VDB-Submapping: Efficient Handling of Retroactive Pose Changes During Large-Scale Volumetric Mapping

Abstract

Volumetric mapping algorithms are widely used in robotics and computer vision applications, such as autonomous navigation and object recognition. The VDB-mapping frame-work is a fast, memory efficient mapping algorithm that uses a tree based voxel structure to represent the 3D environment. In large scale environments small error accumulate over time resulting in larger positioning errors. This is commonly alleviated by a SLAM algorithm in which loop closure constraints are used to correct all robot poses retroactively. However, as the sensor data is already tightly integrated into the map at this point, the position corrections can no longer be applied to the map data. In this paper, we propose to store the sensor data in small locally consistent volumetric submaps. These submaps are loosely linked to each other by the SLAM pose graph. As a result, the map data can be easily shifted, after a graph optimization happened, to account for positioning errors. The proposed method demonstrates comparable performance in terms of integration time and memory footprint, while also effectively accounting for loop closure constraints.