Person Re-identification Method Based On Outdoor Surveillance Scene

Person re-identification is widely used in large public places as a video surveillance technology.Its main task is to locate target pedestrians through monitoring images captured by different cameras to prevent some behaviors that endanger social security.Since the images captured in actual scenes are often interfered by complex environments,it is a hot topic to find a person re-identification system that can accurately identify pedestrians in a monitoring environment.In this paper,the research focuses on the person re-identification system with local features,and in order to make the re-identification better applied to the actual scene,an end-to-end person reidentification system is designed.The main research work of this paper is as follows:(1)Aiming at the problem of the accuracy of person re-identification,such as illumination,pedestrian attitude and resolution,this paper proposes a multi-scale network combining local features and global features.At the same time,in order to effectively improve the discriminating ability of the re-identification network,a triplet loss was introduced,and the joint loss function is used for feature training.In additon,the re-ranking algorithm is used to improve the accuracy of multi-scale networks.In this paper,the comparison experiments are conducted on three datasets.The experimental results show that the algorithm can effectively improve the accuracy of person re-identification under the influence of various disturbances.(2)For the current person re-identification,most of them are identified on the pedestrian image by pedestrian detection,this paper considers actual needs,an end-toend person re-identification method for directly identifying pedestrians in actual scenes is proposed,which improving the entire Faster R-CNN pedestrian detection network framework,and combined with the multi-scale pedestrian recognition network realized in(1).At the same time,in order to make the network more convenient to output pedestrian candidate boxes,this paper introduces the SENet network to process the target pedestrian image features and scene picture features.Finally,the comparison experiments on the scene picture dataset show that the end-to-end pedestrian reidentification method has better recognition performance.(3)In order to make(2)the end-to-end person re-identification system to better applied to real life,this paper uses the pruning algorithm to accelerate the network model,and introduces hard pruning on the residual block of Res Net.Finally,the experimental results show that the pruning algorithm can make the end-to-end network maintain a certain accuracy while having a good recognition speed.