A Human Pose Estimation Method Based on the BlazePose Model

Objective: This study aims to evaluate the performance of the BlazePose model in human pose estimation for sports actions to improve the accuracy and computational efficiency of the model. Methods: The research employed a variety of datasets for training and evaluating the model, investigating the generalization capability of the model under different experimental conditions. Detection and tracking were carried out through the two-stage mechanism of the BlazePose model, identifying the human figure contours and initially predicting the key points, and then refining the positions of the key points through the tracking module. Different parameters such as model complexity and key point smoothing were set, and the steps including image preprocessing, feature extraction, key point detection, naming, and skeleton construction were conducted. Results: The research outcomes yielded the two-dimensional keypoints of single-frame human body images and the construction of 3D coordinate models, and obtained the visualization line charts of PCK experimental data. Through the comparison of PCK scores of different models, it was concluded that on the basketball Dataset, the performance of BlazePose was superior to that of OpenPose, particularly in terms of FPS performance. Although the PCK score of the lightweight version of BlazePose was slightly lower, its FPS performance was higher, making it suitable for scenarios with high requirements for speed. Furthermore, BlazePose effectively reduced the model complexity by offering models of different complexities and using occlusion simulation in training, without sacrificing too much accuracy. Conclusion: This research presents the efficacy of lightweight neural networks in real-time human pose estimation and, through further experimental analyses, investigates the possibilities of enhancing their performance advantages and application effects.