Person Re-identification Based On Improved IDE Network

In order to solve the problem of social security in time and effectively,video surveillance systems have been set up in some large public places to record the behavior information or locus of the dense crowd,which plays an important role in maintaining social order and public security.Person re-identification is to judge whether the persons detected by image data or video data from different cameras are the same person.Person re-identification technology has not only become an important means for law enforcement departments to manage public safety and crime detection,but also can calculate the flow data of public places,so as to improve the design of transportation system or optimize the layout of shopping malls.IDE network is a very important baseline for person re-identification.It regards each person as a category of classification problems by feature description method.But in practical application,IDE network has the problem that extracting features can not adapt to complex environment changes and the ability of generalization of global features is weak.Therefore,aiming at the shortcomings of IDE network,three aspects are improved.In the backbone network,DenseNet and PCB networks are used to extract features of input pictures instead of ResNet.In data augmentation,in order to adapt the model to the changes of complex environment,color jitter and Random Erasing are used to expand the data set.In terms of network structure,two different attention mechanisms,spatial attention and channel attention,are introduced to merge them with IDE network.After experimental analysis,the following conclusions are drawn:(1)DenseNet improved backbone network is superior to ResNet in feature extraction and can prevent model from over-fitting;PCB network improved backbone network can segment feature map horizontally,so that the model can learn the local features of different areas of each pedestrian.DenseNet’s improved IDE network and PCB network’s improved IDE network have higher rank-1 of 1.66% and 3.18% than the original IDE network respectively;(2)color jitter can make the model adapt to the complex background of different lighting conditions,while Random Erasing can overcome the partial occlusion problem of persons and enhance the generalization ability of the model.The rank-1 of improved IDE network with color jitter and Random Erasing is 1.57% and 2.40% higher than that of original IDE networkrespectively;(3)Spatial attention mechanism and channel attention mechanism can obtain more robust features through spatial conversion and re-calibration of features.The rank-1 of the improved IDE network fusing spatial attention and the improved IDE network fusing channel attention is 2.20% and 1.10% higher than that of the original IDE network,respectively.