Evaluating the Robustness of LiDAR-based 3D Obstacles Detection and Its Impacts on Autonomous Driving Systems

Abstract: Autonomous driving systems (ADSs) require real-time input from multiple sensors to make time-sensitive decisions using deep neural networks. This makes the correctness of these decisions crucial to ADSs' adoption as errors can cause significant loss. Sensors such as LiDAR are sensitive to environmental changes and built-in inaccuracies and may fluctuate between frames. While there has been extensive work to test ADSs, it remains unclear whether current ADSs are robust against very subtle changes in LiDAR point cloud data. In this work, we study the impact of the built-in inaccuracies in LiDAR sensors on LiDAR-3D obstacle detection models to provide insight into how they can impact obstacle detection (i.e., robustness) and by extension trajectory prediction (i.e., how the robustness of obstacle detection would impact ADSs). We propose a framework SORBET, that applies subtle perturbations to LiDAR data, evaluates the robustness of LiDAR-3D obstacle detection, and assesses the impacts on the trajectory prediction module and ADSs. We applied SORBET to evaluate the robustness of five classic LiDAR-3D obstacle detection models, including one from an industry-grade Level 4 ADS (Baidu's Apollo). Furthermore, we studied how changes in the obstacle detection results would negatively impact trajectory prediction in a cascading fashion. Our evaluation highlights the importance of testing the robustness of LiDAR-3D obstacle detection models against subtle perturbations. We find that even very subtle changes in point cloud data (i.e., removing two points) may introduce a non-trivial decrease in the detection performance. Furthermore, such a negative impact will further propagate to other modules, and endanger the safety of ADSs.