Sports Motion Analysis : From Competition Videos to Data-Driven Interpretations скачать в хорошем качестве

Sports Motion Analysis : From Competition Videos to Data-Driven Interpretations

Qi Gan's PhD presentation Qi Gan. Sports Motion Analysis : From Competition Videos to Data-Driven Interpretations. PhD thesis. Institut Polytechnique de Paris, 2025. Understanding sports motion is essential for performance analysis and training guidance, and has become a growing research area with the advancement of artificial intelligence (AI) in data analysis. Additionally, due to its analyzable physical mechanisms, sports motion serves as a valuable case study for other motion-related domains, such as pedestrian prediction in autonomous driving, abnormal behavior detection in surveillance, disease diagnosis in medicine, and pose tracking in gaming. However, the black-box nature of modern AI models limits our ability to understand their behavior and decisions. This thesis aims to bridge the gap between sports motion analysis and explainable AI (XAI), leveraging AI's representational power while ensuring interpretability. Two main challenges are addressed: (1) the lack of high-quality sports datasets—despite many online videos, competition footage often suffers from low resolution, motion blur, and noisy backgrounds; and (2) the limited research on interpreting sports motions, with few baseline studies in this area. To tackle these, we focus on long jump, which offers two advantages: the availability of high-quality world-class competition videos and biomechanically well-defined motion sequences suitable for interpretation. To support this work, we built three datasets: (1) Olympic triple jump videos with 2D poses and official distances, (2) World Championship long jump videos with 2D poses and biomechanical features, and (3) long jump videos with 2D poses and jump distances from top level competitions. Our study focuses on two levels of data extraction. First, we estimate athlete poses from video, despite challenges like low frame rates and rare poses. We improve pose accuracy with a post-correction method using 2D sports pose priors, modeled as Neural Distance Fields (NDF) in polar coordinates and trained with a gradient-projection-based augmentation method. Second, we estimate biomechanical features from video using a data-free method, reconstructing athlete trajectories from poses via biomechanical modeling.We explored two paths for interpreting sports motion. The first focuses on biomechanical features: we use a classical interpretable machine learning pipeline by training a quantile random forest regressor and applying SHapley Additive exPlanations (SHAP), Partial Dependence Plots (PDPs), and Individual Conditional Expectation (ICE) plots. This reveals insights aligned with existing literature. However, to capture widespread feature interactions, we propose a new method to estimate not only interaction strengths but also where and how interactions occur.The second path focuses on interpreting pose sequences. We analyze black-box time-series models by generating counterfactual explanations using a sparse autoencoder-based model. Experiments show that this approach yields both faithful and robust explanations, contributing to more interpretable and practical human motion analysis from video data. https://theses.hal.science/tel-05291306/