Unsupervised Person Re-identification Based On Deep Residual Network

Person re-identification is an extended research direction of person tracking,that is,cross-mirror tracking.Combined with person tracking,small segments of motion trajectories can be integrated into the entire motion trajectory,which is conducive to subsequent analysis.Research on person re-identification based on deep learning requires massive amounts of data,and manual image annotation requires a lot of manpower and time.Therefore,unsupervised person re-identification based on unlabeled target data set has attracted widespread attention.In this thesis,aiming at the problem of low accuracy of unsupervised algorithm,an improved deep residual network algorithm based on soft multi-label is proposed.The problem of unsupervised target without label is solved by soft multi-label,and the improved deep residual network is used to get better accuracy.Through experimental verification,compared with related algorithms,this algorithm has a greater accuracy improvement.This thesis describes the research results of person re-identification technology in unsupervised learning and cross data set direction,and improves the research for the existing problems.First of all,for the unlabeled problem of the target data set,this thesis uses the method of reference agent and soft multi-label learning to add soft multi-label to each target,specifically by comparing the characteristics of the target with the characteristics of the reference,so the feature learning is the key point of the research.In feature extraction,the deep residual network is used to extract the deep features;the spatial and channel compound attention model is used to reorder the feature weights to highlight important fine-grained features.Then the multi-layer features are fused to combine the deep features with the shallow features.Feature fusion realizes information complementation and obtains more accurate feature representation.Then calculate the loss function according to the feature representation,where the cross-camera loss function can reduce the difference between the soft multi-labels of the same person under different cameras.The cross-domain loss function reduces the difference in data distribution between the target data set and the reference data set,to make the learned soft multi-label more accurate.Finally,the distance loss function is calculated through the soft multi-label similarity and feature similarity to reduce the intra-class difference and expand the inter-class distance.The final loss value is obtained by weighted calculation of three loss functions,and then the improved adaptive moment estimation optimization algorithm is used to iterate the model parameters to finally obtain the minimum loss value and achieve a faster convergence speed.In this thesis,the above methods are experimentally verified.Firstly,the importance of the labeled reference data set to the experimental results is verified.Secondly,the accuracy of adding attention model to the network is verified.Afterwards,the residual network feature fusion experiment is carried out to find the optimal fusion method.Then the Ablation Experiment of loss function is carried out to verify the function and importance of each loss function.Finally,a comparative experiment is carried out to prove the effectiveness of the optimization algorithm.This thesis provides theoretical support for promoting the transformation of research results.