Automating Wrestling Judgement with AI Vision: A Multi-Angle Deep Learning Approach

This work introduces an AI vision system for automated wrestling judgment that integrates multi-angle video synchronization with deep learning models for detection, pose estimation, activity recognition, and temporal classification. The system includes detection, pose estimation, activity recognition, and temporal classification with spatio-temporal models such as 2D/3D video CNNs (MoviNet, SlowFast, I3D), Frame Wise CNNs, and sequence models (LSTM-transformer hybrids). The introduced system includes three major breakthroughs: (1) a dynamic multi-camera synchronization system to minimize occlusion with clearer motion, (2) an interpretability layer that helps the referees see the timelined explanation and step-by-step description of activity, and (3) a human-AI refereeing system that provides real-time scoring with the flexibility of overriding by explanation from human referees. The system is optimized for low latency to ensure realtime assistance during matches. The proposal integrates automated scoring with a human oversight mechanism that is envisioned to make wrestling refereeing more fair, consistent, and transparent with applicability to sports analytics.