Adaptive shrinkage non-maximum suppression for occluded pedestrian detection
Vision-based pedestrian detection is an important research field in object detection.The outputs of current main-stream anchor-based detectors are redundancy,and the non-maximum suppression(NMS)is necessary before output predictions.Therefore,the effectiveness of NMS will influence detectors performance directly.The biggest challenge in pedestrian detection is the occlusion issue.Heavily overlapping objects make it difficult for NMS to balance between high recall rate and low false positive rate.An adaptive shrinkage NMS(AS-NMS)is proposed to reduce the overlapping degree of occluded objects'boxes by adaptively shrinking the corresponding prediction boxes.Implementing NMS to the shrunken prediction boxes can avoid the impact of occlusion issue.In addition,ambiguous prediction boxes will have negative impact toward AS-NMS,and a center point repulsion loss(CPR loss)is proposed to reduce the numbers of ambiguous boxes hovering between two objects by implement a repulsion force between the center points of overlapping boxes.The experimental results demonstrate that the proposed algorithm can significantly boost the performance of anchor-based detectors towards occluded pedestrian detection.
NETL
NSTL
万方数据
