Abstract

Incorrect movements during marching exercises among older adults can result in acute injuries and chronic complications, highlighting the need for precise and real-time feedback from exercise specialists to mitigate further risks. This study proposes an automated detection system designed to address these challenges by leveraging a convolutional neural network (CNN)based pre-trained model integrated with a cosine-based formula algorithm derived from the dot product. The system extracts anatomical movement data from key joints-specifically the hip, knee, and ankle—to calculate joint angles θ, thereby emulating the evaluative capabilities of exercise specialists. Furthermore, augmented reality (AR)-based techniques are incorporated to visualize the results, offering a practical solution to the scarcity of physiotherapists. The proposed approach is experimentally validated through a study involving 90 elderly participants, 30 in controlled laboratory settings and 60 in real-world environments. The system demonstrates robust performance in detecting marching postures across diverse body weight statuses and genders, achieving F-measure scores exceeding 98%. These findings suggest that the system is highly suitable for real-world applications, providing a self-assessment tool that enables older adults to exercise safely and effectively by simulating continuous professional supervision.