Vision-based Human Motion Correction System

With the development of Internet technology,the role of computer vision has become increasingly prominent.Among video-based human motion comparison analysis attracted attention in areas such as intelligent monitoring,and it can also be applied to video teaching.At present,video action teaching generally uses instructional styles,and the teaching effect is limited.Using computer vision-based methods can provide learners with real-time action correction and obtain the results of action comparison analysis and evaluation.In this regard,the learner only needs to imitate the motion in front of the camera,and then can feedback the motion correction information in real time,and finally obtain the analysis result of the motion comparison.The vision-based human motion correction system selects sports fitness as the application scenario.The system records pre-recorded standard sports videos.After the learner selects this item,the user makes the same motion in front of the camera,and the system corrects the motion in real time,After the end of the exercise,a motion comparison analysis is given.Through the multilayer iterative neural network(CNN)architecture,the results of human pose estimation are obtained,and the position information of human bone joint points is obtained.The collection of skeletal data is regarded as an action sequence,and the difference between the two action sequences is compared as a comparison of the features in the action sequence;the direction of the limbs formed by adjacent joints is used to perform real-time action correction.Then,the dynamic time warping method is used to align the action time series,based on the feature vector and the cosine similarity to determine the action similarity.Finally,the angular motion is used to represent the motion trajectory to obtain the overall action comparison analysis result.The main content of this paper is: Based on the human pose estimation,a human motion correction system is designed.During the exercise,the user performs real-time motion correction,and gives the analysis result of motion comparison after the exercise.Real-time motion correction is based on the position data of human bones and joints,and uses adjacent joints to form the direction of the limb to perform real-time motion correction.Specifically,the coordinates of the joint points of the to-be-detected action are estimated based on the standard action.When the bone data and estimated values of the joint points of the to-be-detected action are within the allowable range,it is judged to conform to the standard action.The sign(positive /negative)of the error between the estimates,to determine the direction of deviation of the action to be detected.Since the difference in the action sequence can be seen as the difference in features in the action sequence,for the action sequence formed by the position data of the bone joints,a suitable feature is selected to describe the process of the movement difference.In order to perform action comparison analysis,considering the existence of action sequences with different lengths,first use dynamic time warping to align the action timing;then extract the feature vectors of the aligned action sequences and use cosine similarity to determine the action similarity.The angular feature describes the motion trajectory and obtains the analysis result of motion comparison.The final experimental results show that the system has good results.This system can capture the user's learning data through the camera,compare the standard video with the action to be detected,so that the learner can correct the action in real time,and give the action comparison analysis result after the action is completed to better guide the user to learn and provide Good teaching platform.