Research On Motion Training System Based On Vision

With the rise of the national fitness boom,more and more people are actively participating in daily training.The current main training methods are classroom teaching and video recording teaching.Although these two methods are simple,they have problems such as time and place limitations,neglect of interaction with users,and lack of scientific exercise guidance for training.It is difficult to achieve ideal results.The training methods of the traditional method are generally implemented based on depth cameras and wearable devices,which have problems such as high cost,complex operation and poor portability,which are difficult to be widely promoted.In response to this situation,this paper designs and implements a training system based on skeleton key point information.The system extracts the human skeleton information in the video through the bone keypoint detection technology,and can automatically identify the type of training exercise and provide guidance.Using an ordinary monocular camera has a good recognition accuracy,simple operation,low price,and is suitable for multiple scenarios.Its realization principle is: using bone key point detection technology to extract bone information in training sports video;using bone information to construct spatiotemporal map,and identifying fitness actions through graph convolution network;performing action alignment through joint feature vectors in bone information,And evaluate the key postures in the action;finally,based on the above results,the training system is implemented.The main work completed in this subject is as follows:1)Construction and optimization of fitness behavior skeleton dataset: Referring to the two existing behavior datasets,the OpenPose algorithm is used to extract skeletal information from the recorded fitness exercise video,and save it according to a specific format specification.A skeleton dataset of fitness behaviors for pulling and push-ups,and using gradient filling to optimize abnormal data in the dataset.2)Research on fitness action recognition based on Spatial Temporal Graph Convolutional Networks(ST-GCN): First,use the skeleton information of the human body to construct a spatiotemporal skeleton map,and use the map to reflect the spatiotemporal features of the key points of the human skeleton.The two-dimensional image convolution method is extended to the spatiotemporal skeleton graph to perform spatiotemporal graph convolution operations,and the attention mechanism is used to further improve the accuracy of the model,realizing the recognition of squat,deadlift and push-up actions.3)Attitude evaluation based on dynamic time warping: For the situation that the speed of the same action is different,first select a specific joint feature vector and apply it to the dynamic time warping algorithm to align the action sequence,so that the action posture and the standard posture can be consistent.One-to-one correspondence,and finally use the angle features of the joints to evaluate the training action pose.4)Design and implementation of the training system: The system improves the interaction with people by collecting training action videos,can automatically identify the types of human training actions,and automatically extract the key poses in the training actions for comparison with the standard poses.Analysis and suggestions for improvement are given,which improves the intelligence of training.At the same time,the system is inexpensive and simple to operate,breaking the time and place limitations of traditional training methods.