Misdetection Risk-Aware Adaptive LiDAR Sensing for Automotive Driving

Abstract

Automotive LiDARs typically have a uniform scanning range over their field of view (FoV). Such a range profile does not account for the varying risk of misdetecting targets in different regions. For instance, prioritizing crosswalks in a LiDAR scan is crucial, as the financial consequences of missing a pedestrian far exceed that of overlooking a distant vehicle. In this paper, we construct a spatial risk map that quantifies the risk of misdetecting targets across different regions around the vehicle. Our risk map incorporates lane semantics, knowledge about previously identified objects, and their potential trajectories. We use this risk map to adapt the LiDAR's scanning range over different sectors in its FoV. Simulations on nuScenes episodes demonstrate that our misdetection risk-aware design reduces the effective risk by about 40% compared to a standard LiDAR.