Recent advances in implicit mapping and positioning have yielded promising results by leveraging the characteristics of Neural Radiance Fields (NeRFs). NeRFs enable the representation of continuous volumetric density and RGB values in a neural network, which can be used to reconstruct the geometry of unknown scenes. However, existing methods have challenges in scaling up to larger scenes and only consider scenarios with a single agent. In this article, we present a collaborative implicit SLAM framework that supports multiple agents running independent implicit SLAM onboard by sharing map information with the server for map fusion. Specifically, we propose a floating-point sparse octree as the structure for storing map information and aligning local maps by transforming three vertices in the octree. To ensure more accurate and efficient map fusion, our backend employs place recognition, implicit alignment, and removal of redundant data. The evaluation results show that our methods can achieve more accurate map fusion through effectively reducing overlap and noise areas.
Efficient Map Fusion for Multiple Implicit SLAM Agents
IEEE Transactions on Intelligent Vehicles ; 9 , 1 ; 852-865
01.01.2024
4505191 byte
Aufsatz (Zeitschrift)
Elektronische Ressource
Englisch