Pose-derived biomechanical feature profiling and proxy pattern classification of basketball player movements

Markerless pose estimation provides a practical approach for extracting movement-related variables from sports video without requiring laboratory-based motion-capture systems. However, pose-derived biomechanical analysis must be interpreted carefully when datasets do not contain ground-truth action labels or independently validated biomechanical outcomes. This study developed a computational pipeline for pose-derived biomechanical feature profiling of basketball player movements using annotated gameplay pose data from the TrackID3 × 3 dataset. Frame-level keypoint annotations were processed to derive biomechanical variables, including elbow angles, segment distances, asymmetry measures, and trunk orientation. These variables were organized into short temporal sequences using a sliding-window approach, enabling descriptive analysis of player posture, limb coordination, subset-level variability, and sequence-level feature patterns. Rule-based proxy biomechanical groupings were constructed to facilitate exploratory evaluation of pattern separability within the pose-derived feature space. The pipeline generated a structured pose-derived dataset of 37,134 frame-player observations and showed measurable variation in joint alignment, limb symmetry, and trunk orientation across players and recording conditions. Classification results should be interpreted as proxy-label separability rather than validated prediction of basketball actions or performance outcomes. Future work should incorporate action-level annotations and independent biomechanical validation.