Research On Pedestrian Detection And Re-identification Technology In Surveillance Video

With the improvement of people's awareness of public security,video surveillance equipment has been rapidly popularized,which plays an important role in the security work of public areas such as traffic,stations,shopping malls and campus.Pedestrian are one of the most important object in the security field,and related pedestrian detection and person re-identification are always the popular issues in intelligent surveillance,which have received wide attention and research.In order to realize the automatic detection and recognition of pedestrians in the surveillance scenario,this thesis separately conducts research on pedestrian detection and person re-identification in the surveillance scenario.The main works are as follows:(1)In order to complete the pedestrian detection in surveillance videos better,this thesis designs a lightweight pedestrian detection model,using Mobile Net V2 as the feature extraction network for pedestrian images.Compared with general object detection models,Mobile Net V2 is more lightweight and maintains good object detection accuracy.In order to make the model adapt to the change of pedestrian scale well,the features of different levels of Mobile Net V2 are integrated across depth,and the concatenated features are used for multi-level pedestrian detection to enhance the detection ability of the model for different sizes and proportions of pedestrians.Meanwhile,for the purpose of maintaining the lightweight of the model,the multi-level detection branches are the fully convolutional structure,which designed by Depthwise Separable Convolutions.Finally,the 88.53% average detection accuracy with 12.76 ms inference time of a single image was obtained by training and testing the model on the pedestrian dataset.By testing the detection effect of the model in the real monitoring scene,the results show that the model can complete the fast and accurate detection of pedestrians in the surveillance videos.(2)The general person re-identification methods extract a simple global feature from the pedestrian image,which leads to the lack of feature representation ability in the complex person re-identification task.Firstly,This thesis proposes a new multiscale feature learning algorithm,which combines global and local feature learning to obtain a better representation of person to improve the recognition ability of the model in complex scenarios.Secondly,considering that the misalignment of pedestrians will limit the performance of the model,in order to realize the spatial correction and alignment of human feature,the feature alignment module is used to perform spatial transformation on the feature maps from different depths of the backbone network,and further enhances the generalization performance of the model.Finally,the proposed method achieved excellent performance on the public large-scale person reidentification datasets,which is better than most of the current person re-identification methods.(3)The lightweight pedestrian detection model was combined with the person rere-identification model to realize the application-oriented person re-reidentification technology.The system can automatically and quickly analyze the surveillance videos to search the given target pedestrian.Finally,the system is tested by using a new test sets collected and made from several monitoring scenes.,the results show that the system can complete the fast search of specific target pedestrians in the surveillance videos.