Research On Person Re-identification Technology In Video Surveillance System

With the development of society,multi-level monitoring network,as an indispensable part of infrastructure,has been steadily completed and expanded year by year under the strong promotion of the country,and the amount of surveillance video data has increased exponentially.Based on the demand of public safety and commercial services,it is time-consuming and inefficient to monitor and track pedestrians in massive video data only by human.Driven by the demand,person re-identification as a key part of intelligent video analysis technology has gradually become a research hotspot.However,such factors as the difference of hardware configuration of cameras,pedestrian posture,illumination,occlusion and the inability to automatically obtain pedestrian identity labels make this research extremely challenging in actual monitoring scenes.Therefore,how to extract pedestrian features with robustness and discrimination,and how to cluster the samples reliably without ground-truth labels as supervisory signals,are the key issues of the current person re-identification task.Focusing on the above problems in the video surveillance system,the main research work of this thesis is as follows:(1)In order to solve the problem of the misalignment caused by pedestrian detection error in person re-identification,the current part-based deep neural networks only learn the adjacent local relationship,resulting in the lack of long-distance local correlation.This thesis proposes a person re-identification network based on first-order and second-order spatial information.On the backbone network,first-order spatial mask is embedded to fine-tune the spatial weight of the input image to reduce the background interference.The second-order spatial mask is used to model the long-distance dependency,and local features are integrated into the dependency model to obtain the global feature representation.In the local branch,dropblock is introduced to regularize the pedestrian features to avoid the network model relying too much on specific part features.In the training stage,the whole network is optimized by the label-smoothed crossentropy loss and the triple loss with positive samples' center.Experiments on Market-1501 and Duke MTMC-re ID data sets show that the feature extracted by the proposed method is more discriminative and robust,and improves the accuracy of person re-identification.(2)In view of the fact that the deviation based clustering method ignores the control of the number of clustering samples,which can easily lead to super-large clusters,this research proposes an unsupervised person re-identification model based on hierarchical clustering by deviation and diversity regularization.In the initial stage,a sample-level pseudo label is given to each pedestrian image,and then the network is trained by the samples of the training set.According to the image features extracted by the network,the intra-cluster and inter-cluster deviations are calculated respectively as the clustering criteria.It is more inclined to combine the outliers first,or the clusters with smaller intra-cluster deviations,so as to effectively weaken the impact of individual deviations on the clustering effect at the initial stage of clustering.At the same time,the diversity regularization is introduced to prevent the number of cluster samples after merging from being too large.The model is fine-tuned and the clustering information is updated in continuous iteration until the performance of the network model is no longer improved.The performance comparison and effectiveness of the proposed algorithm are verified on Market-1501 and Duke MTMC-re ID person re-identification data sets.The experimental results show that the clustering effect of the improved clustering criterion is better,the number of super large clusters is reduced,and the performance of model re-identification is improved.