Research On Person Re-Identification Technology Based On Deep Learning

With the spread of COVID-19,the government’s investigation of the movement of citizens has become one of the key factors to control the epidemic.China’s ever-improving monitoring system is an important part of the entire social security system.In a complex monitoring system,the parameters of the cameras are not the same.At the same time,with the increase of the monitoring range,the facial features of pedestrians are not easy to be captured and distinguished,so it is necessary to analyze the features of the whole body.Person re-identification is proposed in this condition.Based on feature matching,the person re-identification task takes a single pedestrian image after pedestrian detection and cutting as input,and selects the same identity image under other cameras from the set to be matched.The task of person re-identification faces great difficulties: first,the low resolution of the collected images leads to unclear features;second,effective features cannot be extracted under the influence of illumination;third,the feature extraction is insufficient due to occlusion and different camera angles.From the perspective of composition,the person re-identification module is mainly divided into a feature extraction stage and a feature matching stage,and the thesis will focus on the two stages for further research.For the feature extraction stage,the characteristics of the same pedestrian from different perspectives change greatly,so it is inevitable to encounter occlusion in the movement of pedestrians.The thesis adopts the methods such as residual network,pose recognition and model transformation,and proposes a feature extraction method based on pose-guided feature extraction network(PIRA).In this method,the feature extraction based on Resnet network is firstly performed on the initial input.Then,the proposed method combines the technology of generating human key points with pose guidance,extracts feature maps of different parts,and decouples the aggregated human features.The next step is to group human body parts,and use the In-Rest module to make the feature map of each part integrate the information of the adjacent parts,which improves the accuracy of the whole model in the feature matching stage.Aiming at the problem that pedestrian features are not easy to align and have large differences in the feature matching stage,this thesis proposes a feature matching method based on graph matching.First,it combines the generation characteristics of pose-guided features,and uses the graph matching method for all feature maps,so as to obtain the feature difference between the two images to be matched.In order to solve the problem of inaccurate initial ranking in the feature matching stage,the thesis uses the context information of the matching results to optimize the matching and sorting results,narrow the scope of the context used as much as possible,and reduce the computational complexity.Further,this thesis proposes a person re-identification method based on m-reduced neighborhood.Finally,combined with three large public datasets of person Re-identification and two data sets dedicated to occlusion,this thesis experiments on the overall person Re-identification model to verify the high accuracy of this model,and uses market-1501 to carry out parameter experiments and performance experiments on the re-ranking model to verify the effectiveness of the re-ranking model.