PUB - Publikationen an der Universität Bielefeld

Self-localization and navigation in outdoor environments are fundamental problems a mobile robot has to solve in order to autonomously execute tasks in a spatial environ- ment. Techniques based on the Global Positioning System (GPS) or laser-range finders have been well established but suffer from the drawbacks of limited satellite availability or high hardware effort and costs. Vision-based methods can provide an interesting al- ternative, but are still a field of active research due to the challenges of visual perception such as illumination and weather changes or long-term seasonal effects. This thesis approaches the problem of robust visual self-localization and navigation using a biologically motivated model based on unsupervised Slow Feature Analysis (SFA). It is inspired by the discovery of neurons in a rat’s brain that form a neural representation of the animal’s spatial attributes. A similar hierarchical SFA network has been shown to learn representations of either the position or the orientation directly from the visual input of a virtual rat depending on the movement statistics during training. An extension to the hierarchical SFA network is introduced that allows to learn an orientation invariant representation of the position by manipulating the perceived im- age statistics exploiting the properties of panoramic vision. The model is applied on a mobile robot in real world open field experiments obtaining localization accuracies comparable to state-of-the-art approaches. The self-localization performance can be fur- ther improved by incorporating wheel odometry into the purely vision based approach. To achieve this, a method for the unsupervised learning of a mapping from slow fea- ture to metric space is developed. Robustness w.r.t. short- and long-term appearance changes is tackled by re-structuring the temporal order of the training image sequence based on the identification of crossings in the training trajectory. Re-inserting images of the same place in different conditions into the training sequence increases the temporal variation of environmental effects and thereby improves invariance due to the slowness objective of SFA. Finally, a straightforward method for navigation in slow feature space is presented. Navigation can be performed efficiently by following the SFA-gradient, approximated from distance measurements between the slow feature values at the target and the current location. It is shown that the properties of the learned representations enable complex navigation behaviors without explicit trajectory planning.