Effective Visual Place Recognition Using Multi-Sequence Maps

Visual place recognition is a challenging task, especially in outdoor environments as the scenes naturally change their appearance. In this letter, we propose a method for visual place recognition that is able to deal with seasonal changes, different weather condition as well as illumination changes. Our approach localizes the robot in a map, which is represented by multiple image sequences collected in the past at different points in time. Our approach is also able to localize a vehicle in a map generated from Google Street View images. Due to the deployment of an efficient hashing-based image retrieval strategy for finding potential matches in combination with informed search in a data association graph, our approach robustly localizes a robot and quickly relocalizes if it is getting lost. Our experiments suggest that our algorithm is an effective matching approach to align the currently obtained images with multiple trajectories for online operation.