Research On Methods Of Feature Fusion And Metric Learning In Person ReID

Person re-identification is an image retrieval technology for pedestrians,which aims to retrieve specific pedestrians from cross-camera images.With the popularity of high-definition cameras in public places,the amount of monitoring data has multiplied.In the face of massive data,manual analysis and sorting not only consumes manpower and material resources,but also is very inefficient.In recent years,significant breakthroughs have been made in person re-identification methods.The accuracy rate of person re-identification methods has exceeded the human level which proposed in existing literature on public datasets.However,complex backgrounds and changes in pedestrian postures will cause great disturbances in the real world,so the research of person re-identification method is still full of challenges.The goal of person re-identification research is to improve the accuracy of pedestrian recognition,in which the quality of feature extraction method determines the pedestrian feature is discriminative or not.In order to improve the existing feature extraction method,this thesis increase the discrimination of feature through the fusion of part feature and posture feature respectively.Furthermore,this thesis also designed the improved metric learning method to train the feature extraction model of part feature and posture feature,which can effectively enhance the pedestrian recognition accuracy,the main research contents are as follows:(1)This thesis proposed an improved person re-identification method through part-based feature.In this method,the residual blocks in the backbone network were improved by adding SE blocks to strengthen the sensitivity of the features on the channel-wise.At the same time,the Ge M pooling method was used to integrate the spatial-wise features.Global features extracted with the use of the global average pooling in global branch,and this branch was trained with label smooth loss,soft-margin adaptive weighted triplet loss and center loss.In order to reduce the conflict between the losses,the distribution of features was adjusted by batch normalization.The local branch adopted horizontal pooling to extract local features,and the label smoothing loss was set separately for each local setting for training.Finally,the global feature and the local feature were stitched together as the fused feature.This method has achieved good results on Market1501,Duke MTMC-re ID and CUHK03 datasets.(2)This thesis proposed an improved person re-identification method based on human pose feature.In this method,the key points and part affinity fields of pedestrians were extracted by Open Pose.The body pose estimation sub-network was designed for the extraction of body pose features.Then,body pose feature was used for the enhancement of human body foreground and the prediction of part visibility,and part features were matched.The distance between occluded pedestrians was optimized by reducing the mismatched partial difference caused by occlusion,and the person re-identification model was trained by the combination of visibility prediction loss function,part matching loss function,classification loss function and soft-margin adaptive weight triplet loss function.Experimental results on Occluded-Re ID and Partial-Re ID datasets show its effectiveness.