BHunter: Generating Blind-Spot Driving Scenarios for Robustness Testing of Perception Systems

Abstract

For widespread deployment of autonomous driving systems (ADS), perception robustness plays a fundamental role. However, maintaining robust perception performance remains a significant challenge, especially under occlusion-rich conditions, known as blind-spot driving scenarios. In these scenarios, traffic participants are temporarily occluded by surrounding objects, making detection and tracking significantly more difficult. However, existing datasets lack sufficient coverage of such scenarios, hindering comprehensive evaluation and improvement of perception capabilities under such conditions. In light of this, we systematically model blind-spot scenarios for the first time and develop a corresponding framework, BHunter , for generating perception test cases. Specifically, this framework enables controllable synthesis of sensor data corresponding to these scenarios, such as generating data with specified occlusion levels, to effectively test perception models. With the help of BHunter , we construct a dataset of 2,000 blind-spot driving scenarios, covering five occlusion levels with a balanced distribution. Each scenario includes synchronized multimodal sensor data, accurate 3D annotations, and temporally consistent frames, and can be directly used for training and evaluating state-of-the-art perception models. To demonstrate the effectiveness of our benchmark, we evaluate widely used detection and tracking models across various sensor configurations. The results show that the models retrained on our benchmark achieve improvements of up to 30.74 and 36.96 percentage points in detection and tracking tasks, respectively, under high-occlusion conditions. These results confirm that our benchmark fills a critical gap in the perception testing landscape, offering a valuable foundation for improving ADS safety in blind-spot scenarios.