Video Person Search System Based On Fast Re-identification

With the large-scale construction and deployment of urban video surveillance systems,person re-identification technology has gained more and more attention and application as an important part of intelligent video analysis.In actual use,after getting the surveillance image,it is often necessary to perform pedestrian detection first,and then re-identification to judge whether it is the target person.This two-model solution is slow,inefficient,and consumes large amounts of video memory,which cannot achieve satisfactory results.This subject focuses on computer technology in artificial intelligence,researches detection algorithms and re-identification algorithms based on deep learning methods,and designs a person search algorithm that combines detection and re-identification.This paper designs a model for person search based deep convolutional neural network,and proposes a joint training strategy to implement a fast and efficient person search scheme.For the person re-identification algorithm based on deep learning,this paper proposes an improved version of the triplet loss function,whose role is to optimize the distance between samples by adding an absolute distance constraint term.This paper evaluates and tests related performance on public data sets,and finally verifies the effectiveness of our loss function.For the person search algorithm based on deep learning,this paper uses ResNet18 as the basic network to design and implement a person search model that combines detection and re-identification,so that the two tasks share the basic network.After experimental verification,compared with the two model schemes,the person search model can significantly save the size of the model memory and increase the speed by more than 25%with the same accuracy.Finally,this paper designs the person search system software,deploys the algorithm in the background,and implements the fast person search system.