Performance Evaluation of Incremental Eigenspace Models for Mobile Robot Localization

We address the problem of robot localization based on eigenspace representations of a collection of views of the workspace, often referred to as appearance based methods. We study several approaches to build such eigenspaces in an incremental fashion (subspace tracking), thereby allowing the simultaneous localization and mapping of a mobile vehicle. We see how different ways of determining whether or not new information (images) should be included in the eigenspace model lead to different error values and model size (complexity). The performance evaluation allows us to know which criterion and parameters to use in a specific application to meet error specifications subject to computational constraints.