Camera perspective distortion in model-based visual localisation.

Abstract

This thesis starts with a proposal for a collaborative global visual localization system. Then, it centres in a specific visual localisation problem: perspective distortion in template matching.The thesis enriches 3D point cloud models with a surface normal associated with each 3D point. These normals are computed using a minimization algorithm.Based in this new model, the thesis proposes an algorithm to increase the accuracy of visual localisation. The algorithm improves for template matching processes using surface normals.The hypothesis, `Given a 3D point cloud, surface orientation of the 3D points in a template matching process increases the number of inliers points found by the localisation system, that is, perspective compensation.' is objectively proved using a ground truth model.The ground truth is achieved through the design of a framework which using computer vision and computer graphics techniques carries out experiments without the noise of a real system, and prove in an objective way the hypothesis.