Pedestrian Detection and Target Tracking Based on Person Re-identification in Crowded Crowd

As a high-level task of computer vision, pedestrian recognition and target tracking involves many application scenarios has great research significance and practical value. In this paper, we use the Repulsion Loss to solve the problem of the non-maximum suppression operation in the post-processing phase of target detection is sensitive to the threshold value in the crowded and obstructed pedestrian scenarios, which results in mis-judgement and missed detection of the target. In order to solve the undetected pedestrian problem in the improved YOLOV3, we added a deformable convolution to the Mask R-CNN algorithm, which can better meet and adapt to the scale changes of natural pedestrians. The experimental results show that the mAP of the improved YOLOV3 increases from 68% to 73.4%. The mAP of the improved Mask R-CNN increased from 85% to 87.5%. We also present a pedestrian recognition GCPRN, to overcome the problem of insufficient representation of pedestrian appearance features in deep sort target tracking algorithm, we proposed a target tracking FRTN network. The experimental results indicated that the FRTN network is more effective than Deep Sort algorithm on MOTP and MOTA.