Person Re-identification Technology Based On Unsupervised Domain Adaptation

As an important part of intelligent video analysis technology,person reidentification task is used to retrieve pedestrians with the same identity under the crosscamera,which has broad application prospects in intelligent security and criminal investigation.At present,due to the large amount of manpower and material resources needed to manually mark the pedestrian identity,the supervised person re-identification method is limited in practical application.Unsupervised domain adaptive person reidentification method is of great significance for promoting the practical application of person re-identification.However,the two main problems still exist are domain difference and pseudo label noise in target domain.In view of the above problems,this paper conducts in-depth research on the unsupervised domain adaptive person re-identification method.The main work is as follows:(1)Aiming at the problem of pseudo label noise in target domain caused by crosscamera problem,an unsupervised person re-identification method based on camera penalty is proposed.Affected by factors such as light,perspective and background,pedestrians with the same identity may have great differences under different cameras,while similar pedestrians with the same camera are more likely to be assigned the same pseudo label.In order to alleviate the imbalance of sample distance caused by camera difference,triplet loss based on camera penalty was designed to adjust the degree of sample pulling in and pushing out according to camera ID.According to the relationship between sample neighborhood similarity and individual independence,the camera penalty neighborhood pull-in loss and instance independence loss are designed.Experimental results show that this method reduces the influence of cross-camera problem on the quality of pseudo labels and improves the performance of unsupervised person re-identification.(2)Aiming at the problem of complex sources of pseudo label noise generated by clustering in the target domain,an unsupervised person re-identification method based on noise label learning is proposed.Because it is difficult to solve the problem of pseudo label noise comprehensively from the reason of pseudo label noise suppression.Therefore,the recognition performance is improved through noise correction,noise recognition and collaborative training.Based on the pseudo label consistency constraints of samples and their neighborhood samples,a noise correction method based on neighborhood consistency was designed.In order to reduce the influence of label noise on model performance,a noise recognition method based on similarity and confidence relationship was designed to select the correct samples for the two models.Experimental results show that the proposed method can learn more effectively under the condition of label noise.(3)In view of the low quality of pseudo labels in the target domain,an unsupervised person re-identification method based on high quality pseudo labels is proposed to improve the quality of pseudo labels generated by clustering from the perspectives of sample feature expression and similarity calculation.In order to obtain better feature representation of samples,a source domain generalization method based on contrastive learning is designed to learn the inherent consistent structure information between samples,so as to improve the generalization of source domain pretraining model.In order to provide more reasonable similarity calculation for clustering method and obtain more reliable pseudo labels,a soft label similarity degree based on neighborhood information integration is designed.Experimental results show that the proposed method improves the accuracy of pseudo labels generated by clustering,and significantly improves the recognition performance of unsupervised person re-identification.