Research And Implementation Of Non-overlapping Person Re-identification

In recent years,the amount of monitoring data has grown rapidly with the gradual expansion of the monitoring system.And the existing monitoring systems rely on humans to view and process large amounts of data,which is not only time-consuming but also very inefficient,and it is difficult to maintain full-scale monitoring.Therefore,large-scale monitoring systems that can work together have emerged.However,there are problems such as occlusion and changes of the illumination,viewing angle and pedestrian posture in the actual monitoring system,which cause the same person to have a great difference in appearance in different monitoring videos.In order to cope with these problems,this thesis proposes the person re-identification algorithm based on region segmentation and distance fusion,and combined with depth features to applies to an actual system.The main research work and innovations are as follows:1.Firstly,the overall pedestrian images are segmented in horizontal direction,and the color features and texture features are respectively extracted from each of the divided horizontal regions.Secondly,in order to eliminate the influence of occlusion and posture and perspective changes as much as possible,the concepts of invalid region and valid region are proposed.The valid regions are retained and the invalid regions are removed,so the influence of the invalid region on person matching is reduced.2.Most of the person re-identification algorithms concatenate the extracted features directly.However,the direct concatenation of various features does not take into account the differences among different features and the learned measure matrix could not adequately represent the characteristics of different features.Therefore,this thesis proposes the method of distance fusion,which performs the distance measure learning on local features,spatio-temporal features and global features respectively to obtain independent measure matrices.Finally the distance matrices obtained by independent measure matrices are fused to obtain the final matching ranking.3.An actual cross-camera person re-identification system is designed.The global appearance features of pedestrian images detected by the system in real time are extracted and the local traditional features are also extracted by using the proposed region segmentation and invalid region removal methods.Then the deep learning is used to extract the depth features.Finally the traditional features and the depth features are combined with the proposed distance fusion to realize the online warning function of the system.