FAST-AI: An AI-Based Biomechanical Analysis System for Tennis Serve Assessment

Abstract

The tennis serve is a high-velocity and complex movement that plays a critical role in performance and injury risk. Conventional biomechanical analysis relies on laboratory-based motion capture systems, which, although considered the gold standard, are costly, time-consuming, and lack ecological validity. Recent advances in markerless pose estimation and wearable inertial measurement units (IMUs) offer practical alternatives; however, these approaches are often applied independently, limiting validation and interpretability. This study proposes FAST-AI, a hybrid framework combining video-based pose estimation and IMU-based measurements for biomechanical analysis of the tennis serve. The experiment involved 8 tennis players (18-35 years old) conducting 15 tennis serves (N = 120), wherein kinematics data were analyzed using MediaPipe Pose, and IMUs provided reference data for rotations and velocities. Serve accuracy was measured as a binary value. Pearson correlation and error metrics were used to analyze the similarity between data produced by AI algorithms and those recorded with the use of IMU devices. Elbow joint angle showed statistically significant moderate negative correlation with the angular velocity of the forearm (r = -0.302, p = 0.0008) with relatively small absolute error values (MAE = 1.317°, RMSE = 1.614°), making up less than 1% relative error. On the contrary, wrist joint pronation did not show significant similarity (r = -0.084, p = 0.3618). Comparing groups using the Mann-Whitney U test did not show any significant differences between accurate and inaccurate serves; however, some small to medium effect sizes were observed. This can be attributed to the substantial class imbalance, where the accuracy ratio was at 94.2%. Overall, FAST-AI demonstrates the feasibility of a hybrid AI–IMU framework for extracting interpretable upper-limb kinematics in a controlled setting. While agreement is variable-specific, the system shows potential as a scalable, validation-oriented approach for sports biomechanics, rather than a definitive performance assessment tool.