Research On Person Re-identification Methods Based On Deep Learing

Person re-Identification(Re ID)belongs to the field of computervision.It is an effective method to use multiple non-overlapping camerasto retrieve and judge different pedestrians,and then to identify and track matching pedestrians.It is often used in some intelligent security scenarios.It plays an irreplaceable role in criminal investigation.However,pedestrian information exposed in natural scenes is easily interfered by factors such as viewing angle,illumination,resolution,etc.,and the camera will also switch to infrared mode due to insufficient light at night.All these factors bring great challenges to the task of person re-identification.In this paper,we focuse on the person re-identification method based on deep learning,and study the problem of single-modality and cross-modality person re-id.The main contents are as follows:(1)On the basis of understanding the previous work on person re-identification in single-modality and cross-modality,the characteristics of some common methods are summarized,focusing on the feature extraction and existing problems of pedestrian images,pedestrian re-identification The contents and research methods of person re-identification are analyzed and studied systematically.(2)A visual attention model based on cross-latitude interaction is studied.Aiming at the problem of the lack of interdependence between the channel dimension and the spatial dimension in the visual attention mechanism,which affects the robust feature representation of CNN,the experiment proposes to start with a three-branch attention structure,by capturing the spatial dimension of the feature map and the input tensor.Cross-dimensional interaction information between channel dimensions to establish dependencies between attention dimensions.The experimental results demonstrate the importance of multi-dimensional interactions in the attention mechanism.(3)A visual activation function based on spatial characteristics is studied.In the conventional practice of extracting pedestrian image features by convolutional neural networks,most of them are to create spatial dependencies in the convolutional layer,and then perform nonlinear transformation of the activation function respectively,but this will lead to the insensitivity of spatial conditions in the activation process.Faced with more and more complex pedestrian images in life,it is very important to study the fine spatial layout of adding spatial conditions to nonlinear activation functions to extract pedestrian image features.Experiments compare the proposed visual activation function with traditional activation functions and demonstrate the excellent utility of the former.(4)A cross-modal data augmentation method based on generative adversarial networks is studied.In the cross-modal field,there is not only the problem of person recognition in a single modality,but also the difference of pedestrians between different modalities.On the basis of realizing the effective identification of pedestrians in a single modality,an improved cross-modal data set is proposed from the perspective of modal inter-transformation in view of its own problem of differences between modalities.Comparing the data sets before and after the improvement,it is confirmed that the method of modal inter-transfer can alleviate the problem of pedestrian modal differences.