Key Technologies Of Moving Target Detection And Pedestrian Structural Description From Surveillance Video Based On Deep Learning

Recently,with the progressing of major security projects,such as national emergency systems,safe city,safe campus and technological police,video intelligence analysis is playing an increasingly important role in security field.Moving target detection,which is the basic and core function in video intelligent analysis,is both the hotspot and difficulty in the research of video intelligent analysis.Many functions of video intelligence analysis depend on the results of moving target detection.Pedestrians are the most concerned objects in the field of security,and pedestrian attribute recognition is widely used in the field of video intelligence analysis.At the same time,with the continuous innovation and breakthrough development of deep learning technology,deep learning technology has been widely used in the field of video intelligence analysis.This paper studies and implements moving target detection,pedestrian component detection and fine pedestrian recognition based on convolutional neural network,which is deep learning technique.The main work of this paper is as follows:(1)This paper proposes a moving target detection network based on YOLOv2-832 network,which is an improvement on YOLOv2 network.This newly proposed network has a higher recall rate and IOU value in target detection,and also extends the idea of YOLOv2 network by using Anchor mechanism combined with K-means clustering to determine a suitable size of the moving object candidate region,thereby performing position prediction and category classification on the extracted candidate region.(2)This paper studies and implements pedestrian component detection network.Based on the research of target detection algorithms for R-CNN,this paper proposes a pedestrian component detection network based on F-RCN model,and uses ResNet50 as a shared deep convolutional network for RPN and detection network to achieve the precise positioning detection of head,top and down body,bag and feet.Pedestrian component detection network provides high-accuracy pedestrian component data for subsequent fine pedestrian recognition.(3)This paper studies and implements fine pedestrian recognition.Multi-task learning mechanism has been introduced into the classification of pedestrian attributes and a pedestrian multi-attribute classification network is designed.DenseNet121 network is used as the main network of the multi-attribute classification network to complete categories and attributes discrimination of pedestrians' gender,hair,pants and bags.What's more,considering that color feature information can be learned in shallow network,this paper proposes a color classification network based on AlexNet to classify the color attributes of pedestrians' coats,pants,bags,and shoes.At the same time,ResNet50 and VGG16 are selected as the classification network of the attribute and categories of pedestrians' coats and bags.Based on the results of the pedestrian component detection,this paper completes the fine pedestrian recognition and achieves the structural description of pedestrian.Finally,this paper introduces the system of moving target detection from surveillance video and pedestrian structural description and builds Caffe and DarkNet deep learning platforms for this system to provide a computing platform for moving target detection and pedestrian structure description.Display interface of the system is built by Qt to display the results of moving object detection and pedestrian structural description in real time.