Research On Person Re-identification Based On Camera Network Topology

With the steady advancement of safe city and smart city construction in China,a large number of video surveillance equipment are deployed in the streets of the city.These video surveillance devices monitor abnormal behavior,find lost children and track suspects which provide a strong guarantee.But in the face of so much video data,the traditional manual monitoring system is under restrictions,so it is particularly important to find an intelligent video surveillance system that can automatically retrieve and track target pedestrians.Person re-identification technology,as the core algorithm in intelligent video surveillance system,has attracted in-depth research by experts and scholars at home and abroad in recent years.In this paper,two different solutions are proposed to solve the problem that existing person re-identification algorithms only relying on visual information lead to low recognition rate when matching pedestrian images:1.A re-ranking algorithm based on forward and reverse sorting constraints is proposed.The forward sort constraint is that if the image is in the front position of a forward sorted list,its position in the other forward sorted list should be relatively backward.The reverse sorting constraint means that if the two images are the same pedestrian,both of them should be located in the front position of each other's sorting list.2.A spatial-temporal fusion model based on camera network topology is designed.This model first constructs the spatial-temporal relationship of the corresponding camera pairs by counting the movement time of some pedestrians passing through any two cameras in the target monitoring place.Then,using the spatial-temporal relationship,the probability of occurrence of any two pedestrian images capturing time differences is calculated as a spatial-temporal similarity score between them.Finally,the spatial-temporal similarity score is merged with the original visual similarity score as the fusion similarity score between the corresponding pedestrian images.