Research On Classroom Behavior Recognition Algorithm Based On Deep Learning

In recent years,the development of deep learning technology very quickly,was a great success in text processing,speech recognition,object detection and other fields.At present,domestic and foreign scholars have achieved good results in many professional fields using deep learning technology.However,in the field of classroom teaching analysis,deep learning technology still lacks effective applications.In classroom teaching activities,identifying the behavioral status of students in the classroom is conducive to analyzing the efficiency and quality of students ’classroom learning and evaluating the quality of teachers’ classroom teaching.To this end we propose two kinds of classroom behavior recognition algorithms based on deep learning.The first is a classroom behavior recognition algorithm based on wearable sensors.In recent years,smart wearable devices have made certain developments,and human behavior recognition based on sensor data has been realized.In this dissertation,a CNN-GRU deep learning network model is designed by combining a convolutional neural network(CNN)and a gated recurrent unit(GRU).The algorithm collects the acceleration,angular acceleration,magnetic field strength,and heart rate data of the three axes when students make classroom behaviors through a device worn on the students’ wrists and analyzes them to realize classroom behavior recognition.Smart wearable devices mainly include a three-axis sensor,a three-axis magnetometer,a three-axis gyroscope,and a heart rate meter.The accuracy of the algorithm in the collected classroom scene data reached 95.26%.However,as the above algorithm requires the help of smart wearable devices for behavior recognition,it brings some cost problems and convenience problems.Subsequently,this dissertation proposes a classroom behavior recognition algorithm based on computer vision,which realizes behavior recognition by using the built-in cameras in the classroom.Based on the Faster RCNN target detection algorithm,this dissertation proposes two improvements: The first is to adopt a progressive detection structure with multiple thresholds,and gradually increase the Intersection over Union(Io U)threshold of each level of detector through the process of resampling.Improve the accuracy of each detector’s output a bit for use as the input for the next higher-precision detector,and improve the quality of positive samples while avoiding the problem of overfitting;The second is to introduce the Det Net-59 backbone network for feature extraction,and increase the network depth while ensuring the feature map resolution.The accuracy of the improved algorithm on the standard VOC dataset is 77.9%,which is 1.5% higher than the original Faster RCNN algorithm,and is better than most classic target detection algorithms.In addition,in this dissertation,a classroom behavior recognition data set is produced for classroom scenes,which contains 12,300 labeled sample images.The improved algorithm has an accuracy rate of 89.2% in this data set and has excellent detection results.