Michael Horst

Cleaning robot navigation using panoramic views and particle clouds as landmarks

The paper describes a visual method for the navigation of autonomous floor-cleaning robots. The method constructs a topological map with metrical information where place nodes are characterized by panoramic images and by particle clouds representing position estimates. Current image and position estimate of the robot are interrelated to landmark images and position estimates stored in the map nodes through a holistic visual homing method which provides bearing and orientation estimates. Based on these estimates, a position estimate of the robot is updated by a particle filter. The robots position estimates are used to guide the robot along parallel, meandering lanes and are also assigned to newly created map nodes which later serve as landmarks. Computer simulations and robot experiments confirm that the robot position estimate obtained by this method is sufficiently accurate to keep the robot on parallel lanes, even in the presence of large random and systematic odometry errors. This ensures an efficient cleaning behavior with almost complete coverage of a rectangular area and only small repeated coverage. Furthermore, the topological-metrical map can be used to completely cover rooms or apartments by multiple meander parts.

Illumination Tolerance for Visual Navigation with the Holistic Min-Warping Method

Holistic visual navigation methods are an emerging alternative to the ubiquitous feature-based methods. Holistic methods match entire images pixel-wise instead of extracting and comparing local feature descriptors. In this paper we investigate which pixel-wise distance measures are most suitable for the holistic min-warping method with respect to illumination invariance. Two novel approaches are presented: tunable distance measures—weighted combinations of illumination-invariant and illumination-sensitive terms—and two novel forms of “sequential” correlation which are only invariant against intensity shifts but not against multiplicative changes. Navigation experiments on indoor image databases collected at the same locations but under different conditions of illumination demonstrate that tunable distance measures perform optimally by mixing their two portions instead of using the illumination-invariant term alone. Sequential correlation performs best among all tested methods, and as well but much faster in an approximated form. Mixing with an additional illumination-sensitive term is not necessary for sequential correlation. We show that min-warping with approximated sequential correlation can successfully be applied to visual navigation of cleaning robots.

Visual Place Recognition for Autonomous Mobile Robots

Place recognition is an essential component of autonomous mobile robot navigation. It is used for loop-closure detection to maintain consistent maps, or to localize the robot along a route, or in kidnapped-robot situations. Camera sensors provide rich visual information for this task. We compare different approaches for visual place recognition: holistic methods (visual compass and warping), signature-based methods (using Fourier coefficients or feature descriptors (able for binary-appearance loop-closure evaluation, ABLE)), and feature-based methods (fast appearance-based mapping, FabMap). As new contributions we investigate whether warping, a successful visual homing method, is suitable for place recognition. In addition, we extend the well-known visual compass to use multiple scale planes, a concept also employed by warping. To achieve tolerance against changing illumination conditions, we examine the NSAD distance measure (normalized sum of absolute differences) on edge-filtered images. To reduce the impact of illumination changes on the distance values, we suggest to compute ratios of image distances to normalize these values to a common range. We test all methods on multiple indoor databases, as well as a small outdoor database, using images with constant or changing illumination conditions. ROC analysis (receiver-operator characteristics) and the metric distance between best-matching image pairs are used as evaluation measures. Most methods perform well under constant illumination conditions, but fail under changing illumination. The visual compass using the NSAD measure on edge-filtered images with multiple scale planes, while being slower than signature methods, performs best in the latter case.