This thesis deals with the implementation of navigation strategies for a domestic floor-cleaning robot operating on omnidirectional images as primary sensory information. Such navigation strategies enable the robot to efficiently
cover its entire workspace while avoiding both uncleaned areas and repeated coverage. This is accomplished (i) by systematically guiding the robot along meandering lanes, i.e. along straight lanes placed next to each other at a
predefined and constant distance and (ii) by building a map of the robot's environment to distinguish cleaned and uncleaned areas. Since domestic cleaning robots are considered consumer goods, they can only be equipped with a limited number of cheap sensors and restricted computational power. This fact poses additional challenges onto the design of navigation strategies for domestic
floor-cleaning robots.
The navigation strategies described in this thesis use omnidirectional images, in our case panoramic images with a full 360° horizontal field of view. We consider omnidirectional cameras an appropriate choice because they (i) are relatively cheap sensors, (ii) provide dense sensory information about the robot's environment, and (iii) are multi-purpose sensors applicable to further
aspects of cleaningrobot navigation beyond the scope of this thesis (e.g. obstacle detection, visual odometry, or user interaction). We characterize a position in space by the entire omnidirectional image acquired at this place
(hence the methods belong to the class of appearance-based navigation methods) without detecting visible features in the image. Several places are integrated into a dense topo-metric map of the robot's environment. Such maps (i) offer a metrical position estimate required for guiding the robot along meandering lanes, (ii) have a spatial resolution which is fine enough for accurate
navigation, (iii) can be easily built from the available sensor data, and (iv) allow for efficiently operating on the maps. Spatial relations between places stored in the map are estimated by applying a local visual homing method. Such methods are parsimonious yet robust and accurate methods for partial ego-motion estimation from visual information. They recover the direction (but not the
distance) of the translation and the rotation of the robot's motion between two images acquired in direct vicinity of each other without physically moving
between places. As far as we know, our navigation methods are the first application of omnidirectional vision, dense topo-metric maps, and local visual homing for the control of cleaning robots. Hence, this thesis is also a
feasibility study to prove the applicability of these concepts for navigation of cleaning robots. Since a complete control scheme for a cleaning robot is beyond
the scope of this thesis, we propose two essential substrategies of such a control scheme: (i) vision-based trajectory control and mapping and (ii) visual
detection of already cleaned areas.
Regarding trajectory control and mapping, we propose a mostly vision-based controller for covering a rectangular area of the entire workspace by meandering
lanes. While moving along a lane (and cleaning), the robot adds snapshots at regular distances to its dense topo-metric map, which are used on the subsequent
lane to estimate the robot's current distance to the previous lane. For this purpose, the bearing from the current position towards at least two snapshots
stored along the previous lane is taken by applying local visual homing. The bearing information and an odometry-based estimate of the distance between the
two considered snapshots are fused in order to estimate the robot's current distance to the previous lane. The robot is kept on a lane parallel to the previous one by keeping the distance to the previous lane at a predefined value. Instead of estimating the robot's full pose as performed by common navigation strategies, we only estimate the distance to the previous lane and the robot's
current orientation to avoid unnecessary computations. The results obtained from real-robot experiments reveal that the algorithm is capable of guiding the robot
along parallel and meandering cleaning lanes with only a small portion of gaps or overlap between lanes.
Detecting already cleaned areas is essential in order to avoid repeated coverage or uncleaned areas between neighboring cleaning segments. This detection is a
special instance of the loop-closure detection problem usually occurring during map-building whenever the robot approaches an already mapped area. We solve the loop-closure problem by two different approaches both incorporating pairwise image comparisons between the robot's currently perceived image and several images stored in the map. The first approach, referred to as holistic approach, relies on pixel-by-pixel comparisons of the considered images. The second, referred to as signature-based approach, computes low-dimensional signatures extracted from the entire image and compares these instead of the images themselves.
Pixel-based approaches require the application of a compass to align images prior to image comparison. The standard compass method rotates one of the images
step-by-step while keeping the other fixed and repeatedly compares the images to search for the best match. We propose an accelerated variant of this method
operating in the Fourier domain. This method is capable of computing the best match without repeatedly shifting and comparing images. In order to achieve robustness against illumination changes, we preprocess images prior to
comparison and apply illumination-tolerant comparison functions. Loop-closure detection and compass accuracy were assessed by image-database experiments
systematically evaluating a wide range of different preprocessing and comparison techniques. Regarding loop-closure detection, holistic methods achieve very good
detection results even for strong illumination changes. The proposed Fourier-based compass is more efficient than the standard method, but does not achieve its accuracy. Due to their computational complexity, the tested holistic
approaches to loop-closure detection are --at least with the current implementation-- not suitable for a real-robot application.
Signature-based approaches allow for efficient image comparisons because they rely on lowdimensional and rotation-invariant image descriptors. Due to their
rotational invariance, signatures can be compared without prior compass alignment as required by holistic methods. To measure the accuracy of loop-closure detection, we performed image-database experiments which systematically tested different combinations of signatures and comparison
functions operating on the images' intensity information without prior preprocessing. The tested methods allow for accurate loopclosure detection under constant illumination. However, detection is likely to fail under moderate or strong illumination changes. For the most promising combination of signature and comparison function, real-robot experiments were conducted leading to similar results. These results are surprising because we tested several combinations of signatures and comparison functions which should theoretically tolerate
illumination changes better than the combination performing best in our experiments. Despite their low tolerance against illumination changes, which
need to be increased in future work, we favor the application of signature-based approaches because of their low computational complexity.
The overall results of this thesis clearly reveal that omnidirectional vision, dense topo-metric maps, and local visual homing are appropriate building blocks for visual control of cleaning robots because they allow for efficient, accurate, and robust navigation. We conclude that dense topo-metric maps are suitable representations of space --both for trajectory control and for detecting already cleaned areas. Thus, the navigation strategies proposed in this thesis can be used as a basis for a more complex control architecture enabling the robot to completely cover complex-shaped areas. This includes
mechanisms to detect and approach uncleaned areas based on map information and to combine several segments of meandering lanes as obtained from our trajectory
controller. Using omnidirectional vision not only for navigation but also for obstacle detection, odometry, or user interaction could be a promising means to
reduce hardware costs of a potential product by avoiding dedicated sensors.