Research And Implementation Of Video Structured Technologies Driven By Deep Learning

With the continuous advancement of China’s Safe City,Xueliang project and society comprehensive governance system,a large amount of monitoring cameras have been installed in various cities.The massive monitoring video generated by these monitoring cameras has problems of redundant data and disordered information,while video structured technology can organize and manage the video structurally,which can effectively support the needs of current intelligent monitoring information construction.Pedestrian re-identification and pedestrian attribute recognition are two important research directions in video structured technologies.When the video can capture the target image,pedestrian re-identification can recognize the target through the image features.When the video cannot capture the target image,pedestrian attribute recognition can further take the semantic features into account to identify the target.However,current video structured technologies still face three main problems.First,when the video can capture the target image,the existing pedestrian recognition technologies cannot link multiple cameras to accurately locate and track the same target when the pedestrian quickly changes position,resulting in a lower recognition rate for the same pedestrian.Second,when the video cannot capture the target image,the target can only be found through the semantic attributes,but the existing pedestrian attribute recognition technologies does not make full use of the relationship between the inherent structure of human body to predict pedestrian attributes from multiple perspectives,resulting in a high rate of attribute recognition errors.Third,most of the current video structured technologies only have a single use of one technology instead of a cross-use of pedestrian re-identification and pedestrian attribute recognition.To solve the problems above,this thesis first proposes MGMHP-PR(Mask-Guided Method based on Human Parsing for Person Re-identification).Taking the inherent structure of human body into account,this thesis proposes KGMGCN-PAR(Keypoint-Guided Method based on Graph Convolutional Network for Pedestrian Attribute Recognition).Considering the real-world scenarios,in order to further improve the universality of the algorithms and the accuracy of target recognition,this thesis combines the two improved algorithms and designs a deep learning-driven video structured system,which can select different solutions independently in different situations.In the experimental part,the algorithms are validated and compared with the traditional benchmark algorithms.The main work of this thesis is as follows:(1)Aiming at the problem that fast location changes result in low recognition rate,a person reidentification algorithm based on human parsing,named MGMHP-PR,is proposed in this thesis.Firstly,this thesis introduces an improved residual network,multiscale feature fusion and Focal Loss to U-Net to generate human masks for different body parts;then this thesis calculates the weight matrix using human masks combined with attention mechanism to obtain pixel-level local features;then this thesis proposes a region-level triplet loss to synthetically calculate the local features of different body parts;finally,this thesis uses a Siamese network to jointly calculate the loss in instancelevel and region-level.(2)Focusing on the problem that the existing video structured technologies cannot make full use of the relationship between human inherent structures resulting in a high rate of attribute recognition errors,a pedestrian attribute recognition algorithm based on keypoints-guided graph convolutional network,named KGMGCN-PAR,is proposed in this thesis.Firstly,this thesis uses HRNet and residual network to extract the keypoint local features;then this thesis employs the inherent structure of human body to initialize the adjacency matrix,and adaptively train the adjacency matrix to convolute the keypoint local features;then this thesis divides the local feature into different parts for attribute recognition;finally,this thesis employs uniform content label(UCL)to label the structured video.(3)Concentrating on the problem that current video structured technologies only have a single use of one technology,this thesis combines the two algorithms to implement a deep learning-driven video structured system.The system can intelligently select different algorithms to structurally organize and manage the target information.The MGMHP-PR and the KGMGCN-PAR are validated and analyzed through relevant experiments.The experimental results show that the MGMHP-PR has higher recognition accuracy than other commonly used person re-identification algorithms,and the KGMGCN-PAR can achieve more accurate recognition of the attributes compared with the traditional pedestrian attribute recognition algorithms.