Browsing by Author "Chiu, Alexander"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- ItemProbabilistic Outlier Removal for Stereo Visual Odometry(Stellenbosch : Stellenbosch University, 2017-03) Chiu, Alexander; Van Daalen, Corne E.; Stellenbosch University. Faculty Engineering. Dept. of Electrical and Electronic Engineering.ENGLISH ABSTRACT: The field of autonomous navigation is currently receiving significant attention from researchers in both academia and industry. With an end goal of fully autonomous vehicle systems, an increased effort is being made to develop systems that are more efficient, reliable and safe than human-controlled vehicles. Furthermore, the low cost and compact nature of cameras have led to an increased interest in vision-based navigation techniques. Despite their popularity, measurements obtained from cameras are often noisy and contaminated with outliers. A critical requirement for consistent and reliable autonomous navigation is the ability to identify and remove these outliers when measurements are highly uncertain. The focus of the research presented in this thesis is therefore on effective and efficient outlier removal. Many existing outlier removal methods are limited in their ability to handle datasets that are contaminated by a significant number of outliers in real-time. Furthermore, many of the current techniques perform inconsistently in the presence of high measurement noise. This thesis proposes methods for probabilistic outlier removal in a robust, real-time visual odometry framework. No assumptions are made about the vehicle motion or the environment, thereby keeping the research in a general form and allowing it to be applied to a wide variety of applications. The first part of this thesis details the modelling of sensor measurements obtained from a camera pair. The mapping process from 3D space to image space is described mathematically and the concept of triangulating matched image features is presented. Stereo measurements are modelled as random variables that are assumed to be normally distributed in image coordinates. Two techniques used for uncertainty propagation, linearisation and the unscented transform, are investigated. The results of experiments, performed on synthetic datasets, are presented and show that the unscented transform outperforms linearisation when used to approximate the distributions of reconstructed, 3D features. The second part of this thesis presents the development of a novel outlier removal technique, which is reliable and efficient. Instead of performing outlier removal with the standard hypothesise-and-verify approach of RANSAC, a novel mechanism is developed that uses a probabilistic measure of shape similarity to identify sets of points containing outliers. The measure of shape similarity is based on inherent spatial constraints, and is combined with an adaptive sampling approach to determine the probability of individual points being outliers. This novel approach is compared against a state-of-the-art RANSAC technique, where experiments indicate that the proposed method is more efficient and leads to more consistent motion estimation results. The novel outlier removal approach is also incorporated into a robust visual odometry pipeline that is tested on both synthetic and practical datasets. The results obtained from visual odometry experiments indicate that the proposed method is significantly faster than RANSAC, making it viable for real-time applications, and reliable for outlier removal even when measurements are highly uncertain.