YogSaarthiAI: A Vision-Based AI System for Yoga Posture Analysis and Fitness Guide

Abstract

The efficacy of yoga practice is fundamentally tied to precise postural alignment; however, the absence of real-time, personalized feedback in modern digital training platforms frequently leads to suboptimal execution and potential musculoskeletal injury. To bridge this critical gap, this paper introduces YogSaarthiAI, an advanced, real-time virtual yoga assistant that leverages computer vision and deep learning to deliver highly accurate, full-body posture analysis and dynamic correction. Unlike conventional fitness applications that rely on opaque, computationally heavy models or provide limited binary evaluations, YogSaarthiAI utilizes a highly optimized, dual-path hybrid architecture. The system processes live webcam video streams through the MediaPipe framework to extract 33 distinct skeletal landmarks, which are mathematically transformed into ten resolution-independent, scale-invariant geometric angle features. A custom-trained Deep Neural Network (DNN) classifies the user's movements across nine fundamental yoga poses, achieving a robust test accuracy exceeding 95% with a minimal inference time of approximately 10 milliseconds. Operating in parallel, a deterministic template-matching engine compares live biomechanical vectors against expert-derived pose blueprints to generate graduated, joint-specific corrective feedback. A significant contribution of this work is the engineering of a novel Symmetry Mirroring Heuristic, designed to resolve unilateral body occlusion during lateral poses (such as the Plank and Bridge), which empirically reduced false corrective alarms from 40% to under 5% during live execution. Deployed on an asynchronous FastAPI and ReactJS web infrastructure, the platform sustains an end-to-end frame processing latency of 200 to 400 milliseconds, ensuring instantaneous instructional delivery. Comprehensive system evaluations and user acceptance testing confirm that YogSaarthiAI provides objective, pedagogically valuable guidance, establishing a scalable, accessible, and computationally efficient paradigm for automated biomechanical assessment.