Research On Person Re-identification Based On Surveillance Video

With the rapid development of image processing and deep learning theory and technology,video surveillance systems have been used more and more widely,so a large amount of data has also been generated,but at present these data have not been fully and effectively utilized,thus causing huge waste of the resources.In order to improve the application efficiency and level of these resources,this paper is based on monitoring video and carries out the research of person re-identification based on deep learning theory and method,which has important theoretical significance and practical value for enhancing the performance of existing video surveillance systems.Person re-identification refers to the use of a computer to determine whether pedestrians appearing in different cameras belong to the same person.It generally includes two parts: feature extraction and distance metric.For the problems existing in the current two parts research,on the basis of deep learning,this paper first improves the feature extraction effect through network structure design,and simultaneously optimizes the loss function in the distance metric.Specifically,this article has done the following work:Firstly,based on Convolutional Neural Networks(CNN),this paper constructs a multi-branch pedestrian feature extraction network structure.By classifying the feature map horizontally,a targeted local feature extraction branch is constructed,and a multi-branch feature fusion network is formed by the combination with the global feature extraction branch,which gained more discriminating pedestrian feature.Secondly,the effects of different loss functions in the person re-identification task are analyzed.Based on the triplet loss function,we proposed an adaptive triplet loss function and proved it's effectiveness by experiments.Finally,for the problem that the pedestrian appearance and color are similar and difficult to identify in the surveillance video,this paper combines the multi-branch feature extraction network,the optimized loss function and the spatial attention mechanism,and proposes a person re-identification algorithm based on surveillance video.The end-to-end design and implementation of the network model is completed,and the performance of the proposed algorithm is verified by testing on several public data sets.