End-to-End Pedestrian Detection And Person Re-Identification Based On Feature Fusion

With the increasing public security issues,laying video surveillance networks in public places has become an effective means of public safety prevention.These monitoring devices generate massive surveillance videos every day.How to efficiently capture the targets of interest(especially pedestrians)in massive surveillance videos has become an urgent problem to be solved.Therefore,the need to use computer vision technology to intelligently re-identify pedestrians in surveillance video is becoming more and more urgent.However,most of the person re-identification research is based on the already cut pedestrian image block.The cropped pedestrian image block contains only the pedestrian image and does not contain the background information.Such a scenario obviously does not match the actual situation.Because the actual surveillance image contains not only the pedestrians but also the complex background environment,the position in the pedestrian re-image cannot be directly given.Therefore,pedestrian detection is required in the surveillance image before the pedestrian re-identifies.With the deep exploration of deep learning in the field of computer vision,more and more researchers apply deep learning to pedestrian detection and person re-identification.In this paper,we introduce LOMO features on the basis of deep features,and combine this traditional hand-designed pedestrian features with deep features to construct a robust pedestrian feature.The LOMO feature can well suppress the effects of different angles and different illuminations on pedestrian detection and person re-identification in the actual environment.In order to improve the applicability of person re-identification in practical scenarios,this paper designs an end-to-end pedestrian detection and person re-identification system based on feature fusion.The system includes pedestrian detection network and person identity re-identification network,and has good performance.The ResNet-50 network builds pedestrian detection and person re-identification systems.Training and optimization of the network model is carried out on the PyTorch deep learning framework.Learning by migration based on the pre-training model.Experiments show that the end-to-end pedestrian detection and person re-recognition system based on feature fusion in this method can extract more robust pedestrian characteristics,and effectively improve the accuracy and feasibility of pedestrian detection and person re-recognition in actual scenarios.