Abstract
M.Ing. (Electrical and Electronic Engineering Science)
The Visually Aided 3-D Simultaneous Localisation and Mapping (VA-SLAM) is presented as a solution to the SLAM problem for underground environments with the aid of visual information. The system is capable of operating in environments with poor illumination conditions and without GPS localisation. This work aims at furthering the field of robotics in the mining environment, which has not yet benefited from visual techniques.
Using the Kinect sensor data, frames are aligned in 3-D using a robust and efficient variation of the ICP algorithm, which exploits the space-efficient octree data structure. This algorithm is aided with an initial pose estimate whenever possible using sparse visual alignment. The sparse visual alignment system makes use of SIFT feature extraction on 2-D images. The features are matched using kNN-matching and use a distance ratio test on matched features to discard outliers. The matched features are then used to align the images with the RANSAC algorithm. Keyframes, which are sufficiently sparse, are selected and added an optimisable graph and loop closure is attempted for that frame. Loop closure is attempted between the current keyframe and the subset of keyframes which are currently not connected and estimated transformations are within a threshold. Loop closure is attempted on every new keyframe using the same alignment technique as before. Finally, the graph is optimised and used to provide a map at run-time. The system was evaluated on a publicly available data set, using open source benchmarking tools. The results showed that the Octree-ICP algorithm proposed here is vastly superior to previous ICP variations. It is also shown that SIFT is a good choice for feature extraction in visual alignment, comparing it to two other popular techniques in visual alignment; namely SURF and ORB. The final, graphoptimised result is also shown to be comparable to current state-of-the-art visual graph-based SLAM solutions. The system is efficient enough to produce a map at run-time and most importantly, the system is shown to be robust to vastly different levels of illumination.