A Pedestrian Re-identification Algorithm Based On Deep Learning

With the development of economy and society,people's safety awareness is increasing.A large number of surveillance cameras are installed in public places,resulting in a large number of surveillance videos.At present,most of the surveillance video is processed manually,which not only consumes huge manpower and material resources,but also has low efficciency.Therefore,people have a strong demand for intelligent video processing technology.Intelligent video processing technology automatically analyzes and processes surveillance video by computer,which can free people from boring work of viewing surveillance video,reduce costs and improve efficiency.Person re-identification is a key technology in intelligent video processing technology.It refers to retrieving specific pedestrians in the pedestrian images captured by different cameras without overlapping visual angles.It is the premise of pedestrian tracking.lt is difficult to design a robust pedestrian feature descriptor artificially because of the complexity and variety of person re-identification scenes,different eamera types and large pedestrian attitude ehanges.With the rapid development of artificial intelligence,especially deep learning,researchers have gradually applied deep learning to the field of pedestrian re-identification.Convolutional neural network is an important model in deep learning.It has been successfully applied in the fields of face recognition and handwritten character recognition.This paper studies the pedestrian recognition based on convolutional neural network.Aiming at the problem of large intra-class distance caused by different types of cameras,different viewing angles and illumination changes in pedestrian re-identification,as well as the widespread background and pedestrian geometric deformation in pedestrian images,this paper mainly does the following work:1)A convolutional neural network model based on additive margin SoftMax loss was proposed to solve the problem of large intra-class distance in pedestrian re-identification.Compared with the traditional SoftMax loss,the additive margin SoftMax loss enlarges the loss by reducing the score of the input image,thus forcing the network to learn more discriminative features and reducing the intra-class distance.The feature extraction part of the model is mainly based on the ResNet50 network.Because the pedestrian re-identification data sets is small at present,Dropout layer is added to the original ResNet50 network.The Dropout layer temporarily discards neurons from the network with a certain probability,and sets their output to 0.When the next mini-batch,it may resume normal work.This is equivalent to training a new one every mini-batch.The network model can effectively reduce the risk of network overfitting and enhance the generalization ability of the network.Experiments show that title model can effectively reduce the intra-class distance.Compared with the benchmark model using traditional SoftMax loss,Rank1 has increased by 3.59 percentage points,Rank1 has reached 82.51%,and mAP has increased by 7.27 percentage points to 62.30%on the data set Market1501.2)Aiming at the problem of wide background and pedestrian geometry in pedestrian images,a pedestrian feature extraction network based on deformable convolution was proposed.In this paper,deformable convolution was used to improve the residual learning structure of AM-Net network,and an improved residual learning structure is formed.The sampling position of convolution core can be changed dynamically,so that the network can effectively deal with pedestrian geometric deformation and focus the attention of the network on pedestrians,thus reducing the influence of background on the performance of the model.The influence of the number of improved residual learning structures on the performance of the model was studied.When the number of improved residual learning structures is 2,the performance of the model is optimal.on the Marketl 501 dataset,Rank1 reaches 87.47%,mAP reached 68.15%,which was 4.96 percentage points higher and 5.85 percentage points higher than before the improvement.