Sena Sukmananda Suprapto, Sholeh Hadi Pramono, Panca Mudjirahardjo, Fitri Utaminingrum, Mahdin Rohmatillah, Cries Avian
The increasing global adoption of weight training has coincided with a rise in injuries, primarily attributed to beginners’ limited training knowledge and the prohibitive costs of personal trainers. Existing posture monitoring systems, such as 2D pose estimation, often fail to generalize effectively across varying camera angles, underscoring the need for more adaptable solutions. This study presents a generalized AI-based framework for classifying weight training techniques as “correct” or “incorrect” to enhance exercise safety and reduce injury risks. The proposed framework leverages 3D human pose estimation, Strided Transformer, and Long Short-Term Memory (LSTM) networks, addressing angle dependence through the integration of a simple yet effective 3D rotational data augmentation technique. Due to unavailability of public dataset for evaluating the correctness of deadlift techniques, this work introduces a novel dataset consisting of 201 videos from 12 participants. Videos were recorded from two distinct camera angles and labeled as either correct or incorrect by experienced weight-training instructors. Three configurations were evaluated: (1) data from Camera 1, (2) data from Cameras 1 and 2, and (3) data from Cameras 1 and 2 with augmented virtual viewpoints. Results reveal that the third configuration achieved superior generalization, attaining an accuracy of 99.7% and an F1-Score of 0.997, significantly outperforming other configurations. These findings highlight the critical role of data augmentation in improving model generalization across diverse camera perspectives. © (2025), (Intelligent Network and Systems Society). All rights reserved.
Department of Electrical Engineering, Universitas Brawijaya, Malang, Indonesia; Department of Electrical Engineering, Institut Teknologi Kalimantan, Balikpapan, Indonesia; Department of Informatics, Universitas Brawijaya, Malang, Indonesia
