Odhad polohy objektu v prostoru

Abstract

This master's thesis addresses the problem of six-degree-of-freedom (6-DoF) object pose estimation. A survey of current pose estimation approaches is conducted. Two representative methods are selected for practical implementation -- the RGB-based HccePose and the RGB-D-based FoundationPose. Both methods are integrated into a demonstration application written in C++ using the TensorRT runtime for real-time inference. A YOLO26 model is trained for target object detection. For quantitative evaluation, a custom test dataset is acquired using a Stereolabs ZED 2i stereo camera, with ground truth based on an ArUco marker board. The evaluation is performed using metrics defined by the BOP Benchmark. On the custom dataset, FoundationPose achieves higher accuracy (BOP-M 0.924 vs. 0.875) as well as greater robustness against occlusions. The implemented application enables real-time pose estimation using the Stereolabs ZED 2i camera.