A Research On Machine Learning Based Autism Children Afftective Recognition

Autism is a disease that manifests as a barrier to social communication and a repetition of sensory movements.The prevalence of autism continues rising globally,and the number of children with autism is increasing.There are papers showing that the earlier the diagnosis is made,the better the intervention will be.Nowadays,experts majoring in diagnosis and treatment of autistic children continue to propose new ideas and insights,and information technology,especially machine learning is constantly being integrated.The treatment of children with autism in China is mainly based on intervention,but there’s a lack of sufficient experienced teachers.Therefore,this paper focuses on information technologies and makes a system to measure the affect performance of autistic students in the classroom based on the actual application scenarios,hoping that the students’ affective can be effectively judged by analyzing the videos of children with autism in learning.The system focuses on the individual training classroom of autism children,and builds the affective evaluation indicators by modeling the actual teaching block.The feedback given by children with autism in the classroom is mainly presented by actions and expressions,which is the key points of this study.The designed indicators include the engagement score about actions and the affective score about the expressions.The former is based on the time of action response,and the latter is based on the facial expression model.The input data is videos about students in learning.Before the scoring,the action and expression of the autistic children need to be recognized first.Therefore,the core of the project can be divided into two parts.The first part is action and expression recognition.Considering the problem of data transmission and storage,along with the development of feature point extraction method based on attitude estimation,it is better to extract coordinate features from original data before scoring.Recent years,the graph convolutional neural network has performed well with Non-Euclidean Structure data.Therefore,the system is mainly based on the convolutional neural network.In the experiment,the matrix composed of frame sequence and key point coordinates is taken as input,and the spatial temporal graph convolutional neural network is used as the model of action recognition.The accuracy reaches 90%,and then,the Euclidean distance and cosine characteristics are analyzed.After that,the calculation method of the start frame is designed and the engagement score can be seen based on the action response time.During the experiment of expression recognition,it is found that the random forest can train the feature point sequence of the frame and get the accuracy of 99%,so the random forest is chosen as the main classifier of expression recognition.The scoring method is designed after analyzing the relationship among arousal,valence and expression category.