Research On Person Re-identification Technology In Non-overlapping Views

Public security is the prerequisite for social stability and national economic development.Video surveillance plays an irreplaceable role in the field of public security prevention.With the gradual popularization of monitoring equipment,the monitoring data has exploded,and the inefficient manual monitoring methods are no longer desirable.Person re-identification,which can quickly query the target person from the massive video,has become a research hotspot in the field of computer vision.This paper mainly research on person re-identification technology in non-overlapping views,that is,to determine whether the target person in one camera appears in another camera's view in the cross-camera system.Person re-identification technology is studied based on manual features and deep learning methods,and the main research results are as follows:1)To deal with the drastic changes of pedestrian appearance due to the influence of external interference factors such as illumination variation and different viewpoints,this paper proposes normalized metric learning based on multi-feature fusion for person re-identification,integrating three discriminative feature representations and two fast and practical distance metric learning models.The experimental results demonstrate that the fusion of multiple features can fully tap the potential representation capabilities of different features,and the aggregation of different distance metric learning models can further improve the utilization of each feature.2)To cope with the problem that global features pay more attention to the overall appearance description and ignore representative details,a multi-scale features network is proposed.Based on the highly structured of pedestrian images,a multi-branch network is designed to extract local features based on structural information and merge them with global features into a deep feature.Based on this,a combined loss function is designed to release the maximum distinguishing performance of the network.3)In order to deal with the problem of pedestrian misalignment,this paper puts forward pedestrian-aligned multiscale features network for person re-identification.The spatial transformer networks is utilized to re-localize pedestrian images,which is fused with multi-scale features network to extract features from aligned images.The experimental results demonstrate the superiority of the method with certain application value.