Research On Group/Individual Behavior Analysis Technology Based On Video Scene

With the continuous improvement of surveillance / camera coverage,people can get more and more convenient and fast access to a large number of video information,such as classroom video,community gate monitoring and so on.The shortcomings of the traditional manual monitoring and judgment methods are more and more obvious: low efficiency and slow speed of watching video;People's attention is easy to be distracted;The labor cost is too high.And through the action recognition algorithm of artificial intelligence technology,the recognition rate is too low and the recognition result is not credible in complex environment.This paper uses deep learning technology for automatic video recognition and analysis.The accurate actions of the characters in the video can be obtained by joint point recognition and action recognition,and the action information and identity information of the same person can be contacted by gesture Association.This method improves the efficiency and accuracy of multi person recognition in complex environment,and puts forward the evaluation index of classroom and other scenes,which provides a theoretical basis for the construction of intelligent classroom and the automatic evaluation of teaching.In this paper,firstly,the scene modeling is carried out,and the characteristics of common complex video scenes are summarized.Then the specific scene of video is selected according to the characteristics.Combined with the actual needs of the subject,the classroom scene and hall of building scene are selected as the research scope.According to the common actions in the specific scene,we set the action classification and screen the joint point model of the characters,and select the appropriate evaluation index according to the specific background of the scene.Then the classroom video is sampled to speed up the recognition efficiency without affecting the recognition effect.Open Pose is used to analyze the video joint points,and the coordinate information flow of the scene joint point model is obtained.Then,the multi person joint information flow is divided into single person joint point information flow by the way of inter frame attitude correlation,and the continuity is checked.Then,the information flow is preprocessed according to different baseline methods and sent to LSTM network for single person action recognition,and the single person action flow is obtained.The face image of each person's pose is intercepted,and the MTCNN + CNN network is used to recognize and analyze the intercepted image to obtain the individual's identity information;Then,the group action data is obtained by summarizing the single action flow,and the scene evaluation index is calculated.Finally,a visualization system is implemented to demonstrate the algorithm.Because the occlusion of multi person video will seriously affect the work of action recognition,we need to establish a mechanism to improve the robustness of occlusion.In this paper,we use the method of randomly adding occlusion to improve the occlusion rate of the data set,expand the scale of the data set,and interpolate the occlusion in a short time to reduce the impact of occlusion on the recognition effect.