AI Vision-based Solo Training Assistant

The global shift toward independent fitness training and home-based rehabilitation has necessitated the development of automated, vision-based coaching systems. This paper provides a comprehensive survey of the state-of-the-art in Human Pose Estimation (HPE), specifically focusing on the integration of MediaPipe BlazePose and hybrid deep learning architectures. We analyze the evolution of motion tracking from foundational geometric heuristic models (2020) to contemporary 2025 frameworks utilizing Relative Phase analysis, Kolmogorov-Arnold Networks (KAN), and YOLO-hybrid detection for powerlifting and clinical assessment. The survey categorizes literature into four critical domains: strength training, yoga/pilates classification, sports-specific biomechanics, and infant/clinical monitoring. Furthermore, we identify persistent research gaps, including Z-axis depth inaccuracies, temporal rhythm assessment, and view-invariant detection, providing a technical roadmap for the development of an "AI Vision-Based Solo Training Assistant."