Study On Person Re-identification Based On Multi-Branch Feature Fusion And Camera Style Conversion

In recent years,the number of video surveillance systems is growing exponentially,which makes it impossible for people to continuously monitor the monitoring data.This requires an efficient intelligent video monitoring system to automatically analyze the monitoring data.Its main purpose is to effectively extract meaningful information from a large number of monitoring data.Person re-identification(Re ID)is one of the fundamental tasks of intelligent video surveillance system,which refers to tracking a particular pedestrian through a monitoring network composed of multiple cameras.This technology plays an important role in the field of public safety and intelligent transportation.However,the real-life video surveillance scene is very complex,such as the changes of pedestrian posture and appearance caused by different human posture,lighting and camera viewpoints,pedestrian images are blocked by obstacles,camera differences and so on,which will seriously affect the performance of Re ID model.In order to solve the defects of Re ID technology in real life,the following work is done in this thesis:1.The Multi-Branch Feature Fusion Network is proposed which can extract global and local features of pedestrian image simultaneously and to learn jointly.The network structure consists of two branches.One is a global branch for global feature learning,in this branch,the whole-body feature information of human body is extracted to express the pedestrian image as a whole.The other is a local branch for local feature learning.In the local branch,according to the characteristics of pedestrian's human structure,the image is segmented into 6 horizontal strips to bring the local feature representation of the pedestrian image.In addition,a novel integration of multiple loss functions is used in this thesis.The combination of the cross-entropy loss function of label smoothing and the triplet loss function with batch hard mining can further improve the recognition accuracy of the network.2.A Camera Style Conversion Model based on StarGAN is designed to convert pedestrian images with different camera styles to each other on the premise of retaining the original identity tag.And then the generated converted pedestrian image and the original pedestrian image are input into the multi-branch feature fusion network proposed above to identify the final pedestrian category.This method can not only enhance the training database,but also eliminate the influence of different camera styles on the performance of network.The overall recognition ability of the Re ID model can be further improved.The experimental results on the relevant major datasets show that the proposed deep learning model based on multi-branch feature fusion network can achieve state-of-the-art results.For example,on the Market-1501 dataset,the Rank1 precision is95.1%,and the accuracy of mAP is 85.9%.The addition of the StarGAN-based camera style conversion model can achieve higher recognition accuracy.Its Rank1 and mAP accuracy reached 96.0% and 86.5% respectively.