Vision-based Markov localization for long-term autonomy

Lifelong autonomous operation has gained much attention in the field of mobile robotics in recent years. In the context of robot navigation based on vision, lifelong applications include scenarios with substantial perceptual changes due to changes in season, illumination and weather. In this paper, we present an approach to localize a mobile robot, equipped with a low frequency camera, with respect to an image sequence recorded in a different season. Our approach employs a discrete Bayes filter with a sensor model based on whole image descriptors. We compute a similarity matrix over all image descriptors and leverage the sequential nature of typical image streams with a flexible transition scheme in the Bayes filter framework. Since we compute a probability distribution over the entire state space, our approach can handle complex trajectories that may include same season loop-closures as well as fragmented sub-sequences. Furthermore, we show that decorrelating the similarity matrix results in an improved localization performance. Through an extensive experimental evaluation on challenging datasets we demonstrate that our approach outperforms state-of-the-art techniques.