| With the wide application of information technology,artificial intelligence,big data and other technologies in the field of school education,smart education came into being.The change of traditional classroom teaching methods makes teachers’ classroom teaching evaluation face significant challenges,mainly including the pan scene of teaching evaluation space,the multi dimension of teaching evaluation content,and the intellectualization of teaching evaluation methods.Students’ classroom behavior is an important basis for evaluating teachers’ teaching quality and students’ learning level.However,the existing research on students’ classroom behavior recognition is faced with such problems as the lack of data sets,the complexity of model calculation and the difficulty of small target detection.Aiming at the above problems,this paper introduces and improves YOLO algorithm to detect targets,and conducts research around the problems related to student behavior recognition.The main research work is as follows:(1)Construct a data set of students’ classroom behaviors.First of all,we collected students’ real class videos from a university’s smart classroom,and obtained 18900 student classroom data by intercepting video frames.At the same time,we divided students’ classroom behaviors into seven types:listening,playing with mobile phones,reading,writing,standing,sleeping,and raising their hands.Secondly,we use semi-automatic data annotation algorithm and annotation software to create a data set of students’ classroom behavior.Finally,data enhancement was carried out before the training model to enrich the data set.About 430000 student goal annotations,80000 student behavior annotations and40000 student classroom behavior category annotations were obtained in total,which solved the problem of lack of student classroom behavior data set.(2)A student behavior recognition algorithm based on improved YOLOv5 is proposed.For student behavior recognition in small-scale scenarios,H-swish activation function is used to increase reasoning speed,and the original CSP structure is replaced with Ghost module to reduce the number of model parameters.In addition,attention fusion module is introduced to improve the recognition capability of the output.The experimental results show that the classification accuracy of the improved method is higher than that of the original model.(3)A learning behavior recognition algorithm based on two-stage detection is proposed.The first stage is the student target detection algorithm based on YOLOX.In order to improve the problem of mutual occlusion between students,an improved NMS candidate box suppression strategy is used,which eliminates candidate boxes that are highly coincident by calculating Manhattan distance.To solve the problem of undetected small target detection,this paper improves the feature fusion module,and adds a detection layer to enable the model to obtain more small target features,further increasing the small target detection capability.The second stage is the student behavior recognition algorithm based on Efficient Net,which is used to identify the student targets detected in the first stage.The experimental results show that the improved method of two-stage detection and recognition separation can enhance the detection ability of students in large-scale classroom scenarios,and the classification accuracy of students’ behavior is superior to other classification algorithms.(4)Based on the algorithm proposed in this paper,a student behavior recognition system is designed and implemented.This system includes three modules: classroom video upload and query module,behavior recognition module and result display module.Classroom video upload and query module can enable teachers to upload classroom videos on the client and query the detected videos;The behavior recognition module realizes the function of recognizing the behavior of all students in the whole class;The result display module is responsible for returning the results to the client and displaying them. |
PDF Full Text Request