Research Of Pedestrain Retrieval Based On Orthogonal Semantic Feature Decomposition

With the development of science and technology and the promotion of social safety aware-ness,surveillance video is widely deployed in daily life.However,only relying on manpower to analyze a large amount of video data cannot achieve high accuracy and real-time process-ing capability.Therefore,it is of great significance to design intelligent analysis algorithms for surveillance video.This paper focuses on the Object Detection Problem and Pedestrian Re-identification Prob-lem of intelligent surveillance video.Based on Convolutional Neural Network,this paper pro-poses an end-to-end pedestrian retrieval coupling model.For the pedestrian Re-identification Problem,the identity features and the identity-invariance features are highly coupled.This coupling makes the model being interfered by irrelevant infor-mation during the training process,which affects the re-identification accuracy.In this paper,the Pedestrian Semantic Feature Model is proposed.The Schmidt-Orthogonal Feature Decom-position technique is used on the depth feature-map to separate the identity feature from the identity-invariance feature.Representation Learning Strategy and Metric Learning Strategy are used to train decoupled identity features respectively.Experiments show that feature separation in depth network effectively improves the recognition accuracy of Pedestrian Re-Identification Model.In order to realize the end-to-end pedestrian retrieval model,the YOLO-v3 object detection model and the improved pedestrian re-identification model share a common feature extraction network.The common feature extraction network will cause the detection model and the re-identification model to interfere with each other,affecting the accuracy of detection and iden-tification.This paper introduces deep semantic feature model into common feature extraction network and designs Task Separation Layer.Experiments show that the Task Separation Layer effectively separates the common and individual feature of pedestrians and solves the contradic-tion of mutual interference.