Research on feature extraction methods for aerobics jump actions based on machine vision

In order to have an accurate data analysis of aerobics jump actions, this article presents a multi-feature fusion method for feature extraction utilizing machine vision. In the video preprocessing phase, we mix background subtraction with human pose estimation to get the athlete's region of interest (ROI) automatically and apply spatiotemporal coherence to pick and enhance key frames. We propose a multi-channel feature extraction model, which combines geometric features with temporal features and a weighted fusion strategy that effectively integrates information from different dimensions. To ensure that the algorithm runs in real-time, we present the code, the complexity optimization, and the storage of particles that is outlines in great detail. Therefore, the human head's movement has changed the past time that the person behind the direct the behavior of the person and a lot of toe head occur so the movement create the consequent movement. Detailed pseudocode is provided for the algorithm’s implementation, along with complexity-optimization schemes for feature computation and data storage to ensure real-time performance and stability. We evaluate the proposed method on a dataset comprising various jump types under different lighting and background interference conditions using accuracy, recall, and F1-score as metrics. Experimental results demonstrate that our method outperforms traditional single-feature extraction algorithms in both jump detection and feature description, achieving an average F1-score improvement of 12.3%, with strong robustness and real-time capability. Finally, we discuss the method’s advantages and limitations and outline its potential applications in action scoring and sports rehabilitation.