| Action teaching is a method of teaching motor skills,which is widely used in sports training,dance teaching,medical rehabilitation and other fields.Traditional methods are manually guided and corrected by coaches,but due to the different experience and skill levels of coaches,trainees may learn wrong action postures.In addition,the feedback of traditional methods is not precise enough to help trainees find and correct errors in time.Therefore,how to accurately acquire and compare students’ action postures is key to achieving action teaching.This thesis proposes a Transformer-based algorithm for human pose comparison.The algorithm introduces a self-attentive mechanism that enables it to better capture the spatial location and motion trajectory information of 3D pose sequences,thus allowing the algorithm to more accurately distinguish between 3D pose sequences of standard and non-standard actions.The algorithm uses a random masking strategy to train the Transformer encoder,a pre-trained model,to learn the 3D pose sequence features of continuous actions,enabling the algorithm to better represent the 3D pose sequences of new actions,thus further improving the algorithm’s ability to distinguish between 3D pose sequences of standard and non-standard actions.The algorithm uses a voting strategy to classify 3D pose fragments and vote the results to produce more accurate human pose comparison results.In order to objectively evaluate the effectiveness of the above method,this thesis designed and generated a home-made dataset JXNU-Skeleton using a 3D human posture estimation device to detect and track the 3D coordinates of human joint points with high accuracy.The dataset includes 3D posture sequence data and bone length data for ten movements of three aerobics teachers,three martial arts teachers and 15 students.In addition,this thesis proposes a 3D pose sequence data pre-processing method,which aligns standard and non-standard 3D pose sequences of the same action,and uses the aligned 3D pose sequence data as experimental data for the human pose comparison algorithm,so that the human pose comparison algorithm can better learn the pose characteristics of standard and non-standard actions.The experimental results show that the human pose comparison algorithm proposed in this thesis achieves better results on the home-made dataset,with an average accuracy of 0.8254 on ten actions,which is an improvement of 8.63% in the average accuracy compared to the LSTM(Voting)method;an improvement of 29.66% in the average accuracy compared to the method without the voting strategy;an improvement of 29.66% in the average accuracy compared to the method without the Compared with the method without pre-training strategy,the average accuracy of this algorithm is improved by 27.72%.Based on the proposed human posture comparison algorithm,this thesis designs and implements a action teaching scoring system based on the WPF framework.The system mainly includes functions such as action recording,user uploading videos and viewing scoring results,which can help teachers better evaluate students’ aciton performance and provide more effective action teaching aids for students. |
