The Research And Application Of Action Detection And Recognition In Online Video Surveillance

Computer vision is a focus area in the current wave of artificial intelligence.With the development of computer hardware systems and deep learning theory,intelligent algorithms can be used to realize the visual function of the human eye.Nowadays,the advanced research results obtained through the unremitting efforts of scholars in the industry have been applied in reality to help solve people's life problems.In recent years,the rapid advancement of deep learning technology has played a huge role in the field of computer vision.Among them,the most prominent convolutional neural network has achieved remarkable achievements in image classification and object detection with its powerful performance.Further research is aimed at video data,and the surveillance cameras are everywhere in the living or working scenes.The massive monitoring video generated every moment only relies on manual monitoring,which consumes huge manpower,financial resources and material resources to realize intelligentization.Video surveillance will effectively solve the above problems and meet people efficient and convenient daily needs.This paper focuses on the deep learning technology to achieve multi-model fusion action detection and recognition algorithm,and improve the action recognition algorithm based on Two-Stream convolutional network,the specific work is as follows:(1)Studying the data processing and analysis from the long-term untrimmed video,propose a multi-model fusion action detection and recognition algorithm.The temporal region proposals of the input video according to the regular interval,and combine the action detection algorithm to generate the binary classification of background and action.Then the action start and end time will be clarified.The action sub-segment is sent to the action recognition algorithm to implement the class classification of the action.In this paper,the self-built database is used for experiments.The results show that the selected sub-segment has a recall rate of 76% when the IOU value is 0.5,and the optimal performance is achieved.It can be applied to practice to achieve real-time detection.(2)Improving the action recognition algorithm of Two-Stream convolutional network,and the network model based on visual attention mechanism is proposed,AttsNet.The model extracts the spatial features and temporal features of the video sequence separately,trains the network with two different forms of input data,and attaches the Attention weight to the feature map channel,so that the spatial stream network has higher recognition ability for the video content information.Finally,the scores of the two morphological networks are combined to obtain action categories.The improved AttsNet model has stronger learning ability on the self-built dataset,and the experimental results show that the recognition accuracy is 95.6%,and the recognition accuracy is 93.7% on the UCF-101 dataset.(3)Designing intelligent online video surveillance system,which is supported by multimodel fusion action detection and recognition algorithm to achieve real-time monitoring of surveillance video under working scene.It can also alarm the abnormal or dangerous actions that occur in the monitoring,instead of the manual monitoring work.