Abstracts

Vision based mapping and navigation: modelling, geometric analysis and quality control

by Zeyu Li




Institution: University of New South Wales
Department:
Year: 2017
Keywords: Navigation; Vision; Mapping; Geometry; Quality control
Posted: 02/01/2018
Record ID: 2188250
Full text PDF: http://handle.unsw.edu.au/1959.4/58300


Abstract

For the last three decades, vision based mapping and navigation has become one mainstream to localize platforms and mapping environments especially where Global Navigation Satellite System (GNSS) is not available. By using the camera measurement, vision based mapping and navigation is able to determine cameras poses and objects coordinates with regard to a specific coordinate system. However, since the inputs of such systems are merely images with ambiguities, a number of potential issues still exist, deteriorating its accuracy, integrity and continuity. This thesis aims to analyze and improve the performance as well as extend the application of vision based mapping and navigation from sparse mapping to dense mapping. The main contribution of this thesis can be summarized as follows:a) Performance evaluation of robust estimators and outlier detection algorithms is conducted in navigation using reality based 3D maps as outliers will deteriorate its integrity. The resources of outliers are analyzed, and then a simulated environment is created to evaluate the performances of these outlier detection algorithms. The experiments demonstrated their feasibility to detect and exclude outliers varies with the magnitudes of outliers and environment characteristics. It shows that Least Trimmed Squares (LTS) and Modified M (MM) estimator have the most balanced performance in Detection Capability (DC) and False Alarm Rate (FA).b) The concept of geometry in vision based mapping and navigation is defined by the geometric components between the camera and objects. On the mapping side, the geometrys influence on global redundancy number, Dilution of Precision (DoP), Minimum Detectable Bias (MDB) and Minimum Separable Bias (MSB) are analyzed through a simulated environment. It shows that the number and distribution of Ground Control Points (GCPs) will significantly affect global redundancy number. Besides, the number of images is the most important factor that affects MDB and MSB. On the navigation side, detectability and separability in both single-outlier and multiple-outlier scenarios are analyzed using real indoor images. It shows that better geometry will have better detectability and separability. Among them, the number and distribution of Pseudo Ground Control Points (PGCPs) are the most important factors.c) A new environmental feature named as Line Segment Intersection Feature (LSIF) is proposed in this thesis to improve the continuity of vision based mapping and navigation. The proposed design of LSIF detection, description, matching and validation algorithm guarantees its performance in texture-less indoor environments. By using real indoor images, LSIF shows superior performances compared with traditional keypoints such as Scale Invariant Feature Transform (SIFT) in number of matched keypoints. Under the framework of navigation using reality based 3D maps, LSIF shows better accuracy and continuity compared with traditional keypoints.d) A new keypoint mismatch removal approach for images in vision based mapping andAdvisors/Committee Members: Wang, Jinling, Faculty of Engineering, UNSW, Rizos, Chris, Faculty of Engineering, UNSW.