Research On Key Technologies Of Person Re-identification

Aiming at matching the target person under different non-overlapping cameras,person re-identification has become an important task in the computer vision.With the development of deep learning,person reidentification has achieved promising results,and possessed a great practical value in the field of video surveillance and public security.The main research contents and innovations of this thesis include:1.In the person re-identification based on image data,a new spatial and channel attention module is proposed,which enhances pedestrian features at different level.2.In the person re-identification based on occluded data,a person representation algorithm is proposed,which combines the human key-points model and graph convolutional network to reduce the effects of occluded region on the global features and improve the generalization performance of occluded person re-identification.3.In the person re-identification based on sequence data,a temporal convolutional module is proposed,which can extract effective temporal information.In addition,a feature-learning model is designed to extract gait information.Through the features of the four limbs,the effective extraction of gait features is realized.Finally,the performance of the algorithm is compared with the current representative algorithms on multiple public person datasets,and the results verify the effectiveness of the three proposed algorithm.