🤖 AI Summary
Current 3D object detectors for autonomous driving lack quantifiable safety evaluation metrics, hindering rigorous assessment of collision risk. Method: This paper proposes Uncompromisable Spatial Constraints (USC), requiring detection bounding boxes to strictly cover ground-truth objects in both front-view and bird’s-eye view (BEV) projections—thereby explicitly mitigating collision risk. We formalize rigid spatial coverage as a safety-oriented, differentiable metric grounded in perspective projection and BEV geometry, and—uniquely—integrate it directly into the training loss function to enable safety-driven fine-tuning. Experiments are conducted on nuScenes, with validation via closed-loop simulation. Contribution/Results: The USC-aware detector achieves significantly improved detection robustness and a marked reduction in closed-loop collision rate, empirically demonstrating that perception-level safety metrics exert a direct, positive impact on end-to-end system safety.
📝 Abstract
In this work, we consider the safety-oriented performance of 3D object detectors in autonomous driving contexts. Specifically, despite impressive results shown by the mass literature, developers often find it hard to ensure the safe deployment of these learning-based perception models. Attributing the challenge to the lack of safety-oriented metrics, we hereby present uncompromising spatial constraints (USC), which characterize a simple yet important localization requirement demanding the predictions to fully cover the objects when seen from the autonomous vehicle. The constraints, as we formulate using the perspective and bird's-eye views, can be naturally reflected by quantitative measures, such that having an object detector with a higher score implies a lower risk of collision. Finally, beyond model evaluation, we incorporate the quantitative measures into common loss functions to enable safety-oriented fine-tuning for existing models. With experiments using the nuScenes dataset and a closed-loop simulation, our work demonstrates such considerations of safety notions at the perception level not only improve model performances beyond accuracy but also allow for a more direct linkage to actual system safety.
