Research On Key Technologies Of Integrated Pedestrian Re-Identification

Integrated pedestrian re-identification technology is a computer vision technology that queries target pedestrians from the uncropped scene gallery,also known as pedestrian retrieval technology,which combines two sub-tasks of pedestrian detection and pedestrian re-identification,and is widely used in intelligent video surveillance and Intelligent security and other fields have high research value.Thanks to the huge advantages of deep learning methods in recognition accuracy compared with traditional recognition methods,the use of deep learning models is the theoretical basis of the current integrated pedestrian re-identification technology.However,in real-world application scenarios,the integrated person re-identification technology has problems such as subtask optimization target conflicts,poor recognition rate of complex similar occlusion targets,and lack of high-level semantic information.In view of the above problems,this thesis conducts research on the key technologies of integrated pedestrian re-identification for the detection sub-task framework,re-identification sub-task feature matching method,and pedestrian attribute recognition model.The specific content is as follows:(1)Aiming at the problem of subtask optimization target conflict,this thesis proposes an anchor-free integrated pedestrian detection network framework with instance-guided branches in the detection subtask part.The framework first uses ResNet50 as the backbone feature extraction network,and then builds an improved anchor-free target detection network to achieve effective multi-scale fusion feature extraction of pedestrians.Finally,an additional guidance branch based on the example pedestrian image is added,and the example-guided method is used to balance the optimization goal conflicts of the detection and re-identification tasks in the network.The framework is experimentally verified on two public datasets,CUHK-SYSU and PRW,and uses a general re-identification method to form an overall model.Experimental results show that,compared with the overall model composed of the original anchor-free target detection network,the framework can effectively enhance the model’s feature extraction ability for re-identification identity matching.(2)Aiming at the poor recognition rate of complex similar occlusion targets,this thesis proposes a feature matching method based on random occlusion and hard sample guidance in the re-identification subtask.This method first adds random occlusions to the re-identification features of the detection branch,uses the method of contrastive learning to perform additional occlusion supervision on the features,and then designs an improved loss function guided by difficult samples,and enhances the model by introducing difficult samples for feature update.The discriminative power of difficult sample features during training.This method is verified experimentally on two public datasets,CUHK-SYSU and PRW.The results show that using this feature matching method on the basis of(1)can effectively improve the accuracy of model recognition,and the recognition accuracy exceeds some other current methods.An all-in-one pedestrian re-identification model approach.(3)Aiming at the lack of high-level semantic information,this thesis proposes an identity-assisted pedestrian attribute recognition model in pedestrian attribute recognition.The model first builds a multi-branch network that uses hierarchical clustering to generate pseudo-IDs,uses ID information to assist attribute identification,and then designs a five-tuple loss function,including attributes,IDs,and attribute-IDs.Feature relationship,integrate pseudo-ID information into attribute space,and finally perform attribute identification.This method is verified on the PETA public dataset,and compared with the common pedestrian attribute recognition model,it has achieved a 1-5% improvement in the important attribute experimental indicators.Adding this model to the overall model in(2)can achieve additional pedestrian attribute output in the integrated pedestrian reidentification model.In summary,this thesis conducts research on integrated pedestrian re-identification technology from three aspects: subtask balance,similar occluded target scenes,and highlevel semantic information,and proposes an anchor-free integrated pedestrian detection network framework with instance-guided branches.A feature matching method based on occlusion and hard sample guidance and an identity-assisted pedestrian attribute recognition model to alleviate the subtask conflicts of the integrated pedestrian re-identification model,poor recognition of similar occluded targets,and lack of high-level semantic information,which has practical significance for the development of integrated pedestrian re-identification technology.