Analýza lidské chůze s využitím synteticky generovaných hloubkových map

Abstract

This bachelor thesis focuses on the use of synthetically generated depth maps for human gait analysis. A modular system was designed and implemented that combines human body keypoint detection with seven monocular depth estimation models and a three-level calibration scheme. Reference data were collected using an Intel RealSense D455 sensor in two indoor scenarios. The experiment demonstrated that step length can be estimated with a deviation of 0.1–2.6 cm compared to reference data, while step height is more sensitive to the absolute accuracy of the depth map and shows larger deviations. Gait phase detection was robust with respect to the choice of depth model. The best results were achieved by VDA-Base with per-frame calibration in scenario A and Marigold with spatial calibration in scenario B.