A Pedestrian Object Detection Method from Point Cloud Based on Skeleton-Aware Network

This work addresses the problem of low accuracy in the object detection of pedestrians from point cloud in road scenes, a detection method based on human skeleton-aware network is proposed. The whole framework is based on the two-stage object detection method PointRCNN. The stage-1 sub-network uses DGCNN to segment the foreground points of the whole scene point cloud, and generates 3D proposals from foreground points. The stage-2 sub-network applies the topology and the prior information of the human body model to pre-train a skeleton-aware network to capture the body shape and posture information hidden in the human point cloud and estimates the parameters of posture for accurate box refinement. Experiments show that our 3D detection method achieves better performance than other classic methods on pedestrian object detection.