Three-Dimensional visual SLAM for Unstructured Domains

Publication Type:
Thesis
Issue Date:
2010
Full metadata record
Files in This Item:
Filename Description Size
Thumbnail01Front.pdf3.65 MB
Adobe PDF
02Whole.pdf52.69 MB
Adobe PDF
NO FULL TEXT AVAILABLE. Access is restricted indefinitely. ----- This thesis is concerned with finding a vision-based solution to the simultaneous localization and mapping (SLAM) problem in an urban search and rescue scenario (USAR). The contribution of this thesis is three-fold. Firstly, a novel two-dimensional (2D) stereo-based algorithm is presented. This algorithm demonstrates the viability of a stereo vision solution to the SLAM problem, and with application to the less challenging part of a search and rescue environment. Using features extracted from the scale invariant feature transform (SIFT) and depth maps from a stereo head, SLAM can be tackled by an Extended Kalman Filter (EKF) based algorithm. This allows the independent use of information relating to the depth and bearing of natural visual features. By means of a map pruning strategy for managing the computational cost, the production of statistically consistent location estimates are demonstrated. However, a two-dimensional representation of the environment becomes insufficient as the robot travels further into a less structured area such as a demolition site. A three-dimensional SLAM solution is therefore required, thereby constituting the second contribution of this thesis. By combining a conventional camera and a range imager, the robot is capable of making reliable range, bearing and elevation observations as input to the three-dimensional SLAM approach. This is achieved independently of the scene texture, which is often lacking in a search and rescue environment. In addition, a computationally efficient algorithm is proposed to give the robot the ’intelligence’ to select the maximum informative observations used in the estimation process, from the steadily collected data. It is shown that while the actual evaluation of the information gain may introduce an additional computational cost, the overall efficiency is significantly increased by keeping the information matrix compact. The noticeable advantage of this strategy is that the continuously gathered data is not heuristically segmented prior to being inputted to the filter. Quite the opposite infact, as the scheme lends itself to be statistically optimal. To this end, it is capable of handling sizeable datasets collected at a realistic sampling rate. The third contribution rises from the discussion of the realistic constraint imposed by the deployment of the proposed methodology in a search and rescue environment. Under these conditions, a number of tasks with varying priorities can be executed by the rescue robot in a timely manner and throughout the course of a rescue mission. The best outcomes will not only be measured by the accuracy of the robot’s location and the generated exploratory map, but can also represent a trade-off between a number of other key parameters. These can include the ground coverage, the feature distribution, and so on. A methodology is presented in the second last chapter which addresses the adjustment of the SLAM outcome through altering the amount and nature of the information being processed.
Please use this identifier to cite or link to this item: