Robust and Accurate Detection of Mid-level Primitives for 3D Reconstruction in Man-Made Environments

Abstract

The detection of geometric primitives such as points, lines and arcs is a fundamental step in computer vision techniques like image analysis, pattern recognition and 3D scene reconstruction. In this thesis, we present a framework that enables a reliable detection of geometric primitives in images. The focus is on application in man-made environments, although the process is not limited to this. The method provides robust and subpixel accurate detection of points, lines and arcs, and builds up a graph describing the topological relationships between the detected features. The detection method works directly on distorted perspective and fisheye images. The additional recognition of repetitive structures in images ensures the unambiguity of the features in their local environment. We can show that our approach achieves a high localization accuracy comparable to the state-of-the-art methods and at the same time is more robust against disturbances caused by noise. In addition, our approach allows extracting more fine details in the images. The detection accuracy achieved on the real-world scenes is constantly above that achieved by the other methods. Furthermore, our process can reliably distinguish between line and arc segments. The additional topological information extracted by our method is largely consistent over several images of a scene and can therefore be a support for subsequent processing steps, such as matching and correspondence search. We show how the detection method can be integrated into a complete feature-based 3D reconstruction pipeline and present a novel reconstruction method that uses the topological relationships of the features to create a highly abstract but semantically rich 3D model of the reconstructed scenes, in which certain geometric structures can easily be detected.

True Orthophoto Generation Based on Unmanned Aerial Vehicle Images Using Reconstructed Edge Points

“…The researchers who are working on 3D line reconstruction methods have attempted to create point clouds from images and fill them with extra points, such as structural edge points. 3D line reconstruction methods in real datasets may be classified into two groups based on whether the camera poses are known or unknown (Wolters, 2018). The majority of the proposed scenarios are based on weak SfM pipelines (Bartoli et al., 2004; Bartoli and Sturm, 2005; Bay et al., 2005; 2006; Schindler et al., 2006; Zhang and Koch, 2014), which only use two images to match line segments (LSs) and solve both camera pose estimation and 3D reconstruction, resulting in poor geometry that requires more additional constraints.…”