Research On Person Re-identification Algorithm Based On Deep Convolution Neural Network

Pedestrian re-identification(Re-ID)is a technology that can track,match and identify target pedestrians across time and space dimensions.Re-ID has extremely high application value in the field of criminal investigation,and plays an increasingly important role in combating crime and security.It can also calculate the flow of people in public places to achieve the purpose of designing and optimizing the traffic system.In addition,the use of Re-ID technology can obtain the movement trajectory of customers when shopping,helping business operators to analyze the true needs of customers,thereby further digging for commercial value.However,in the practical application of Re-ID technology,there are still many limitations,such as the number of samples in the data set is insufficient,low image resolution,varying illumination intensity,diverse pedestrian poses,and complex backgrounds.All these greatly affect the recognition accuracy of the model.How to solve these problems effectively is still a challenge for Re-ID.Deep learning has developed rapidly in recent years.How to use intelligent technology to solve the aforementioned problems,achieve efficient processing and analysis of massive video data,extract valuable information from it,and identify targets in video has become a hot research issue.In this dissertation,we will research Re-ID based on deep convolutional neural network.The main research work includes the following parts:(1)When using the traditional Softmax classification loss,the in-class distance and the inter-class distance of the sample will be ignored,which will affect the effect of Re-ID.To solve this problem,we applied the ranked list metric learning loss to the field of Re-ID for the first time,and combined it with the Softmax classification loss to design a composite loss.The combination loss function designed by us can not only learn a good metric method to ensure that the distance between the same categories is small enough and the distance between different categories is large enough,but also can obtain the distinguishing pedestrian characteristics through network model learning.The metric learning loss function and the traditional classification loss function have their own characteristics.They can complement each other when they work together as the combined loss function.Experimental results on multiple Re-ID datasets show that the deep metric learning algorithm based on the combined loss function has obvious advantages over other algorithms and can effectively improve the performance of Re-ID.(2)The methods based on deep learning rely on a large number of training data,so the performance of the model is closely related to the dataset used to train Re-ID model.But at present,many public dataset of Re-ID is small,the number of samples in a single category is scanty,and the diversity of pedestrian pose is insufficient.This leads to the Re-ID model cannot adapt to complex background changes.To deal with these problems,this dissertation designed a pedestrian image style transfer model for Re-ID specifically.It can automatically transform the style of the images in the original data set.Then the images generated after the style conversion and the original images constitute a new training set.The model we designed can effectively expand the data set,increase the diversity of samples.It also helps to learn discriminative pedestrian features,and reducing the impact of overfitting.(3)We improve the backbone network of the IDE model and use it to train Re-ID model.The IDE model is one of the most important benchmarks for Re-ID.However,the IDE model has poor adaptability and generalization ability to complex backgrounds in practical applications.Therefore,we have made improvements to the deficiencies of the IDE model.First of all,the backbone network of feature extraction in the original IDE model is improved to Res Ne Xt network which is more concise,efficient and lower computational complexity.Then,the internal structure of Res Ne Xt network is improved so that it can adaptively set the output dimension of the full connection layer according to the number of pedestrian categories in the training set.The training of the pedestrian image style transfer model designed in this article and the Re-ID model after improving the IDE backbone network will form a confrontation.With the quality of style transfer images is continuously improved,the re-identification model can improve the ability to extract the details of the image.The two models constitute a comprehensive and integrated Re-ID system.Experiment results show that our method finally achieves three improvements: the accuracy of recognition,the robustness of pedestrian pose,and the generalization ability.