Improving data association in vision-based SLAM

This paper describes an approach to solve the Simultaneous Localization and Mapping (SLAM) problem for autonomous mobile robots using visual landmarks. Our map is represented by a set of three dimensional landmarks referred to a global reference frame. We use significant points extracted from stereo images as natural landmarks, in particular we employ SIFT features found in the environment. Each landmark contains a visual descriptor that partially differentiates it from others. Our method is based on a Rao-Blackwellized particle filter, thus the problem is decomposed into two parts: one estimation over robot paths, and N independent estimations over landmark positions, each conditioned on the path estimate. We actively track visual landmarks at a local neighbourhood and select only those that are more stable. When a visual feature has been observed from a significant number of frames it is then integrated in the filter. By this procedure, the total number of landmarks in the map is reduced, compared to prior approaches. Due to the tracking of each landmark, we obtain different examples that represent the same natural landmark. We use this fact to improve data association. Finally, efficient resampling techniques have been applied, which reduces the number of particles needed and avoids the particle depletion problem