Analýza rozsahu pohybů člověka

Abstract

This thesis addresses the automated measurement of human range of motion from video recordings using computer vision methods. The primary motivation is to replace manual goniometry, which exhibits inter-rater variability of up to 25° and requires trained personnel, with an objective and repeatable procedure. A survey of available human pose estimation algorithms is conducted, covering both convolutional and Transformer architectures operating in 2D and 3D space. A synthetic dataset comprising 1,128 video recordings with ground truth accurate to tenths of a degree is then generated using the biomechanical model SKEL, the parametric body model SMPL-X, and photorealistic textures from the BEDLAM project within the Blender environment. On this dataset, 31 pose estimation model configurations are benchmarked using MAE, RMSE, PCK, and standard deviation metrics. Among the evaluated models, RTMP-M achieves the lowest mean absolute error (2.45°); however, considering the trade-off between accuracy, inference speed, and robustness to BMI variability, YOLO26l-pose is selected as the optimal model (MAE = 2.72°, PCK7.5° = 96.89%). This model reaches accuracy on synthetic data comparable to the reliability of manual measurement performed by a single experienced clinician, while providing full repeatability and eliminating inter-rater variability. Based on these results, a desktop application is implemented in Python with a PyQt6 graphical interface, supporting both video analysis and real-time measurement with export to JSON, XML, and PDF formats.