Research On Classroom Abnormal Behavior Recognition System Based On Target Detection Algorithm

With the continuous enhancement of China's economic,social and comprehensive national strength,whether the quality of education can keep up with the pace of the times has become a key factor for China to move to a higher level.In recent years,with the increasing scale of college enrollment and the popularization of entertainment electronic products,it is necessary to further improve the supervision of college classroom quality.At present,it is not enough to supervise students' classroom behaviors only by manually viewing classroom video data and teachers.The above-mentioned methods need to consume teachers' classroom energy,inconvenient and time-consuming operation,very low efficiency,easy to miss detection,false detection and other situations,and can not detect students' sleeping,playing mobile phones and communication behaviors in real time.It is of practical significance to apply the technology of deep learning to the supervision of classroom quality in Colleges and universities.This paper mainly studies how to recognize the abnormal behavior of students by using video monitoring in the classroom scene of colleges and universities,and how to help teachers judge the abnormal behavior of students by using computer vision technology.In this paper,we will use the combination of hardware and software to realize the design of the classroom quality assistant system.The main work of this paper is as follows:1.From the application scenario of the system,it should have the characteristics of low cost and high efficiency.Therefore,the system uses embedded technology such as system transplantation,arm linux application development,etc.to design the equipment terminal based on TQ2440 development board and realize the real-time sampling of camera.The real-time image data collected by the equipment terminal(camera)is compressed and encoded by using x264 encoding library,and then transmitted to the server.The main function of the server is to use ffmpeg and H.264 to decode the received video data in real time,then to recognize the image in real time,and to store the video data.When the image processing module detects the abnormal behavior of students,it will transmit the image data to the mobile app terminal of the teacher,and alarm by pop-up message box.In this paper,the system server students and mobile phone target detection and abnormal behavior recognition of students will be the focus of this paper.2.The target detection of students and mobile phones in the image processing module of the system server is the basis for identifying abnormal behavior of students in the classroom.The system will return the position information of students and mobile phones in the image during the process of target detection.Because the YOLOv3 algorithm has outstanding performance in the detection of small targets such as mobile phones,and the speed of processing image data is faster,it is more in line with this design.The algorithm design of this paper adopts the Tensor Flow architecture.After the original image is input into the trained YOLOv3 network,the target target recall rate and accuracy rate of the students after feature extraction are 95.0% and 92.3%,respectively.86.9% And 93.5%,the recall rate and accuracy rate of the overall target detection reached 90.95% and 92.9%,respectively.3.Recognition of students' abnormal behaviors in class is the ultimate goal of image processing module.For the three kinds of abnormal behavior recognition,the method of combining depth network and traditional manual features is used.The depth network is mainly used for target detection.The traditional manual features mainly define the characteristics of three kinds of abnormal behaviors,and the three processes are synchronous for behavior recognition.For the behavior of "playing mobile phone in class",the method of bounding box regression is used,and the idea of non maximum inhibition is used to see that the student's bounding box overlaps the maximum area of the bounding box of the mobile phone target,then the student has the behavior of playing mobile phone.For the behavior of "sleeping in class",we use Haar feature to detect the face of each detected Student object.If no face is detected,we conclude that the student has sleeping behavior.The behavior of "class communication" adopts the way of setting threshold.When the overlapping area of two or more students' goal boundary boxes is larger than the set threshold,the group of students is judged to have communication behavior.Experiments show that the comprehensive recall rate and accuracy rate of the three behavior recognition are more than 80%,and the detection speed of the three behaviors is within 1 second,which is suitable for the real-time requirements of the system.4.In addition to detecting abnormal student behavior in the classroom,a complete classroom abnormal student behavior detection system should also have a corresponding real-time alarm function.The alarm module of the system is mainly that the APP of the teacher's Android mobile terminal receives the JSON abnormal behavior data of the server and stores it in the Handler message queue,and then uses the Android internal mechanism to let the APP realize the alarm in the form of sound effects,vibrations and pop-up message boxes.Experiments show that the APP terminal can realize the real-time alarm function when students have abnormal behaviors.