Unsupervised Learning of Metric Representations with Slow Features from Omnidirectional Views

Unsupervised learning of Self-Localization with Slow Feature Analysis (SFA) using omnidirectional camera input has been shown to be a viable alternative to established SLAM approaches. Previous models for SFA self-localization purely relied on omnidirectional visual input. The model led to globally consistent localization in SFA space but the lack of odometry integration reduced the local accuracy. However, odometry integration and other downstream usage of localization require a common coordinate system, which previously was based on an external metric ground truth measurement system. Here, we show an autonomous unsupervised approach to generate accurate metric representations from SFA outputs without external sensors. We assume locally linear trajectories of a robot, which is consistent with, for example, driving patterns of robotic lawn mowers. This geometric constraint allows a formulation of an optimization problem for the regression from slow feature values to the robot’s position. We show that the resulting accuracy on test data is comparable to supervised regression based on external sensors. Based on this result, using a Kalman filter for fusion of SFA localization and odometry is shown to further increase localization accuracy over the supervised regression model.