Research And Implementation Of Action Detectionbased On Joint Deep Learning

In the rapid development of artificial intelligence and automation operations,so that at the moment when artificial intelligence and automation operations are rapidly developing,the current video surveillance system can no longer meet the needs of society,and gradually develops toward the direction of intelligence.In the process of development,human behavior detection technology based on deep learning has become a hot topic for researchers at home and abroad.Human behavior detection refers to the identification of human behavior while also positioning the human body.Based on the existing behavior detection technology,this thesis studies the behavior recognition and human body location by means of deep network by means of the powerful feature extraction ability of deep learning,and proposes a new behavior detection algorithm.For behavior recognition,because behavior has time characteristics,when character learning of video,it is necessary to learn spatial characteristics such as behavioral gesture distribution,and also learn the temporal characteristics of behavior,making behavior recognition more challenging than object recognition;For human body positioning,it is a special case of object detection,which mainly predicts the external contour of the human body,and requires less object classification than ordinary object recognition.This thesis focuses on behavior recognition algorithms based on improved convolutional neural networks and behavior detection algorithms based on multi-task joint learning.The specific work is as follows:1.Based on the improved residual network(TP3D ResNet)and the behavior recognition network of the Two-stream framework,the improved residual network is used as the feature extractor,and the Two-stream is used for human behavior recognition.For the T-TP3 DResNet network architecture.The two-stream network of this network architecture adopts TP3 D ResNet,which learns the appearance and time domain of the behavior in the video through 3D convolution,and uses the bilinear model to fuse the network.Due to its depth and good learning ability of the two-stream network for video features,this network architecture has greatly improved the performance of the behavior recognition.2.Based on the T-TP3 DResNet behavior recognition network,this paper adds a human-oriented network structure and designs a network architecture based on joint multi-task learning.This network architecture can not only perform behavior recognition but also human body positioning.By introducing multi-task joint learning in behavior detection,training in data sets in different domains alleviates the problem that the current behavior detection data sets are insufficient and the scale is small.Therefore,this network solves the problem of joint learning of heterogeneous data,and also provides a new method for video and image joint learning.3.According to the market demand of the society,this paper designs an intelligent anomaly behavior monitoring system based on C/S architecture.In this system,the background behavior detection algorithm is based on the multi-task joint learning behavior detection algorithm designed in this paper,and the intelligent behavior server and database server are designed to serve the client's operation.The distributed management method is applied to the system.The embedded system operates.