Định danh và nhận dạng vận động trong thời gian thực sử dụng mô hình học sâu

The increasing demand for advanced artificial intelligence (AI) and deep learning solutions in physical fitness monitoring, training, and evaluation highlights the limitations of traditional, manual assessment methods, which often suffer from subjectivity, human error, and lack of scalability. Addressing this gap, the development of integrated systems for real-time identification and action recognition becomes paramount. Such systems are crucial not only in sports and education but also in medical rehabilitation and smart security. This paper introduces an innovative deep learning-based framework designed to automate physical fitness assessment. It combines robust facial identification with precise movement analysis, offering an objective and scalable approach to evaluate and track user performance in real-time environments.

Researchers and practitioners in artificial intelligence, deep learning, computer vision, sports science, physical education, and healthcare technology interested in automated fitness assessment and human activity recognition.

This paper presents an integrated deep learning system for real-time identification and action recognition, specifically tailored for automated physical fitness assessment. The proposed framework features two core functionalities: secure user authentication via facial recognition and precise analysis of human movements. For identification, the system utilizes CNN-based feature extraction to generate discriminative facial embeddings, ensuring reliable verification against pre-registered identities. Concurrently, for action recognition, the system employs Mediapipe to extract skeletal keypoints from video streams in real-time. These keypoint sequences are then processed by LSTM (Long Short-Term Memory) models to accurately classify and count repetitions of common exercises, such as gập bụng (sit-ups) and chống đẩy (push-ups). The research showcases a high-accuracy, stable, and scalable system capable of deployment in educational, athletic, and health monitoring contexts. Experimental results demonstrate the superior performance of LSTM compared to traditional methods in recognizing and counting repetitions, achieving high accuracy (e.g., ~99% for sit-ups and ~91% for push-ups). By seamlessly integrating identity verification with physical activity analysis, the system ensures that performance data is accurately linked to the correct individual, enhancing transparency and data integrity. This approach represents a significant advancement in leveraging thị giác máy tính for objective fitness evaluation, reducing reliance on subjective human observation and opening new avenues for personalized and automated health management. Future work aims to expand the dataset diversity and incorporate advanced spatial-temporal models to address current limitations.