| With the accelerated pace of life and the normalization of epidemic prevention and control,people are more inclined to exercise at home.However,ordinary self-study methods lack professional guidance,and irregular or even wrong limb motions will lead to decreased exercise effect and may also cause physical injury.At present,motion analysis mostly relies on professional sensor equipment,and daily exercise is difficult to popularize.With the rapid development of computer vision technology,vision-based human detection technology is widely used in security monitoring,sports,games and entertainment.The application of human skeleton key point detection algorithm to limb motion recognition and evaluation can assist individuals to exercise and increase exercise interest.In this paper,human pose motion analysis is carried out from the perspective of computer vision.The main work is as follows:(1)The research status in the field of human action recognition and evaluation at home and abroad is analyzed,and the development process of two-dimensional human pose estimation algorithm based on vision is expounded.Open Pose is selected as the algorithm for pose estimation in this paper by comparison.In view of its complex network structure and large number of model parameters,a lightweight improved Open Pose network based on Oct-Mobile Net is proposed,and the data set is enhanced in the network training stage.The experimental results show that the calculation amount of the improved model is reduced to 12% of the original model,the detection speed is increased by 300%,and the detection accuracy is reduced by only 1.2%.(2)An attitude recognition method based on bone information and multi-feature fusion is proposed.The improved Open Pose algorithm is used to extract the limb key point information and normalize it.According to the characteristics of human motion,the spatial geometry and temporal motion features are designed,including coordinate position,joint angle,joint distance and motion speed.The features are fused by sliding window algorithm.Finally,the best combination of classifier and feature type is selected by comparative experiments.The recognition accuracy of LSTM-based fusion feature action classifier is 98.8% in self-made Eight-section brocade dataset and 97.2%on public dataset KTH.It is proved that the multi-feature fusion method can better carry out attitude analysis,and the method has certain universality.(3)Taking Eight-section brocade as the research object,an action sequence scoring method based on multi-modal information is proposed.The posture vector is obtained after the key point coordinates are preprocessed,and the chord distance between the vectors is used to represent the action similarity.The overall similarity between the test sequence and the template sequence is calculated by the dynamic time warping algorithm,so as to measure the overall similarity between the exercise action and the standard action.The action video is transformed into action label sequence by classifier,combined with the periodic characteristics of Eight-section brocade fitness action,a sequential pattern mining method based on interval length is proposed to calculate the periodicity and synchronization of exercise action and standard action.Finally,the application system of Eight-section brocade action recognition and evaluation is developed by integrating the results of the study.The system can recognize and quantify the signage score for the input video,and compare and display the joint angle difference between the standard posture and the practice posture,so as to assist the practitioners to adjust their postures,improve the exercise effect,and achieve smart fitness.This paper contains 63 figures,16 tables and 68 references. |
