Author(s): LIU, J., CUI, Y., Institution: BEIJING SPORT UNIVERSITY, Country: CHINA, Abstract-ID: 1766

INTRODUCTION:
Shooting is a multi-joint basketball movement where the temporal organization of action phases and proximal-to-distal sequencing of the kinetic chain influence release stability and accuracy. Traditional evaluations rely on expert judgment or laboratory-based protocols, lacking ecological validity and scalability for real-game settings. The study aims to: (i) automatically detect in-game shot events using markerless 3D skeleton-pose data from a multi-camera tracking system; and (ii) localize action-phase timing via a finite-state machine (FSM) and quantify kinetic-chain sequencing to derive process-level metrics that distinguish shot types and performance outcomes.
METHODS:
Six collegiate games were recorded using a synchronized 12-camera tracking system. Multi-view videos were used for 3D skeletal pose and ball-trajectory reconstruction. Shot events were identified through ball-flight physics and 3D rim constraints to identify apex, net-entry, and release frames, while kinematic rules filtered false positives (e.g., passes, lobs). The shooter was assigned based on hand-ball proximity and inferred offensive direction. A reverse-anchored FSM segmented each shot into four phases: gather, loading, setpoint, and release, using the net-entry frame as a temporal reference. Phase boundaries and key timepoints were validated against two experts' annotations (event-timing ICC=0.921). For kinematics, knee/hip/shoulder/elbow/wrist joint-angle time series were computed; Savitzky-Golay smoothing and differentiation produced angular velocity/acceleration profiles and peak timepoints. Proximal-to-distal kinetic-chain timing was quantified from peak timings and validated against expert labels (Cohen's kappa=0.961).
RESULTS:
A total of 630 shots were identified. Manual verification of two games yielded detection performance: Precision = 98.7%, Recall = 90.5%, and F1 = 94.4. Shot-type classification (five classes) achieved substantial agreement (kappa=0.89), kinetic-chain timing kappa = 0.901, and FSM timepoint ICC(2,1) = 0.974-1.000. For jump shots (Catch-and-Shoot and Off-Dribble; N=290), the made group showed greater release height than misses (2.84 +/- 0.57 vs 2.65 +/- 0.37 m; t=3.55, p<0.01; ES=0.43). For layups (N=288), made shots showed a shorter distance to the basket (t=-4.13, p<0.01). Power generation differed by shot type, with significant effects in peak knee angular velocity (F=29.72, p<0.01; Catch-and-Shoot 367 deg/s vs Layup 269 deg/s) and peak elbow angular velocity (F=15.18, p<0.01).
CONCLUSION:
This study presents an in-game framework for evaluating shooting mechanics through automatic event detection, phase segmentation, and kinetic-chain timing. The approach reveals performance-relevant biomechanical differences in jump-shot release height and layup release distance between successful and missed attempts and characterizes shot-type-specific power-generation patterns, providing objective metrics for skill development and real-time performance monitoring in basketball.