Cumulative Evidence for Scene Change Detection and Local Map Updates

Accurate localization and mapping is a key to applications such as Metaverse, Augmented Reality, and Autonomous Driving using positioning technology to determine their global position in consistent world coordinates. As the scene changes with time, map descriptors become outdated, affecting Visual Positioning System localization accuracy. Previous studies have primarily relied on direct comparison of point clouds for change detection, which is a slow process due to the need to build a new point cloud every time. Image-based comparison requires keeping the map images - a privacy issue and is sensitive to viewpoint differences. In this work, we propose a novel approach based on point-clouds descriptors comparison, which can detect structural and texture scene changes followed by the process of local map update. This approach is more robust under appearance changes, even in illumination differences, and more efficient for local map updates as it provides better localization accuracy and faster run times. The cumulative evidence approach eliminates the need for a dedicated mapping process. In addition, our work provides state-of-the-art performances for image-to-image change detection compared to previous research.