Research On Person Re-identification Technology Based On Deep Learning

Person re-identification is a very important research topic in the field of computer vision.By using the technology of pedestrian re-identification,the video surveillance data can be retrieved efficiently,and the clues of the target person can be obtained quickly,so as to win the valuable time opportunity for the timely investigation of the case.Computer vision technology is used to determine whether there is a specific object or pedestrian in the image or video sequence,which is called pedestrian recognition,and is also usually considered as a sub-problem of image retrieval.The process of person re-identification is usually as follows: firstly,given a person image of interest,and then matching features with person appearing under different cameras,and then judging where the person appear in sequence according to the results of the feature matching.However,in the actual scene,the person's trajectory is unconstrained and can appear under different cameras in different time periods.The same person may have different feature representations due to the changes of illumination,scale,posture,angle of view,and the problems of pedestrian being blocked and background clutter.This paper mainly studies the technology and algorithm of person re-identification under two different data sets of single image and video sequence.This paper proposes two different network models based on deep learning to solve the technical problems encountered by person in the corresponding situations.The main work of this article is as follows:For image-type data sets,this paper proposes a convolutional neural network model based on mask alignment to solve the problem of person image misalignment and background complexity in person re-identification.The algorithm in this paper uses a fully convolutional neural network and global average pooling operation to implement the mask alignment module,and then solves the problem of person misalignment through the mask alignment module.At the same time,considering the different levels of convolution kernels in the convolutional neural network,the extracted feature information is different.In this paper,a multi-feature fusion module is added to the proposed convolution model,and a multi-level feature information is fused to obtain a robust person feature representation.For video-type data sets,this paper proposes a spatial-temporal attention network model,which is used to solve the problems of pedestrian occlusion and background clutter in person re-identification.For the spatial-temporal attention network model,the spatial attention module and the temporal attention module are used to fully mine the person feature information in the video sequence.Through the spatial attention module,the key feature information of person in the person image is extracted to enhance the feature extraction ability of the person re-identification model.At the same time,through the temporal attention module,the model can effectively focus on the important frames in the video sequence,so as to fully extract person features.