Research On The Detection Method Of Classroom Behavior Patterns In Classroom Scenes

In recent years,big data smart classrooms have become the focus of discussion in the education field.Applying artificial intelligence technology to create an intelligent classroom learning environment is an innovative teaching model that will be implemented in various schools in the future.With the increase in computing power of computer hardware and the advent of the era of deep learning,behavior detection technology will be widely used in fields such as smart classrooms and cloud education.When the existing object detection algorithms are adopted for behavior detection of the surveillance image of the classroom scene,problems such as false detection,missed detection,and inaccurate positioning will occur.To solve these problems,this thesis constructs a new classroom scene image data set,and proposes three new behavior detection methods.The main research contents of this thesis are as follows:1.This thesis constructs a new classroom scene image data set.Through observing common behaviors in daily learning classrooms,this thesis has identified 13 behavior categories,and completed the image collection work through a variety of ways.Since each image contains multiple different classroom behavior patterns,there is no such data set in the existing public data set.Therefore,this thesis analyzes the new data set from the perspectives of the number of images,the behavior annotation boxes,and the annotation box areas,which proves that the classroom scene behavior data set constructed in this thesis can serve the behavior detection task.2.This thesis studies the behavior detection method of geometrically distorted objects based on spatial transformation.In view of some inherent geometric distortions caused by the imaging characteristics of the camera in the classroom scene,this thesis performs spatial transformation operations on the distorted object features,i.e.,predicting the transformed parameters to perform pixel resampling on the object to offset or reduce the Influence caused by distortion,and increase the consistency of object behavior characteristics.3.This thesis studies a multi-scale attention behavior detection network based on feature pyramid.In view of the limitations of multi-scale features fusion when feature pyramid network is employed,this thesis proposes to adopt multiple scale features for adaptive resampling and concatenating and the concatenated features are weighted and fused through the attention of the channel and space dimensions.This method considers the importance of features of different scales to different size objects during fusion,and further enhances the expressive ability of features on the basis of feature pyramid.4.This thesis studies the behavior detection algorithm based on the object positioning loss weighted by the intersection ratio parameter.Starting from increasing the training weight of difficult samples,this thesis first introduces the development process of the existing positioning loss function.Then a positioning loss function weighted by the Intersection over Union parameter is proposed on the basis of the commonly used loss functions.Finally,a set of appropriate loss function parameters is determined through experiments.In this thesis,the above methods are trained and tested on the constructed data set.The experimental results show that the method proposed in this thesis can better describe the behavior characteristics of the classroom scene and can effectively complete the behavior detection task of the classroom scene.