Research On Person Re-Identification Method Based On Feature Correlation

Person Re-Identification technology is one of the research hotspots in the field of computer vision.It is mainly applied to pedestrian image retrieval across cameras.With the wide use of high-definition cameras in public places,the amount of monitoring data also shows explosive growth.However,it takes time and effort to analyze and sort out these data manually.With the development of deep learning technology,the Person Re-Identification method based on deep convolution neural network framework has made a significant breakthrough and gradually replaced the traditional method.However,in practical applications,the complex background and the change of pedestrian posture are still important challenges for Person Re-Identification research.The goal of Person Re-Identification research is to improve the recognition accuracy of the target pedestrian.The realization of this goal is inseparable from the quality of feature extraction methods,so the quality of extracted features determines the discrimination of pedestrian features.In order to improve the existing feature extraction methods,the global feature extraction and local feature extraction are fused and improved to obtain more discriminative features.At the same time,we use the optimized metric learning method to train the extracted features,thus effectively improving the accuracy of the Person Re-Identification model.Specifically,the main research contents of this paper are as follows:(1)A Person Re-Identification method based on multi-scale feature fusion is proposed.This method improves the feature extraction network to obtain more discriminative features.IBN-Net network is introduced to the backbone network Res Net-50 to increase the sensitivity of features in the channel.In the global branch,multi-scale feature fusion method is used to extract and learn the global features of the pedestrian image.For the global features of different scales,label smoothing ID loss function and difficult sample mining triple loss function are used for joint training.At the same time,batch normalization operation is used to adjust the distribution of features and reduce the conflict of loss functions.The local feature branch uses the segmentation of the global feature map at different scales to achieve the depth extraction of local detailed information,and uses the tag smoothing ID loss function to train each local feature.The final fusion of multi-scale features can solve the problems of insufficient detail information extracted by global features and incomplete learning of local features.The experimental results show that this method can effectively improve the algorithm performance on both Market1501 and CHUK03 data sets.(2)A Person Re-Identification method based on local feature correlation learning is proposed.This method introduces a human posture estimation model to improve the confidence of local feature extraction of pedestrian images,and designs a correlation learning module after local feature extraction to fully integrate the correlation information between local features,taking into account the relationship between pedestrian body parts and other parts,so as to enhance the discrimination of each local feature,thus improving the retrieval ability of similar pedestrian images.In training,the cross entropy loss function and adaptive soft interval triple loss function are used to optimize the model.Experiments are carried out on Market-1501,CHUK03 and Duke MTMC-Re ID data sets,and it is proved that this method also has excellent recognition performance for pedestrian images with similar pedestrian attributes.