Research On Person Re-identification Technology Based On Deep Learning

Person Re-identification is a crucial research area in the field of computer vision,which aims to solve the problem of pedestrian identification and retrieval across cameras and scenes.It can be widely applied in fields such as security surveillance and criminal investigation.In this thesis,we use deep learning techniques to study the difficulties in person re-identification,and the main contributions are as follows:(1)To address the problem of poor recognition performance of current lightweight person re-identification models,this thesis proposes a lightweight person re-identification method that combines global and local features.The method is based on OSNet(Omni-scale Network),and designs CGLF-OSNet(OSNet that Combines Global and Local Features),which includes global feature extraction and local feature extraction: in global feature extraction,an improved PAN module is introduced,which can better fuse features from different levels;in local feature extraction,a sliding window idea is used to divide the feature map into stair-like blocks,which enables the model to learn fine-grained features of local regions while preserving the correlation information between regions.At the same time,by jointly guiding the model training with cross entropy loss(Cross Entropy)and triplet loss with batch hard mining(Triplet Loss with Batch Hard Mining),we can better mine the fine-grained difference information between person images.Experimental results show that our proposed method can achieve good recognition accuracy while having much fewer parameters than conventional networks.(2)In view of the problem that the performance of the current cross-modal person re-identification model is generally poor,this thesis proposes a cross-modal person re-identification method based on triple attention mechanism(TAM),which uses Res Net50 as the backbone network and designs TAM-Net.TAM-Net adopts a dual-stream network structure to better extract features of different modalities,and uses the method of network parameter sharing to map multi-modal features to a common feature space.In addition,by designing a triple attention module to enhance feature learning,we can make the model pay more attention to the salient features of the target during training.Finally,we use cross entropy loss,weighted regularization triplet loss and hetero-center loss to jointly guide the model training,so that the model can better learn the shared features between different modalities.Experimental results show that our method has higher recognition accuracy in cross-modal person re-identification tasks.(3)Most of the existing person re-identification methods can only handle segmented pedestrian images,and cannot perform retrieval on surveillance videos.Based on this,we design and implement a person re-identification system for video surveillance scenarios based on the re-identification methods proposed in Chapter 3 and Chapter 4.Since the video data collected in surveillance scenarios contains a lot of background interference information,it cannot be directly input to the person re-identification model.We combine the object detection algorithm with the person re-identification algorithm,and use the real-time object detection network YOLOv7 to detect the pedestrians appearing in the video,and input the detected pedestrian information to the person re-identification model,thus achieving the retrieval of the target pedestrians in the video.