Research On 3D Human Posture Estimation Algorithm In Treadmill Fitness Scen

Three-dimensional human pose estimation could help treadmill users obtain more accurate motion data,such as step count,stride length,and posture,thereby providing data support and enabling users to scientifically formulate fitness plans.Therefore,the algorithm for three-dimensional human pose estimation in the treadmill fitness scenario was studied in this paper,and the performance of the algorithm was improved by utilizing temporal information and human body topology.Intelligent step counting strategies and improper posture recognition strategies were designed,incorporating pose analysis techniques to achieve step count statistics and reminders for improper postures.To address the issue of inaccurate three-dimensional pose estimation caused by occlusion,a grouped human skeletal structure was utilized for body modeling,and a grouped spatial-temporal attention network was designed for three-dimensional human pose estimation.A spatial encoder was designed to extract local inter-joint relationships within the region and global inter-region relationships using self-attention mechanisms,increasing the model’s understanding of spatial joint information,and improving the possibility of predicting rare poses.A temporal encoder was designed to capture the temporal relationships of the input frame sequence,obtaining pose representations with temporal features,thereby improving the occlusion problem.Qualitative and quantitative analyses were conducted on the Human3.6M dataset,proving that the algorithm could improve the accuracy of three-dimensional human pose estimation.Qualitative analysis and visualization of results were performed on the treadmill fitness dataset,demonstrating the effectiveness of the proposed method for estimating three-dimensional human pose.To address the issue of excessive redundancy in video sequences with similar poses,this paper designed a three-dimensional human pose estimation method based on interframe information fusion.Based on stride convolution,a temporal encoder was designed to gradually aggregate local temporal information,enabling the model to focus on predicting intermediate frames,thereby reducing the model’s parameter count and improving training and optimization speed.Qualitative and quantitative experiments on the Human3.6M dataset demonstrated that the algorithm could reduce the parameter count,decrease computational costs,and improve computational speed.Qualitative analysis and visualization of results on the treadmill fitness dataset confirmed the effectiveness of the proposed method.The three-dimensional human pose estimation algorithm was applied to step counting and improper posture recognition.An intelligent step counting strategy was designed to track the number of steps during a certain period.An improper running posture detection strategy was designed to real-time detect improper postures such as "leg swinging forward," "bent back and hunched posture," and "incorrect arm swinging" in videos.The designed methods were validated on a self-made dataset,proving the feasibility and effectiveness of the methods,and displaying results in real-time.