Research On Person Re-identification Method Based On Attention Mechanism And Feature Complementarity

With the rapid development of social economy and technology,people pay more and more attention to the stable social public security environment,and the intelligent video surveillance technology based on computer vision is widely used in various security fields.Person re-identification is a technology that can identify pedestrians with the same identity across cameras,that is,given a picture of a pedestrians to be queried,the pictures taken by the pedestrians under multiple cameras with nonoverlapping perspectives can be retrieved from the pedestrian picture library.However,in actual scenes,factors such as different shooting environments,changes in pedestrian posture,and similar appearance of pedestrians will affect the performance of the person re-identification model.How to reduce the interference of these factors so that the model can learn the invariant features of pedestrians within the class and the discriminative feature of pedestrians between classes is an urgent problem to be solved in the task of person re-identification.The main research work of this thesis is as follows:(1)In order to improve the diversity of the person re-identification dataset and reduce the impact of the shooting environment and pedestrian posture change on feature extraction,this thesis proposes a person re-identification model based on Automatic data Augmentation and Linear Shared Attention network(AALSA-Net).First,this thesis utilizes a generative adversarial network architecture to search for suitable data augmentation strategies for person re-identification datasets.Then,the person reidentification backbone network is optimized using instance normalization and batch normalization neck,and a channel attention module and a linear shared attention module are further designed and embedded in each convolutional stage of the network.The channel attention module uses two different pooling methods to extract channels with high response to the pedestrian foreground.The linear shared attention module not only strengthens the salient regions inside a single image,but also introduces shared external prior knowledge to capture the connection between samples.This connection makes pedestrians with the same identity have similar feature representations and improves the robustness of the model.(2)In order to extract the potential and discriminative details of pedestrians,this thesis proposes a person re-identification model based on Multi-Granularity Salience feature Complementary network(MGSC-Net).MGSC-Net consists of backbone network,global feature extraction branch and local feature extraction branch.Each convolutional stage of the backbone network embeds a secondary salient feature excavation module after the attention module to capture latent features that are easily overlooked.The global feature extraction branch uses the feature fusion module to integrate salient features at all levels,and the local feature extraction branch divides pedestrian features into multiple regions for feature encoding.Features are extracted at multiple granularities through two branches to form pedestrian features with rich and discriminative semantic information.(3)This thesis conducts experiments on two public datasets Market1501 and Duke MTMC-Re ID.The experimental results show that the method proposed in this thesis can effectively extract distinguishing and robust features of pedestrians and improve the accuracy of pedestrian retrieval.