Research For Video-based Person Re-id Based On BRNN And Attention Mechanism

In recent years,how to find the target person in the surveillance video or determine whether they are the same in two different pedestrian videos has become the research focus of current video research.This issue is called person re-identification(person re-id).Person re-identification problem has very important application value in the field of video surveillance,especially public security criminal investigation and security monitoring.Person re-identification problems are generally divided into image-based and video-based.Because the video data is more complex than the image data,and the video-based form is more in line with the actual pedestrian re-identification application scenario,the video-based pedestrian re-identification method has broader research value and application prospects.This paper focuses on the relevant algorithm for recognizing pedestrians on the basis of a given pedestrian video,especially based on deep learning,and explores how to effectively improve the accuracy of pedestrian re-identification.In this paper,based on the introduction of convolutional neural networks(CNN)and recurrent neural networks(RNN)and some of their extended structures,with the main issues in person re-id and the characteristics of video data,we proposed two methods based on deep learning fo video-based person re-id.One method focuses on how to learn the effective temporal information in the video,and another algorithm is to explore how to automatically learn the effective information and interference information in pedestrian video,so as to improve the re-identification accuracy rate.The research content and main work of this article are as follows:First,based on the research of existing algorithms,the problem of insufficient learning of temporal information in the method of video pedestrian recognition using only one-way recurrent neural network(RNN)is improved,and a bidirectional recurrent neural network(BRNN)is proposed to simultaneously learning the temporal cues of the video in two directions,to achieve further effective learning of the overall video temporal characteristics,thereby improving the recognition accuracy of pedestrians.On the basis of the spatial features extracted by CNN,the person re-identification method based on BRNN has achieved a higher recognition accuracy of pedestrians than the existing methods.Second,aiming at the problem of scale changes and occlusion in pedestrian videos,a attention mechanism based on deep neural network is proposed to reduce the influence of interference information.Combined with the idea of multi-scale learning,the attention mechanism in the paper is based on the spatial feature of each frame in the video,and automatically learns the validity of the corresponding region of the feature,making the algorithm focus on the important areas in the pedestrian video,and at the same time reducing the effect of invalid or even interference information,thereby further improving the effectiveness of model feature extraction.This method achieved state-of-art performance in some person re-id datasets.