Research On Person Re-identification Method In Video Surveillance

In recent years,with the improvement of people’s public security awareness,the number of video surveillance equipment has increased rapidly.Video surveillance technology has been widely used in urban public places and public security investigation.However,for the massive video monitoring data generated every day,the traditional manual monitoring method can no longer meet the demand.So it is very necessary to develop an intelligent video monitoring system that can assist manual identification.Person Re-Identification(Re ID)technology can automatically retrieve the target pedestrian under more complex conditions.Therefore,in the monitoring scene,the pedestrian recognition technology is more valuable than the technology that relies on high-definition face for identity recognition.In order to realize the intelligent retrieval of the target pedestrian under monitoring,this thesis studies pedestrian detection and pedestrian re-recognition.The main work is as follows:(1)In the pedestrian detection part,this thesis constructs a pedestrian data set under the monitoring scene,and enriches the data set through the brightness change to improve the generalization ability of the model.In view of the problem of small target pedestrian missing detection due to the large change of pedestrian scale in the monitoring scene,this thesis introduces the weighted bidirectional feature pyramid on the basis of YOLOv5 algorithm to achieve the fusion of high-level features and low-level features,and at the same time,give different weights to the features of different scales,so as to reduce the loss of small target information and improve the detection accuracy.In view of the problem that pedestrian occlusion is easy to miss detection in the monitoring scene,this thesis uses DIo U-NMS to replace the NMS non-maximum suppression algorithm in YOLOv5,and adds the constraint of the center point distance of the target frame when filtering the overlapping frame,thus improving the recognition accuracy.The experimental results show that the proposed algorithm is 2.5% higher than the original algorithm m AP.(2)In the part of pedestrian re-recognition,the pedestrian image detected in the monitored scene is often disturbed by the background of objects or other pedestrians,which can not effectively determine which part of the pedestrian image is an important area,resulting in the problem of low recognition accuracy.This thesis proposes a pedestrian recognition method based on the dual relationship perception attention mechanism.By embedding the dual relationship-aware attention mechanism in the lightweight network Res Net18,the attention matrix is obtained by using the pairwise relationship between each feature point and the global feature on the channel and space,so that the network can focus on the feature points that have a large relationship with the global feature,thus effectively eliminating useless background interference.At the same time,the lightweight network also ensures the real-time performance of the system.In the aspect of loss function,the four-tuple loss function is used to optimize the inter-class distance,and the center loss is combined to make the intra-class distance more compact,so that the model can better learn the representation of pedestrians.The experimental results show that compared with the original algorithm,the improved algorithm in this thesis improves m AP by 4% and Rank-1 by 1.9% on the Market-1501 dataset.On the Duke MTMC-re ID dataset,m AP increased by 3.1% and Rank-1 increased by 2.1%.(3)By combining the pedestrian detection algorithm and pedestrian re-identification algorithm proposed in this thesis,an intelligent video monitoring system is developed based on Py Qt library,which includes the functions of traditional video monitoring,pedestrian detection and pedestrian re-identification,and can switch the required functions as required.The algorithm is deployed on the PC and verified under the monitoring platform.The results show that the comprehensive accuracy rate of the system measured by the improved model in this thesis is increased from 64.92% to 69.23%,which is 4.31% higher,which verifies the effectiveness of the improved algorithm and the feasibility of the system.