A Research On Person Re-Identification Based On Convolutional Neural Network

Person re-identification(re-ID)is mainly used to identify the same person in different frames of the different/same camera.It is very important in many fields such as public safety,photo album clustering,and intelligent identification.Especially in security work,it is very important to use the current advanced technology to automatically analyze the massive datasets captured by cameras in real world to provide effective evidence for criminal investigators.Convolutional Neural Nsetworks have been proved to be very effective for large-volume image processing.Therefore,this thesis aimed at research on person re-identification based on Convolutional Neural Network(CNN)and explored methods to improve the performance effectively.Because in the real situation,influenced by light,person posture,person detector and occlusion,the direct use of CNN to extract global features cannot meet the practical application.Therefore,the current part-based re-ID methods using CNN are analyzed in this thesis.A novel method-person body internal structure distance(ISD)is proposed to learn the potential internal structure information of pedestrians by using the local distance of different parts of the pedestrian in the same picture.According to ISD,a small branch is added to the traditional classification network.For the first time,it is proposed to use the internal structure constraints to guide the representation learning of global features.Experiments have been carried out on the current general datasets.Proposed method can effectively improve the baseline.The internal structure distance of the person can guide the network to learn better representation.In addition,our branch overhead is only 0.8 percent of the baseline parameter,but can effectively accelerate network convergence.In addition,for the commonly used triplet loss in CNN,an improved triplet loss(ITLoss)is proposed in the thesis,which can achieve a smaller intraclass distance and a greater interclass distance.In order to speed up the convergence,we introduced batch hard training mechanism.Person re-identification can be seen as a ranking task or a clustering task.Therefore,performance of ITLoss on these two tasks are evaluated in this thesis.Experiments shows that our loss function has better generalization ability and can effectively improve accuracy.Finally,the person re-ranking method is an effective step to improve the recognition accuracy after the CNN learns an effective feature representation.Therefore,this thesis analyzes the current re-ranking method and proposes an improved re-ranking method(Ire-ranking),which automatically adds more context information in each iteration of re-ranking.Our method can effectively find more matches.In addition,in order to make better use of deep features,we propose a feature fusion method-FFM.Through experiments,we verify that both methods can effectively improve performance.In particular,the combination of the two methods can lead to even further performance gains.Improvements of the network architecture,loss function and re-ranking method are proposed in this thesis,proposed methods are compared with the current related methods on three general datasets,which proves that proposed methods can effectively improve the performance of person re-identification.