A Method and System for Real-Time Acquisition of Player Speed and Attacker-Defender Distance in Football Video Using Deep Learning (Patent Application No. 10-2026-0018076)

Purpose: This study aims to develop a deep learning–based system, FootballVideoProcessor, designed to automatically extract real-time player speed and attacker–defender distance from standard football broadcast video. The system eliminates the need for dedicated wearable tracking equipment or specialized multi-camera setups. Method: The system architecture consisted of five modules: ObjectTracker for YOLO-based detection; KeypointsTracker for homography-based coordinate transformation using Direct Linear Transformation (DLT) and RANSAC algorithms; ClubAssigner for RGB-based team classification; BallToPlayerAssigner for identifying ball possession; and SpeedEstimator for real-time velocity calculation (distance÷time×3.6) with moving-average smoothing. Nearest-opponent distance was computed using Euclidean distance in real-world coordinates. System optimization incorporated batch-size tuning, GPU parallelization, and a threaded queue pipeline to ensure processing stability. Results: The system achieved a mean average precision (mAP) of 0.934 and maintained a processing speed of 24 frames per second (FPS). Team classification accuracy ranged from 89.4% to 97.8%. The output video generated real-time annotations for each object, including unique IDs, bounding boxes, team colors, trajectories, speed, and nearest-opponent distance. Representative performance data included player ID 13 recorded at 31.05km/h with a 5.88m gap to the defender, and ID 4 at 31.56km/h with a 2.75m gap. Conclusion: The proposed system enables reliable real-time extraction of player speed and inter-player distance from standard football video and has potential applications in tactical coaching, match performance analysis, highlight generation, and player load monitoring.