Research And Implementation Of Person Re-identification Based On Deep Learning

It is extreme difficult to find specific pedestrians in the surveillance video by relying solely on human resources due to the popularity of video surveillance systems and the increasing number of surveillance videos,person re-identification technology came into being to solve this problem.Person re-identification technology refers to judging whether it is the same pedestrian by distinguishing the same pedestrians appearing under multiple cameras.Based on the above problems,the work that has been done of this thesis mainly includes the following aspects:1.First of all,a brand new algorithm named ISSE for image segmentation of person re-identification based on skeleton has been proposed,the 18 person skeletons are extracted by exploiting the OpenPose platform,and then the person images will be segmented into three parts according to skeletons extracted above,the three parts are a small amount of overlapped image content between each part,which can effectively enhance the image features so that solve the problem that the performance of extracting features is not good due to the low resolution of image.2.The input layer of the original Resnet50 convolutional neural network has been improved,and the input layer of the original Resnet50 residual network is changed from 224x224 to 64x64,so that the improved input layer of network can adapt to the dimension of segmented pedestrian image.The residual block of the original Resnet50 convolutional neural network also has been improved,including the convolution residual block and the identity residual block,which solves the problem of network degradation caused by the increase of the network layer.The loss function of the original Resnet50 convolutional neural network has been improved in purpose of solving the problems caused by parameter updating,and the performance of the original residual network has been optimized.The 1x1 and 3x3 convolutional kernels have been added into the original Renet50 residual network,so as to reduce the number of parameters in the network,which also makes the features extracted by each layer in the network much better.A new fully connected layer has been added to the original residual network Resnet50,which makes the network more flexible while enhancing network performance.3.The optimized convolutional neural network Resnet50 is trained on Market1501 data set and DukeMTMC-reID data set,the result indicates that the performance and identification accuracy of new network are much better than the original one,the improved neural network is competitive with other neural networks,even with the networks of international level.