Risk-Based Filtering of Valuable Driving Situations in the Waymo Open Motion Dataset

This repo contains situation example data for the paper https://hri-eu.github.io/RiskBasedFiltering. The example data can be found in the folder data.

For the full dataset of valuable and non-valuable driving situations, please contact: [email protected].

🧾 Data Format

baselines_xxx

These files contain the baseline methods for retrieving valuable vehicles in one scenario. The file includes the following keys:

• tracks_to_predict (baseline 1)
• objects_of_interest
• kalman_difficulties (baseline 2)

Value meanings:

• 0: not valuable
• 1: valuable
• -1: no vehicle for this entry

For kalman_difficulties, the actual Kalman difficulty values are stored. You can convert them to boolean using a threshold (e.g., value > 30 is 1, otherwise 0).

---

interactions_xxx

These files contain the risk model output for identifying interesting vehicles in a scenario. The file includes:

• interaction_first
• interaction_second
• all_states_id

interaction_first:

• 0: low risk with any other vehicle
• 1: high risk with one or more other vehicles (valuable first-order situation)

interaction_second:

• 0: low risk
• 1: the vehicle has high risk with another vehicle, and that other vehicle also has high risk with a third one (valuable second-order situation)

Example application for ML predictor:

✅ Recommendation: Start testing with the interaction_first results.
Use all vehicles with a value of 1 for training and evaluation of the ML predictor.

The data always includes the data name and scenario index of the corresponding Waymo data. Make sure to use the all_states_id array to map the array index of interaction_first to the actual vehicle ID in the Waymo data. Otherwise, you may select the wrong vehicles.

training_tfexample.tfrecord-00000-of-01000-current-0:

• training_tfexample.tfrecord-00000-of-01000: data name
• current-0: scenario index

---

risk_results_xxx

These files contain the actual risk values between one vehicle (ego vehicle index) and all other vehicles (other vehicle index). The values range between 0 and 1.

ℹ️ This file is mainly for debugging and likely not important for your task.

License

This dataset is licensed under the Open Data Commons Attribution License v1.0 (ODC-By 1.0).

You are free to use, modify, and share the data, including for commercial purposes, as long as you give proper attribution.

📚 BibTeX

If you use the data, please cite our work as follows:

@inproceedings{puphal2025,
    author    = {Puphal, Tim and Ramtekkar, Vipul and Nishimiya, Kenji},
    title     = {Risk-Based Filtering of Valuable Driving Situations in the Waymo Open Motion Dataset},
    booktitle = {IEEE International Automated Vehicle Validation Conference (IAVVC)},
    year      = {2025}
}