AI-based human pose estimation in Fitness: A Review of posture correction, repetition counting, and range-of-motion methods
Abstract
This review examines the application of AI-based markerless human pose estimation (HPE) to exercise coaching, focusing on four essential functions: posture correction, repetition counting, range-of-motion (ROM) analysis, and exercise recognition. The review investigates which methods are most effective for each task and whether real-time solutions that scale across a variety of exercise types and application domains are available. Following a PRISMA-guided scoping protocol, 43 studies published between 2020 and 2025 are included. To map the tasks and methods of each study and to summarize their key findings, a master comparison table was made, and datasets, benchmarks, and reported hardware/FPS metrics are analyzed. To construct accurate, scalable, and end-to-end physical exercise coaching systems, this paper identifies present constraints and suggests future research areas. Datasets, failure modes (occlusions, user variation), and edge hardware real-time performance are analyzed. The review finds that AI-based systems may offer correct feedback for certain exercise sets in controlled circumstances, but dataset bias, small sample numbers, and unstable generalization to real-world home environments limit their efficacy. The paper outlines future directions, including hybrid architectures with deep backbones and interpretable rule-based heads, self-supervised learning, personalization and domain adaptation, uncertainty-aware feedback, and multi-task frameworks for posture, ROM, counting, and exercise recognition.