Research On Pedestrian Detection And Person Re-Identification Based On Deep Learning

In recent years,video surveillance has been widely used in public safety,smart city and other fields.Surveillance camera networks with extensive coverage generate a large amount of data every day.Due to limited storage space,surveillance cameras in public places usually have lower resolutions,resulting in blurred pedestrian images.Moreover,it is difficult for the camera to clearly capture the frontal face pictures of pedestrians,which makes it an almost impossible task to find and confirm pedestrians through face recognition technology.Person re-identification technology is a powerful supplement to face recognition technology.Research on person re-identification technology is of great significance for social security and the development of smart city.This paper studies pedestrian detection and person re-identification based on deep learning.The specific work and innovations are as follows:(1)A deformation-aware pedestrian detection model based on attention guided is constructed to solve the problem of pedestrian discrimination and localization at different scales in the pedestrian detection task.A feature fusion module guided by channel attention is designed to adaptively fuse deep features and shallow features to generate features with both high-level semantic information and rich contextual information,and improve the ability of the detection model to adapt to pedestrian scale changes.The deformation-aware detection head module is designed.It uses deformable convolution to flexibly sample pedestrian boundaries to model pedestrian deformation,generates robust features that adapt to changes in camera perspective and pedestrian’s own posture,and enhances the deformation representation capability of the detection model.The Caltech pedestrian detection dataset is used to design ablation experiments.The experimental results show that feature fusion module guided by channel attention and the deformation-aware detection head module can effectively improve the overall performance of the pedestrian detection model and reduce the miss rate of the detection model.This verifies the validity of the model designed in this paper.(2)Aiming at the problem that the same pedestrian has very different appearances and different pedestrians have similar appearances in the person re-identification task,a global feature enhanced person re-identification model based on group spatial attention is constructed.Instead of the traditional global average pooling,a global feature enhancement module is designed to extract the global features of pedestrian images,which can obtain richer pedestrian context information and solve the problem of insufficient global information extraction.A channel group spatial attention module is designed,which generates spatial attention to guide the person re-identification model to focus on certain local spatial regions of pedestrians.The model not only pays attention to the pedestrian features under a certain horizontal block,but also pays attention to the features of local regions with important semantic information.Market1501 and Duke MTMC-re ID are used to design ablation experiments.The experimental results show that compared with the traditional global average pooling,the designed global feature enhancement module can extract more sufficient and effective pedestrian global features;the designed channel group spatial attention module effectively captures the key local region features of pedestrians,improving the limitation of extracting local features based on horizontal block.The experimental results prove that the model designed in this paper has learned discriminative pedestrian features,and effectively distinguishes different pedestrians with similar appearances from the same pedestrian with different appearances.