A Deep Learning-Based Framework Using CNN+LSTM for Karate Kata Classification and Correctness Evaluation

Abstract

This paper presents a deep learning-based framework for automated classification and correctness evaluation of karate kata sequences using human pose data. The method utilizes MediaPipe BlazePose to extract 3D body keypoints from video frames, which are subsequently transformed into 132-dimensional vectors and temporally normalized into fixed-length sequences. Two neural architectures are evaluated: a baseline Convolutional Neural Network (CNN) and a hybrid CNN combined with Long Short-Term Memory (CNN+LSTM). Both models perform dual-output predictions: multi-class kata identification and binary correctness classification. Experimental results demonstrate that the CNN+LSTM model performs better in classification accuracy and movement assessment, with up to 95.74% accuracy and F1-scores exceeding 95% for several kata classes on unseen data. The findings highlight the importance of temporal modeling for structured movement analysis and establish a foundation for intelligent martial arts evaluation systems.

Access to Document

Other files and links

• Link to publication in Scopus
• Open Access Version Available

Fingerprint