DEVELOPMENT OF A REAL-TIME PULL-UP VALIDATION SYSTEM FOR COMPETITIVE SPORTS USING YOLO26N-POSE

Abstract

This paper describes the development and experimental validation of an automated real-time pull-up counting system for competitive sports, built on the YOLO26n-Pose model. The system integrates 2D keypoint detection, geometric feature extraction, exponential smoothing of elbow-angle measurements, and a rule-based finite-state machine to count repetitions against competition-compliant criteria. A two-level validation architecture separates visual feedback, four discrete states: DOWN, TRANSITION, UP, and CHEATING, from the underlying repetition-counting logic. A calibrated floor-contact detector prevents repetitions from being credited when the athlete’s feet touch the ground. Running on a standard CPU-only laptop (Intel i5, 8 GB RAM), the system achieves roughly 25 FPS interface refresh. Unlike wearable sensor setups, the system requires no athlete-worn hardware, making it practical for judging at small competitions and training sessions.