Efficient and elastic LiDAR reconstruction for large-scale exploration tasks

Thesis

High-quality reconstructions and understanding the environment are essential for robotic tasks such as localisation, navigation and exploration. Applications like planners and controllers can make decisions based on them. International competitions such as the DARPA Subterranean Challenge demonstrate the difficulties that reconstruction methods must address in the real world, e.g. complex surfaces in unstructured environments, accumulation of localisation errors in long-term explorations, and the necessity for methods to be scalable and efficient in large-scale scenarios.

Guided by these motivations, this thesis presents a multi-resolution volumetric reconstruction system, supereight-Atlas (SE-Atlas). SE-Atlas efficiently integrates long-range LiDAR scans with high resolution, incorporates motion undistortion, and employs an Atlas of submaps to produce an elastic 3D reconstruction.

These features address limitations of conventional reconstruction techniques that were revealed in real-world experiments of an initial active perceptual planning prototype. Our experiments with SE-Atlas show that it can integrate LiDAR scans at 60m range with ∼5 cm resolution at ∼3 Hz, outperforming state-of-the-art methods in integration speed and memory efficiency. Reconstruction accuracy evaluation also proves that SE-Atlas can correct the map upon SLAM loop closure corrections, maintaining global consistency.

We further propose four principled strategies for spawning and fusing submaps. Based on spatial analysis, SE-Atlas spawns new submaps when the robot transitions into an isolated space, and fuses submaps of the same space together. We focused on developing a system which scales against environment size instead of exploration length. A new formulation is proposed to compute relative uncertainties between poses in a SLAM pose graph, improving submap fusion reliability. Our experiments show that the average error in a large-scale map is approximately 5 cm.

A further contribution was incorporating semantic information into SE-Atlas. A recursive Bayesian filter is used to maintain consistency in per-voxel semantic labels. Semantics is leveraged to detect indoor-outdoor transitions and adjust reconstruction parameters online.

Authors: Wang, Y

• DOI: 10.5287/ora-zbmpv7vyx
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: active mapping; LiDAR; semantic mapping; reconstruction; robotics