Standardized Evaluation of Athlete Movements Based on Computer Vision and Cnn Algorithms

To address the drawbacks of subjective scoring, subpar motion detail capture, and cross-athlete standardization barriers of manual observation or basic sensor data, this paper proposes using computer vision and Convolutional Neural Networks (CNNs) to develop a motion standardization evaluation method, to be used in place of subjective scoring. In order to make the keypoint detection process more accurate, the optimized pose estimation model incorporates anti-occlusion map constraints and time smoothing constraints. In this model, individual body shapes are eliminated in a normalization mechanism, and the dynamic motion features are obtained in 3D convolutional networks and optical flow weighting. Standard motion templates constructed based on motion semantic alignment and high-dimensional feature vectors enable precise matching between model scores and standard motions. At a PCK10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−1</sup> threshold, the matching accuracy for weightlifting motion reaches 0.95, while the matching accuracy for rotational motion is slightly lower at 0.90. The results of the experiments prove that the model detects with an accuracy of more than 0.90 at a PCK at 0.1 threshold. At the same time, keeping a low variance of scoring in the different lighting and occlusion conditions, the model demonstrates the capability to identify the difference between the motions of the athletes of different skills. This offers the reliable technical assistance to the intelligent training monitoring and motion evaluation through the obtaining of the accurate control of the motion details and cross-individual scoring standardization.