Abstract
Reliable detection of nearby vehicles is critical for autonomous driving. Problems thus arise when a vehicle is occluded and becomes undetectable, which may lead to accidents. In this paper, we develop mmOVD+, a millimeter-wave radar-based system that enables fully occluded vehicle detection by leveraging ground-reflected signals that propagate beneath the blocking vehicle's chassis. To eliminate ghost points caused by frequent multi-path reflections, we devise a set of neighborhood behavior-based point features and integrate them into GhostNet, a customized deep network that enables accurate point classification through combined feature learning. In addition, we propose a cumulative clustering algorithm to effectively aggregate the unstable ground-reflected radar points over consecutive frames to derive the bounding boxes of the vehicles. We have evaluated mmOVD+ in both controlled and real-world environments. In an underground garage and two campus roads, we conducted controlled experiments in 70 scenes with 9 vehicles, including a minibus, a motorcycle, and an electric moped. Our system accurately detects occluded vehicles, with a 92.7% F1 score for occluded vehicle detection and a 100% success rate for occlusion event detection. More importantly, we drove 324 km on crowded roads at a speed up to 70 km/h and show we could achieve a 93.8% occlusion detection success rate and a 2.4% false alarm rate with only 10% of the training data.
Funding Information
  • National Natural Science Foundation of China (62332016)
  • Key Research Program of Frontier Science
  • Chinese Academy of Sciences (ZDBS-LY-JSC001)

This publication has 43 references indexed in Scilit: