@inproceedings {pub5584,
	title = {Risk-Based Filtering of Valuable Driving Situations in the Waymo Open Dataset},
	author = {Tim Puphal AND Vipul Ramtekkar AND Kenji Nishimiya},
	year = {2025},
	month = {May},
	abstract = {Ensuring the safety of Automated Vehicles (AVs) requires extensive testing in a variety of challenging driving situations. However, most real-world driving data consists of routine, low-risk scenarios that offer limited value for AV testing. To address this, we propose a novel risk-based filtering approach that can automatically select valuable driving situations for testing AVs. Specifically, we use a probabilistic collision risk model to identify high-interest driving situations within the Waymo Open Dataset. Our approach stands out by considering both first-order vehicle interactions (where one vehicle directly influences another and inducing risk) and second-order vehicle interactions (where influence or risk propagates through an intermediary vehicle). We evaluate our method against established baselines, such as Kalman difficulty, and the metrics from the Waymo Dataset - Tracks-To-Predict (TTP) and Objects-of-Interest (OOI). Our results show that our model identifies complex, complementary driving situations that can lead to significant behavioral changes in vehicles. Notably, our approach highlights valuable multi-vehicle interactions that would be missed by Kalman difficulty, TTP and OOI. To support further research, we will open-source the outputs of the risk model and scenario IDs from our filtered dataset. This work advances AV validation by ensuring that evaluation datasets contain diverse and safety-critical driving scenarios.},
	publisher = {IEEE},
	booktitle = {IEEE International Automated Vehicle Validation Conference (IAVVC)}
}