Sledenje človeške drže v 3D prostoru pri simuliranih vojaških vajah

Simulated military training exercises require pose estimation systems that can handle unusual body configurations and viewpoints. Trainees adopt three key positions during tactical drills: standing, crouching, and prone, with the latter presenting fundamental challenges due to extreme foreshortening and self-occlusion patterns that differ dramatically from standard training datasets. This thesis systematically evaluates state-of-the-art pose estimation methods for tracking human pose in 3D during simulated military drills, examining both 2D-to-3D lifting approaches that reconstruct skeletal joint positions and parametric mesh recovery methods that generate complete body surfaces. The evaluation analyzes accuracy, failure modes, computational efficiency, and practical deployment constraints. Results reveal a clear hierarchy of difficulty, with standing and crouching positions handled robustly by most methods, while prone positions cause most methods to fail. CoMotion emerged as the optimal solution with its tracking-by-attention approach, achieving the best overall performance across all poses while maintaining predictions during occlusions and showing superior precision, though the most extreme prone cases remain challenging.