Integrating multimodal biomechanical analysis and machine learning for key factors identification in standing long jump performance

The increasing adoption of AI-assisted physical fitness testing has underscored the need for precise computer vision-based sports movement analysis. This study presents a biomechanics-oriented multimodal analytical framework for standing long jump (SLJ) assessment, designed for practical video-based pose-estimation sequences. The framework addresses two key challenges: accurate detection of takeoff and landing instants and examination of the associations between multimodal variables and jump performance. A baseline-referenced event inference strategy integrating foot–ground interaction cues and ankle kinematics was employed to identify takeoff and landing moments from routine video sequences. A biomechanical characterization procedure integrating physiologically weighted center-of-mass (CoM) estimation, trajectory fitting, and interpretable multimodal association analysis was further employed to analyze potential performance-related factors. First, a dual-criteria detection algorithm, combining foot–ground clearance and ankle acceleration, is developed to identify representative takeoff and landing timestamps (e.g., 4.20 s and 4.50 s for Athlete 1). For the aerial phase, a physiologically weighted CoM trajectory (based on segmental mass distribution: head 0.07, torso 0.46, upper limbs 0.18, lower limbs 0.29) and parabolic trajectory fitting are employed, yielding trajectory patterns that remained relatively consistent under the adopted preprocessing configuration and with the adopted quadratic motion formulation across the analyzed high-resolution frame sequences. Second, a comprehensive feature set encompassing eight technical descriptors (e.g., takeoff angle, arm swing frequency) and individual physiological parameters was constructed. An interpretable multimodal association framework integrating these variables with an ensemble feature-importance mechanism was then employed to examine the associations between the analyzed variables and jump performance. The empirical analysis indicates that physical constitution variables exhibited comparatively stronger overall associations with jump performance, while takeoff angle and arm swing frequency showed comparatively stronger associations among the examined modifiable technical descriptors within the analyzed dataset. Rather than pursuing fundamental algorithmic modifications, the present study emphasizes an application-oriented multimodal integration framework that prioritizes coordinated biomechanical interpretation, offering a practical paradigm for AI-assisted sports training analysis.