Research And Application Of Class Action Recognition Based On Skeleton Data

Class action recognition through the analysis of students' action in the class environment to form effective teaching feedback.It can improve the quality of classroom teaching and student learning effect.The traditional teaching evaluation method is unitary and relies on subjective experience.By applying computer vision technology and artificial intelligence method to analyze students' class action.The results are more objective.Also the efficiency of action recognition can be improved.This paper conducts an in-depth study on the two aspects of multi-person interactive group discovery and individual action recognition based on the students' action skeleton data obtained from the classroom environment.The contents are as follows:(1)To solve the problem of multi-person interactive group discovery in the class environment,an interactive group discovery algorithm based on skeleton node trajectory aggregation model was proposed.Firstly,in order to reduce the impact of body size and initial position on 3D bone data collection,skeleton data were standardized.Secondly,in order to describe the trajectory information of skeleton action,a trajectory optimization method of speed regulation was introduced,which used the estimated skeleton action speed to adjust the optimal contraction position.Finally,the aggregated skeleton trajectories were clustered to realize the discovery of interactive population.In the experimental part,class measured data were used to verify the results of interactive group identification of the method and comparison method in this paper.The results show that the track-based method has a good accuracy in the discovery of multi-person interactive group.(2)To solve the problem of limited number of videos collected in class and high computational efficiency,a multi-kernel SVM action recognition method based on artificial features was proposed.The distance and rotation angle between skeleton coordinates were defined as the characteristics of skeleton data,and PCA was used to reduce the dimensionality of the high-dimensional data.Finally,the feature of data was classified by multi-kernel SVM classifier,which made full use of the feature that the individual motion state satisfies the nonlinear manifold space.In the experimental part,the effects of the proposed method and the comparison method were verified by using the public test data and the class measured data respectively.The results showed that for the data of nonlinear manifold structure,a better recognition rate could be achieved on the medium-scale data set by using the kernel based nonlinear classifier.Compared with the manual feature extraction method of Lie Group and ST-GCN,the accuracy of the measured data set was 43.4% and 53.5% higher.At the same time,the operation efficiency of this method was high,which can meet the real-time performance requirements of motion recognition of middle school students in the class environment.(3)A prototype system of classroom behavior evaluation based on bone data was designed and implemented according to the method proposed in this paper,which could realize real-time posture recognition and evaluation of classroom video collection.The system functions include teacher login,classroom recording,classroom evaluation,and display through the image interface.