Video Pedestrian Re-identification Based on Human-Machine Collaboration

The purpose of pedestrian re-identification is to find pedestrian targets from many real surveillance videos. It has multiple realistic application scenarios such as criminal search, intelligent security, and cross-camera tracking. However, the current stage of pedestrian re-identification faces many challenges, including blurring of surveillance video, changes in surveillance angles, changes in pedestrian posture, and pedestrian occlusion. Although the existing research methods have a good performance on some datasets, there is still a big gap from the practical application. Taking into account the advanced intelligence, flexibility, and adaptability of human beings in the face of complex problems and environmental changes, this paper attempts to establish a model based on human-machine collaboration, in order to improve the accuracy and enhance the practicability. Along this line, we first construct a new human-machine collaboration video pedestrian re-identification model (HMRM) and design the task allocation strategies and result fusion strategies accordingly. Then, based on the YOLOv3 object detection algorithm and the ResNet-50 network structure, we perform the pedestrian detection part and the pedestrian re-identification part respectively. Moreover, since HMRM allocates human-machine tasks based on overlapping thresholds, we do exploratory experiments on human workload, overlapping thresholds, and model accuracy. Finally, based on pyqt5 and multi-threading technology, a human-machine collaboration video pedestrian re-identification interface is developed. Experiments on the public datasets show that HMRM is superior to other pedestrian re-identification approaches.