Intelligent Recognition Of Student Behavior In Classroom Video Based On Deep Learning

In recent years,with the rapid development of artificial intelligence,The emergence of intel-ligent cameras greatly reduce the labor costs.Behavior monitoring as a hot topic in computer vision,drives the intelligent video detection technology with the monitoring device popular-ity and rapid development.Since today's students are very individual and more difficult to manage,if a behavior-recognition camera is installed in the classroom and in the exam room,the camera can automatically capture the student's behavior with an alarm to ask students correct their behavior.Because of the widespread application of convolutional neural networks in computer vision,in this thesis we will use iDT and TSN algorithms to determine whether the behaviors similar to students in class meet the regulations including several actions.The difficulty lies in monitoring the complexity of video scenes,and the timeliness and accuracy of student action judgments.This article mainly studies the following issues:(1)The basic theory of behavior recognition is studied,including the basic knowledge of how to perform feature representation and deep learning.It also briefly introduces the basic algo-rithm and process steps of CNN.Using these methods to cut the video,that is,how to extract dynamic and static features from the video that can better describe the video judgment.(2)This thesis studies the basic algorithm of iDT,uses optical flow field to obtain some trajectories in the video sequence,and then extracts the four characteristics.Using Fisher Vector method encode the features,and then trains the SVM classifier based on the encoding results.(3)Using the TSN algorithm,for specific scenarios of student behavior,including students'listening,eating,playing mobile phones,talking,etc.,the dense optical flow is calculated for every two frames in the video sequence and the dense optical flow sequence is obtained.Then the CNN is trained separately The model,after making a judgment,directly fuses the class score of the network's feature to obtain the final classification result.