Methods Research On Object Detection Of Classroom Students’ Behaviors Based On CNN

As one of the important application fields of object detection technology,students’ behavior detection in classroom environment combines classification tasks and positioning tasks,reflects students’ participation in classroom teaching,and plays an important role in promoting the development of education and teaching information.The traditional object detection algorithm is difficult to meet the requirements in the aspects of high efficiency,big data processing,multiple intelligence and so on.At the beginning of the 21 st century,the object detection algorithm based on convolutional neural network is more and more widely used,and the detection effect exceeds the traditional object detection algorithm,and then becomes the mainstream hot spot.This paper focuses on the detection methods of five kinds of behavioral goals in the classroom scene,such as writing,listening,raising hands,sleeping and answering questions.This paper focuses on the multi-scale problems and the intra-class occlusion problems of students’ behavior detection.The main research work is as follows:(1)In view of the multi-scale problem of large difference between the front row and the rear row students in the classroom scene,this paper focuses on the two aspects of network structure and candidate frame size reconstruction.In this thesis,a multi-scale detection network based on SSD algorithm is proposed for multi-scale problems with large difference in size between front row and back row students in classroom scene.On the one hand,the basic feature extraction network is designed as a convolutional structure of cascaded multi-expansion rate,which uses multi-scale receptive field to replace the single receptive field of the original network,and the residual unit is introduced to increase the correlation between adjacent feature layers;On the other hand,the SYM-FPN detection network with symmetry structure is designed for the lack of shallow detail information in the deep feature map of the feature pyramid,and then the multi-scale feature fusion is carried out after the top-down and bottom-up structure of the six sized feature maps,which makes each of them contain the details and semantic information of the other five sized feature maps.In addition,the scale and width of the open data set in the field of object detection are different from the scale of the classroom student behavior data set made in this paper Therefore,aiming at the reconstruction of candidate frame size in multi-scale problems,this thesis proposes a candidate box reconstruction algorithm based on K-means clustering to solve this problem.The algorithm uses Intersection-over-Union distance instead of euclidean distance as the similarity measure of K-means algorithm,and designs the fitness function based on Intersection-over-Union distance.The genetic algorithm is used to optimize the initial center point of clustering.Using the candidate box setting after clustering can add good prior information to the detection,increase the matching degree between the prediction box and the real box,and improve the detection effect to a certain extent.(2)Aiming at the problem of class occlusion of students’ behavior in crowded classroom situations,this thesis proposes a candidate box comprehensive screening algorithm based on center distance and penalty function to improve the occlusion problem and complete the detection and positioning of students’ behavior based on the study of multi-scale students’ behavior detection.When the overlapping area of the two candidate boxes is greater than a certain threshold,the center distance of the two boxes is used to determine whether the candidate box is a redundant box for different behavior or for the same behavior,and then decide whether to carry out exponential weight penalty or suppress the redundant box.This algorithm solves the problem that the behavior of the occluded students in the Non-Maximum suppression algorithm is easily suppressed because of the large area of overlap with other students.The Focal Loss classification loss function is introduced,controlling the proportion of positive and negative samples,reducing the relative loss of easily classified samples,so that the network can focus on training difficult samples to further improve the performance of the whole model in detecting the behavior of occluded students.The experimental results show that the proposed algorithm can effectively improve the detection accuracy of the five kinds of students’ behavior,and can meet the real-time detection requirements,and the detection accuracy of the self-made classroom students’ behavior data set is 86.5%,and the detection speed is 37 FPS.In addition,the proposed multi-scale detection network and the improved occlusion algorithm have a good generalization ability on the Pascal VOC data set,and the detection accuracy is 81.5%,which is better than that of other contrast methods.