Research On Person Re-Identification Methods Based On Deep Feature Embedding

Person re-identification is an important task in video surveillance systems.It aims to retrieve all images or videos of a specific pedestrian from the gallery set.It can retrieve target images collected by surveillance cameras across regions.It is used in related tasks such as security construction,person tracking and trajectory analysis.Due to the huge research and application value,person re-identification has received a lot of attention.However,complex factors such as changes in camera resolution and angle,pedestrian walking posture and lighting differences have brought challenges to the accuracy of person re-identification.This thesis studies person reidentification based on deep feature embedding and discusses person re-identification methods in visible-visible and visible-infrared image scenes respectively.The main research and contributions are as follows:(1)This thesis proposes a visible-visible person re-identification method based on an OmniScale Feature Aggregation(OSFA).Due to the phenomenon of occlusion,pose and angle change of pedestrians during walking,a convolutional neural network is used to build a model based on full-scale feature aggregation to learn discriminative pedestrian features.During the training process,the network extracts the discriminative local features and global features of pedestrians for re-identification.In view of the distance in intra-class is greater than that in inter-class,OSFA combines multiple losses to reduce the intra-class feature distance of the same person under crosscamera.On the visible-visible person re-identification datasets,OSFA learns the representation information of pedestrians and performs well.(2)This thesis proposes a visible-infrared person re-identification method based on a Hierarchical Cross-modal Feature Network(HCFN).The method uses a convolutional neural network for training on visible-infrared image datasets.HCFN includes two parts: an Intra-modal Feature Extraction Module and a Cross-modal Graph Interaction Module.A Hierarchical Attention Module is proposed in the Intra-modal Feature Extraction Module,which can help the network learn intra-modal discriminative features of visible and infrared images.The Cross-modal Graph Interaction Module can narrow the pedestrian images of the same identity in different modalities,reducing the modal gap problem caused by attribute differences between visible light and infrared images It can supervise the networks learn identity-related but modality-irrelevant representations.Extensive experiments on the visible-infrared person re-identification datasets demonstrate the advancement of HCFN and the effectiveness of each components.