Research On Detection Of Abnormal Behavior In Classroom Video Based On Deep Learning

The continuous development of science and technology provides a platform and guidance for network technology to promote the progress of human civilization.In today’s information age,the widespread application of surveillance video has made intelligent video an unprecedented development.However,the market’s demand for more humanization has not been met。For example the application of real-time behavior recognition and detection in some fixed scenarios,behavior alert systems applications.More and more people are concerned about developments in the field of video detection to identify and carry out scientific research and analysis.In the aspect of behavior detection,compared with the time-consuming and laborious traditional methods,the convolutional neural network algorithm because of its high speed,accurate identification advantage,causes many related fields researchers to shift the research center to this direction.This paper detects the abnormal behavior of students in classroom video,analyzes and researches the project according to the functional requirements,and the builds a detection model based on deep learning.Due to experimental needs,this paper will make a self-made behavior data for the project,improve the accuracy of small target recognition and solve the occlusion problem in target detection by improving and optimizing the behavior detection algorithm.In order to verify the effect of the proposed model in this paper,the proposed model is trained and tested on the public data and the self-made classroom video behavior data.And the results are analyzed and compared.The main research work of this paper is as follows:(1)Aiming at the situation that the open data set on the network can’t meet the need of this project,this paper collects the video data by recording the actual classroom scene,and uses the professional LabelImg tool to make the behavior tags to meet the experimental needs.(2)According to the actual needs of the students’ abnormal behavior recognition and detection in the classroom background,this paper is based on the convolutional neural network structure of deep learning,selects Yolov3 with real-time features for target detection,and makes full use of the shallow information to represent the feature information of small targets.The first residual block of Darknet-53 part is convolved with the feature map after 8 times down sampling through the hollow convolution of the feature information,and then cascade the improved RFB module,that is,add a branch to the RFB to perform peripheral features on the basis of the original module,a branch is added to extract the peripheral features,thus increasing the reference to the peripheral field of view and making the model more accurate.In order to make full use of the feature information in the pyramid,a convolution of 3*3 with a step size of 2 is used instead of the maximum pool operation in the upper sampling of the lateral connection.This article named the model Rs-YOLOv3.Through the optimization of the original network,the small target recognition effect has also been improved,and the missed detection rate has been reduced to a certain extent.(3)In the classroom scenes,students are easy to be occluded by other students in the real world.With the deepening of the neural network layers,the information will be lost in feature extraction.This paper will combine Res2Net network to the receptive field of each layer of the network,express feature information in a more granular manner,realize multi-layer feature reuse,and make the designed network more suitable for the scenarios in this paper.At the same time,considering that the frame regression loss function used in YOLOv3 cannot accurately describe the intersection ratio relationship between the frames.In this paper DIoU＿Loss is proposed,which increases the reference of the overlap ratio and the scale,and makes the model regression more stable.Through theoretical analysis and experiments,it is known that the behavior detection convolutional neural network is suitable for students’ abnormal behavior detection in classroom videos,and has faster recognition speed and higher recognition accuracy.At the same time,this paper is of great significance to the further research of human behavior detection algorithms in similar complex and fixed-scene videos.