Pose guided predictive ballistics for body part–targeted football training

Abstract

This thesis presents the design, implementation, and evaluation of a vision-guided ball-launching system for body-part-targeted football training. Conventional sports ball-launching machines operate in open-loop mode: a trainer manually sets the launch angle, speed, and interval, and the machine repeats the same trajectory regardless of the athlete’s position. This work addresses that limitation by developing a perception-guided system that detects and tracks an athlete in real time, reconstructs selected body-joint positions in three dimensions, and computes the launcher aim parameters needed to direct a ball toward a chosen target joint. The proposed system uses four fixed low-cost USB cameras installed in a domestic garage arena and calibrated using ChArUco boards and AprilTag fiducial markers. Real-time ball and human-pose detection are performed using YOLO-based models, while multi-view triangulation reconstructs 3D positions in the arena coordinate frame. A constant-velocity Kalman filter smooths the measured joint positions and predicts their short-term motion to compensate for perception and actuation delay. A ballistic solver then converts the predicted target position into pitch, yaw, and wheel-speed commands for a custom Ball Launching Machine controlled by an ESP32-based firmware state machine. The system also includes a safety-gated runtime architecture, structured decision logging, emergency-stop handling, and an offline voice-command interface for hands-free target selection. The system was evaluated through static ball-localisation tests, human joint-localisation trials, dynamic tracking experiments, and supervised live integration tests. A 36-point static ball ground-truth grid achieved a corrected mean 3D localisation error of 95.17 mm, while an 81-trial joint-touch protocol with 62 valid trials achieved a mean 3D joint-localisation error of 143.38 mm. Both results satisfied the accuracy thresholds defined for the project. The integrated live test demonstrated that the system could aim at and fire toward named body joints under operator supervision while maintaining safety constraints. Fully autonomous closed-loop firing at a moving subject remains the main future integration milestone. Overall, this thesis demonstrates that low-cost multi-camera perception, real-time pose estimation, predictive tracking, and safety-gated actuation can be combined into a practical prototype for adaptive football training. The work contributes a complete pose-to-launch pipeline, a quantitative validation methodology, and a foundation for future research on intelligent sports-training machines that respond to athlete movement rather than relying on fixed pre-programmed trajectories.