Object Relevance for Radar Object Detection on Autonomous Driving Scenarios

Ensuring safety for driving assistance (DA) systems is a crucial aspect that must be taken into account to prevent hazardous situations and undesirable behaviors. Given that many DA systems heavily rely on object detection, this paper specifically focuses on object detection on radar point clouds and examines it from a safety perspective. More concretely, we propose a weighting scheme during training by modeling object relevance to bias the detection model towards prioritizing safety critical objects. We propose criticality functions, derived from physical assumptions regarding distance and velocity, to compute object relevance. We provide evidence of the effectiveness of the proposed relevance weighting with numerical experiments conducted on the nuScenes dataset. The results show improvements in both mAP and IoU measures compared to a baseline approach.