reproducible research.