From Pixels to Performance: A Survey of the Estimation-to-Biomechanics Pipeline for Vision-Based Human Pose Estimation in Sports and Movement Health

Traditional quantitative movement analysis relies either on the subjective visual observations of a coach or laboratory-based motion capture involving markers. Human Pose Estimation (HPE) using deep learning could provide a third solution where markers can be omitted, and anatomical landmarks can be extracted from regular RGB videos through cameras, leading to movement analysis in non-laboratory environments. The potential offered by HPE is sometimes exaggerated relative to empirical results. This survey reviews 45 articles, published from 2021 to 2026, covering the four stages of the process from video to movement analysis: landmark estimation, transformation of coordinates to biomechanical metrics, sport-specific analysis, and validation against a gold standard. Our contributions include a taxonomy based on the pipeline, a comparative study of landmark estimation and metric transformation approaches, landmark-to-joint angle error propagation analysis, and temporal validity budget modeling.Three results come through clearly. First, accuracy is the best-optimized and most widely reported metric, but it is not the constraint driving adoption. Second, the translation and validation phases combine to occupy less than one fifth of the corpus, despite being essential to every practical application. Third, the reported accuracy of different frameworks is not comparable between implementations of the same framework. We summarize six converging gaps, and present our reference implementation on consumer hardware, including performance metrics (CPU throughput of 29.2 frames per second and GPU throughput of 74.1 frames per second for a 33-landmark estimator; CPU throughput of 4.4 frames per second and GPU throughput of 102.0 frames per second for a 19-landmark estimator), and what these metrics demonstrate and what they do not demonstrate