Based On Spatial Temporal Convolution To Tai Chi Action Quality Evaluation Study

With the rapid development of social and economic development,people are facing health problems while their living standard is improving.Lack of exercise is a major factor affecting health.The fast-paced lifestyle limits the types of exercise that people choose.The fact that Tai Chi has the ability to be unrestricted by time and space gives it a wider audience.Artificial evaluation and instruction of tai chi’s actions relies on the experience and subjective judgement of professionals,which is costly and limitedly.Therefore,a convenient and objective method for assessing the execution of Tai Chi actions is of significant research value.The Tai Chi action quality assessment algorithm,utilizing skeleton and spatial temporal graph convolution,is proposed to counterbalance the influence of background noise on assessment models performance and the disregard for spatial and temporal semantic information comprehension of movement sequences in current actions assessment studies.The main tasks are as follows:(1)Improving pose estimation algorithm.Improving pose estimation algorithm.In this paper,a pose estimation algorithm is used to obtain Taiji action skeleton data,thereby eliminating redundant background information from the RGB video’s original source.A lightweight Open Pose skeleton data extraction algorithm is proposed to tackle the issues of low accuracy,intricate model,and an abundance of parameters in Open Pose skeleton network models.To begin with,Open Pose’s lightweight backbone network,the Micro Net,guarantees a high-quality extraction at a low computational cost.Repeated operations in the pose refinement network are then merged,while small-sized convolutional kernels are used in the network to reduce the amount of parameters and operations in the Open Pose model.Finally,joint information is extracted from Tai Chi action’s videos using lightweight Open Pose to construct a taijiquan skeleton action dataset(2)spatial temporal feature extraction of skeletal sequence data.Skeleton graph-based taijiquan action sequences are rich in spatial temporal semantic information.The single type attention matrix in ST-GCN,which performs well in human skeleton action sequence analysis,is not suitable for all data and network layers,and the one-dimensional temporal graph convolution learns a simple action time information.The adaptive multi-scale temporal graph convolution network model is proposed to solve the problems in ST-GCN in order to extract richer spatial temporal feature.The network first add an adaptive matrix and node relevance matrix that are learned jointly with the network training parameters,focusing on nodes with higher relevance to the target node when aggregating nodes belong of neighborhood.The temporal convolution is then extended using multi-scale windows to enrich the information features of the action sequences in the temporal domain.Finally,the impact of heterogeneous nodes on network performance is reduced through causal inference.It is also experimentally validated on NTU RGB+D dataset,Kinect skeleton 400 dataset and self-built taijiquan skeleton dataset.(3)Tai Chi action quality Assessment Network.A model of movement quality assessment,based on spatial temporal graph convolution,is suggested to tackle the issues of incomprehensibility and inadequate performance of current evaluation methods.The model simultaneously learns the similarity and quality scores between training and standard actions on the basis of a Siamese network structure,and demonstrates the effectiveness of the action assessment algorithm thought comparing it with the method of learning feature regression directly on deep neural networks.