Top-down Attention Supports Visual Loop Closing

In this paper, we present a method to improve the loop closing behaviour for visual SLAM. Landmarks consist of a combination of attention regions and Harris-Laplace corners. The attention regions are detected by a visual attention system which combines image-based, bottom-up and target-related, topdown information. The ability to perform target-directed search is used to search for expected landmarks. We analyze the amount of correct and false matches for bottom-up and top-down matching depending on different matching thresholds. It shows that whereas bottom-up matching is useful for situations in which the scene changes only slightly like during tracking, top-down matching has advantages in loop closing situations by detecting a much higher amount of correctly matched landmarks.