Research On Person Re-Identification Technology Based On Deep Learning

The issue of social public safety has always been the focus and hot issue of people's attention.Video surveillance systems based on camera equipment have been widely used in various public places.The processing technology of public security surveillance video is crucial.It ensures that video surveillance can be used as an important auxiliary means to maintain public safety and play its role.Improving the processing capacity of surveillance video can effectively improve the efficiency of processing surveillance video.It also ensures the accuracy and real-time of information.Especially in the public security monitoring process,it plays an important role.Finding the target pedestrian is one of the purposes of processing surveillance video.The meaning of the person re-identification problem is to match the target pedestrian in the non-overlapping area of the surveillance cameras.In other words,it is determined whether the person photographed at different locations and at different times is the target pedestrian.With the development of deep learning,person re-identification based on deep learning has become the mainstream of person re-identification research.Person re-identification based on deep learning can improve the efficiency of person re-identification by computers and reduce costs.Based on the analysis of the current domestic and foreign research in the field of person re-identification,this thesis chooses to train the deep neural network DenseNet to perform the task of person re-identification.The main research contents and contributions of this thesis are as follows:1.The pre-stage of person re-identification is person detection.Person detection is the basis for person re-identification.Therefore,this thesis studies the theoretical basis and training method of Faster R-CNN,a target detection network based on deep learning,and applies the Faster R-CNN network to person detection.The coding realizes the person detection algorithm based on the Faster R-CNN network.2.This thesis studies the network structure and characteristics of the deep neural network DenseNet,studies the characteristics and advantages and disadvantages of three commonly used loss functions,then applies them to the DenseNet network.The problem with the commonly used loss function is that it cannot well reduce the intra-class spacing of the same pensor and expand the out-of-class spacing between different pensor.Therefore,this thesis proposes a composite loss function model based on three commonly used loss functions,and studies its characteristics.Experiments verify the improved DenseNet network based on the composite loss function model can effectively improve the accuracy of person re-identification.3.Aiming at the problem of the deep depth of the DenseNet network,which causes the low-level features of the image to be easily lost,this thesis studies the network structure and characteristics of the VGGNet-16 network,and uses the VGGNet-16 network to extract feature images of pedestrians.This can achieve the purpose of keeping the low-level features of the image well.This thesis combines the features of the VGGNet-16 network and the DenseNet network in parallel.It is verified by experiments that the parallel model of VGGNet-16 network and DenseNet network can improve the accuracy of person re-identification.4.Aiming at the problem of cross-view person re-identification images with different and more complex backgrounds,this thesis studies the attention mechanism algorithm.Attention mechanism algorithm can reduce the influence of different backgrounds and complex backgrounds.This thesis introduces the attention mechanism module into the deep learning network DenseNet.And through experiments,it is verified that the DenseNet person re-identification model with the attention mechanism can improve the accuracy of person re-identification.5.This thesis applies the compound loss function proposed in this thesis.This compound loss function and the attention mechanism are introduced into the parallel model of VGGNet-16 network and DenseNet network.Therefore,a re-identification model based on DenseNet network is studied.The modified person re-identification model based on DenseNet can reduce the intra-class distance of the same pedestrian and expand the out-of-class distance between different pedestrians.The model also retains the low-level features of pedestrian images,while reducing the influence of different backgrounds and complex backgrounds.Through a comparative experiment with the original unimproved DenseNet-based person re-identification model,it is verified that the DenseNet-based person re-identification model can significantly improve the accuracy of person re-identification.