Enhanced YOLOv8-Based Recognition of Yoga Hasta (Hand) Mudras Using Attention Mechanisms for Smart Healthcare Applications

• Yasir Altaf,
• Abdul Wahid,
• Mudasir M. Kirmani,

Vol. 20, No. 1, January 31, 2026
10.3837/tiis.2026.01.017, Download Paper (Free):

• Human activity recognition; Yoga hand mudras; Computer vision; YOLOv8; smart healthcare; Deep Learning.

Abstract

Human activity recognition in computer vision has gained significance in healthcare, especially in yoga, where the correct recognition of hand mudras aids in the physical and mental benefits of the individual. Nevertheless, real-time recognition of intricate gestures is difficult because of user-to-user variation, occlusions, and lighting or background differences. In this study, we present an enhanced lightweight object detection framework for real-time yoga hand gesture recognition based on the YOLOv8n architecture augmented with a Convolutional Block Attention Module (CBAM). The integration of the CBAM improves the model’s ability to focus on salient spatial features, such as subtle hand shapes and partially occluded gestures, while maintaining a low computational overhead. Our proposed CBAM-YOLOv8n model achieved a mAP@50 of 99.12% and mAP@50–95 of 79.45% on the test set. These compelling results, combined with its demonstrably lightweight architecture (significantly fewer layers, parameters, and GFLOPs compared to the baseline YOLOv8n and even other standard YOLOv8 variants), clearly surpass the performance of the baseline YOLOv8n and many State-of-the-Art object detection models, thereby providing robust evidence in support of our claims. The hyperparameter was optimized using Bayesian optimization. We curated a fully annotated dataset of ten essential yoga hand mudras from publicly available YouTube videos under diverse conditions to aid our research. The results of the present study indicate that the model can be employed in real-time applications for digital health, yoga teaching, and interactive wellness platforms. Code is available at: This GitHub Link.

Statistics

Cite this article