Karate Kata Scoring and Performance Evaluation Using Video Analysis and Deep Learning

Abstract

Karate kata, a fundamental element of traditional martial arts, consists of choreographed sequences of stances, strikes, and blocks performed against imaginary opponents, requiring precise posture, timing, and technique. However, self-practice often lacks objective feedback, leading to unnoticed errors and uneven skill development, and even competition scoring can be influenced by subjective judgment. To address this, we present a vision-based scoring system that evaluates kata from structured video recordings without wearable sensors, providing accurate, consistent scores to help practitioners track progress. The system processes videos through a multi-stage pipeline: frames are extracted at 10 frames per second using OpenCV and segmented with YOLOv8-seg to remove background clutter. Pose estimation is performed with MediaPipe, generating 99 normalized 3D keypoints, while hip-centered normalization and pelvic-width scaling ensure consistency across different body sizes. A modified ResNet50 classifies frames into nine fundamental stances with high accuracy and near real-time processing (~150 ms per frame at 1080p), and a custom Convolutional Neural Network evaluates the angular accuracy of 14 critical joints using trigonometric analysis against reference poses. Each kata is decomposed into 19 key positions, with a stance marked correct only if both classification and joint angles (≤ ±2.5° deviation) meet the threshold, and final scores are calculated on a 5.0–10.0 scale to align with traditional judging. Experimental results show that ResNet50 achieves 95.97% accuracy, while the CNN reaches 96.11%, demonstrating that this approach offers a low-cost, accessible, and consistent scoring tool. It supports remote training, reduces judging bias, and provides practitioners with reliable feedback to improve their kata performance.