Research On Person Re-identification In Multi-camera Network

With the rapid development of video surveillance technology,the application of large-scale intelligent video surveillance composed of multiple cameras is becoming more and more widespread.How to accurately identify and track specific goals in non-overlapping regions of multi-camera has become an urgent need and a huge challenge for the industry,The research of person re-identification in multiple camera surveillance network has become a research hotspot because of the difficulty,but the urgency of social needs.Person walking through the surveillance network of multiple camera fields are affected by various external environmental factors,especially the interference of similar targets and the diversity of person' attitudes.How to improve the accuracy of person re-identification in surveillance network becomes a very difficult problem.For this reason,the residual network model is used to obtain the depth characteristics of the target sample and the test sample to improve the robustness of the feature.At the same time,the more similar samples are ranked higher based on the multiple confidences re-ranking,and the matching results are more accurate,which improves the accuracy of person re-identification.Firstly,the descriptive characteristics of the target sample and the test sample are obtained;Afterwards,the similarity between the target sample and the test sample is initially sorted.When re-ranking,firstly use the mutual similarity discrimination principle to delete the wrong match in the initial sort,then pull back the real match in the initial sort to the front of the list based on the confidence intervals and the confidence weights,and use the Jaccard distance to re-rank the similarity,finally similarities were merged between the target sample and the test sample,the cluster center and the test sample.The experimental results on the three standard test data sets Market-1501,CUHR03 and MARS verify the effectiveness of the algorithm.