AI-Driven Vision-Based Fitness Coaching Using Adaptive Skeleton Matching and Real Time Feedback System

View at publisher Summarise with VIRAT
Authors: Charan M, A. Mary Mekala, Manoj D, Mohammad Alsamad Dandu Shaik, Shaik Mohammed Suhail
Cited by: 0
References: 11

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The growing adoption of unmonitored home workouts has increased the demand for reliable, real-time posture correction systems that improve form and reduce potential injury risks through real-time guidance. Current vision-based solutions mainly use two types of methods: angle-threshold techniques and deep learning-based approaches. While threshold-driven methods are simple and easy to use, they struggle with different camera angles, body types, and joint occlusion. These methods also need manual adjustments for various joint limits. On the other hand, optimized neural network-based pose estimators can function on regular devices. Nevertheless, deep-learning posture correction systems require labeled datasets, exercise-specific adjustments, and added costs for analyzing sequential motion over time, making them less practical than lighter heuristic options for quick multi-exercise use. This paper presents a real-time fitness coaching system that melds canonical skeleton alignment with a single joint-angle tolerance. Instead of adjusting several thresholds for different joints, this system saves reference positions for key exercise positions. It aligns these to user body types using rotation, scaling, and translation transforms. The system checks posture correctness by comparing joint angles to a chosen baseline tolerance of 30 degrees. Temporal consistency mechanisms provide stable recognition of movement progression, continuous corrective cues with highlighted joint deviations, and fully automated repetition counting with audio feedback. Experimental evaluation on multiple participants performing bicep curl exercises demonstrates 85% posture-matching accuracy and 88% feedback relevance, measured through stage-level correctness and manual verification of feedback relevance.

Fields of Study & Subjects

Human Pose and Action Recognition Context-Aware Activity Recognition Systems Ergonomics and Musculoskeletal Disorders

Publication details

Journal / source
International Conference on Computing, Communication, Control and Cyber-Physical Systems (I5CPS)
Pages
1–6
Publication date
26 February 2026
Language
EN