Human Motion Similarity Assessment Based On Pose Estimation And Its Application In Sports Training

With the advent of the information age,the field of human motion recognition and analysis has made Significant progress and breakthroughs,aided by massive amounts of data and high-performance computing devices.As one of the crucial research branches in the field of human motion recognition and analysis,human motion recognition technology has developed rapidly with the aid of the powerful feature learning ability of deep learning.Despite initial accomplishments in human motion recognition technology,there has not been a highly effective solution for the problem of human motion similarity assessment,particularly regarding human motion similarity assessment from different perspectives,within the field of human motion recognition and analysis.Meanwhile,with the improvement of living standards,people are increasingly involved in sports,and utilizing human motion similarity assessment technology to achieve normative sports motion assessments holds tremendous practical significance.Against this background,this Thesis focuses on the task of human motion similarity assessment and studies the problem of human motion similarity assessment based on human pose estimation technology.The main tasks of the research are as follows:(1)The traditional human motion similarity assessment method based on Euclidean distance has problems such as inconsistent human skeleton size and different time axis.To address these issues,this Thesis proposes a human motion similarity assessment method based on human joint Angle features and DTW algorithm.This method utilizes human joint angle features as the fundamental characteristic to assess human motion similarity and combines the DTW algorithm to perform temporal alignment on the human joint angle feature sequences,thus effectively eliminating the impact of inconsistent human skeletal size and time axis differences on human motion similarity assessment and obtaining more reliable and accurate human similarity assessment results.(2)This Thesis proposes a deep metric learning-based method for assessing human motion similarity under different camera perspectives.The method maps the human skeleton sequence to three low-dimensional latent spaces using an autoencoder deep neural network model.It extracts a motion information feature vector that is independent of camera perspective and human skeleton structure information,which is used as the basic feature for human motion similarity assessment.Experimental results show that this method effectively eliminates the effects of camera perspective factors and differences in human skeleton size on human motion similarity assessment.It can generate more reliable and subjective similarity evaluation results for human movements with different camera perspectives or inconsistent skeleton sizes,which are closer to human subjective perception.(3)Combined with the actual application scenario of sports training,this Thesis takes the standing long jump motion as the research case.Based on the aforementioned method for assessing similarity in human motion,a software system was designed and implemented to analyze and assess the standardization of standing long jump motion.The system provides functions such as skeleton key point extraction and preprocessing,assessment of motion standardization and identification of incorrectly positioned limbs.This system can assist trainees in completing standing long jump movements more standardly and improve their athletic performance.