Research On Pedestrian Re-identification Algorithm Based On Deep Learnin

With the rapid development of computer technology and artificial intelligence,intelligent security,intelligent search,unmanned shopping malls and other fields have sprung up,which has brought great convenience to People’s Daily life needs.Therefore,as the core of these technologies,person re-identification has become the focus of research in recent years.Person re-identification is essentially the task of retrieving a person given an image of a target person on different devices.However,the person re-identification task has great limitations in practical applications due to the interference of background factors,occlusion of non-target pedestrians,and dislocation of pedestrian parts.Aiming at the above problems in person re-identification task,the main research contents of this paper are as follows:(1)In order to reduce the interference of background information and effectively enhance the foreground information,a person re-identification algorithm based on information complementary attention module is proposed.An information complementary attention module is designed to connect the channel dimension and the spatial dimension,and suppress the background information.The proposed model has a concise structure,simple and strong interpretability,and does not need additional branches to guide the auxiliary training.Six loss functions,including identity loss and triplet loss,are used to constrain model training.In addition,the degradation layer was analyzed to explore parameter Settings more suitable for person re-identification tasks to enhance the overall generalization ability of the model.The heat map visualization of the trained network was performed to analyze the differences between the network learning of different pedestrian features.(2)Occlusion is common in practical applications,and the standard person re-identification method cannot be applied to all scenes.Aiming at the person re-identification task under occlusion conditions,this paper proposes a hierarchical fusion network framework to fuse the mid-level features and high-level semantic information of the network model,which enhances the attention of the network to the low-dimensional features to better learn the details of non-occluded areas.An attention module is embedded in the baseline network to further enhance the information representation of channel and spatial dimensions.Different nonlinear activation functions are introduced in the pedestrian classification layer to prevent neuron death and improve the overall performance of the network.(3)Aiming at the problem of the dislocation of pedestrian parts under occlusion conditions,a person re-identification algorithm combined with human pose estimation was implemented.Alphapose is used to estimate human pose and detect key points,predict the positions of pedestrian parts,generate heat maps and match the global features of pedestrians to achieve the effect of parts alignment.Combined with the local information forced by pedestrian horizontal segmentation as the final scoring standard,the model has strong interpretability.