Pedestrian Identification Based On Global And Local Information

With the advent of the big data era and the development of deep learning,the field of target recognition has made a breakthrough.As a branch of target recognition,pedestrian identification has been widely used in security,finance,education and other fields.Pedestrian identification relies on face recognition,which can achieve high accuracy for pedestrians in cooperative environments.However,for pedestrians walking freely in the video,their faces may have occlusion,motion blur and various deformations,which makes face recognition more difficult.It is difficult to get good recognition performance only using face recognition algorithm.At present,some methods use face tracking algorithm to obtain face image sets and adopt face recognition of multiple face image frames to improve the recognition accuracy.However,due to the low resolution and large deformation of the face,the robustness of face tracking is hard to guarantee.This paper proposes an identification method based on global and local information for pedestrian in non-cooperative environments,which can improve the accuracy of pedestrian identification in complex situations such as deformation or occlusion.The main work of this paper includes two aspects:Firstly,taking the pedestrian on the road as the identification target,this paper proposes a single-camera pedestrian identification method based on global and local information.Compared with the face,the body has higher resolution and smaller image differentiation when it is deformed.In this paper,global appearance features are used to improve the robustness of tracking.Through pedestrian detection and tracking,we can get body image sets and identity numbers of different identities.Face detection is performed on each pedestrian to obtain the face image set.The pedestrian identity number is mapped to the face image set,and the face image set of each identity number is recognized.So,we realize the combination of global pedestrian information and local face information.At the same time,this paper uses the prior knowledge of face distribution to remove the branch and optimize parameters for the face detection network,which reduces the network computation and improves the performance of face detection in this scene.The experiment shows that the proposed method has higher recognition accuracy than the method which only uses local face information to realize pedestrian identification.Then,in order to solve the problem that it is difficult to identify pedestrians who moving away from the camera,this paper proposes a cross-camera pedestrian identification method based on global and local information.This method realizes selective identification based on the velocity direction.In order to ensure the accurate pedestrian movement direction,the sliding window method is adopted to filter the velocity to eliminate the abrupt change in the velocity direction.According to the direction of pedestrian movement,different identification methods are selected.For pedestrians moving towards the camera,the single-camera pedestrian identification method is used.For pedestrians moving away from the camera,we combine the pedestrian re-identification with the online pedestrian gallery construction.This method is applied to multi-camera scenes,and extends the function of identifying the pedestrian based on the back view.In conclusion,the pedestrian identification method based on global and local information proposed in this paper has good performance for pedestrians in non-cooperative environment,which improves the robustness and accuracy of pedestrian identification to some extent.This method can not only realize the identification of the pedestrian moving towards the camera,but also expand the function of identifying the pedestrian moving away from the camera.