Rozpoznávání objektů na základě jejich 3D modelů: volba klíčových bodů

Abstract

This thesis focuses on 3D point cloud processing with an emphasis on object recognition and pose estimation. It analyzes the influence of key steps of the processing pipeline, namely keypoint detection, local feature descriptor computation, and correspondence matching, on the accuracy and robustness of the results.

Selected methods, including PFH, FPFH, and SHOT descriptors, are implemented and compared, while transformation estimation is performed using the RANSAC and ICP algorithms. Experiments were conducted on scenes of varying complexity.

The results show a significant influence of parameter selection as well as scene characteristics on the success of pose estimation and provide recommendations for parameter selection.