Improvements To The YOLO Algorithm And Its Application To Student Classroom Behaviour

Deep learning algorithms are constantly evolving,accelerating the development of intelligence in different fields,and in some areas have surpassed the human level,building smart classrooms with a very bright performance.In traditional teaching,the quality of classroom learning is ensured through teacher observation of student behaviour,which is both time-consuming and inefficient.Therefore,the application of deep learning algorithms to classroom behaviour detection is a question that researchers have been thinking about.In this work,we combine classical deep learning algorithms in the detection of students’ classroom behaviour and improve them to increase the accuracy of the algorithm so that classroom behaviour recognition becomes more easily applicable in real life.This thesis focuses on the following aspects:(1)A dataset of student classroom behaviour is constructed.As the dataset involves privacy,the data is collected through online public classes and processed with appropriate mosaic and other techniques to ensure data security.The collected video data were processed through image processing to create a training dataset SCBD(Student Classroom Behaviour Data),which includes five types of student behaviour,namely raising hands,standing,writing,skimming and listening to lectures.(2)The research methodology was identified and improved accordingly.Different deep learning target detection algorithms were compared and finally the YOLOv5 deep learning network outperformed other neural network target detection algorithms with its average accuracy of 65.1%,so YOLOv5 was identified as the base network for the subsequent research.By adding an attention mechanism to the backbone network,the neural network directly calculates the useful feature vectors,avoiding the impact on model accuracy due to information such as occlusion and small targets,and the improved model is named A-YOLOv5(Attention-YOLOv5).The model improves accuracy while throwing away the interference of useless features on the fitting results.The m AP of the model reaches 78.3%,which is 13.2 percentage points higher compared to the m AP of YOLOv5.(3)To further improve the accuracy of the model,A-YOLOv5 was improved again.In this thesis,the neck of A-YOLOv5 was improved by citing the Efficient-Rep GFPN structure and making corresponding improvements to obtain the new model S-A-YOLOv5(Simple-Attention-YOLOv5).The m AP of the S-A-YOLOv5 model was improved from78.3% to 81.9%,which is the same as that of A YOLOv5,the accuracy of the S-A-YOLOv5 model improved by another 3.6 percentage points,and the accuracy of the S-A-YOLOv5 model improved by 16.8 percentage points when compared with YOLOv5.