SBC-SLAM: Semantic Bioinspired Collaborative SLAM for Large-Scale Environment Perception of Heterogeneous Systems

Recently, there have been efforts to enable collaborative simultaneous localization and mapping (SLAM) of heterogeneous unmanned aerial vehicle (UAV)/unmanned ground vehicle (UGV) in large-scale dynamic environments. However, the challenges of collaborative localization introduced by the view differences between aerial and ground robots, high cost of storing, reusing the metric map, greatly limit the application of traditional visual SLAM algorithms in cross-domain robot collaborative SLAM. To address these issues, this article introduces a semantic-based bioinspired SLAM frame work for operating mobile robots equipped with a visual system in unstructured large-scale outdoor environments. This system forms a semidense semantic map through bioinspired tightly coupled visual inertial odometry, which can encode the sensory cues and self-motion cues with biological neural models in a compact and efficient manner. The topological information of semantic feature points serves as visually robust scene descriptor. It enables UAVs and UGVs to recognize the same scene with view variation. The proposed system is tested in both synthetic and real-world large-scale unstructured environments. Proposed system achieving a similar level of accuracy with about 20% of the key frame information required by traditional algorithms. It successfully registered heterogeneous robot’s trajectory in both simulation and real environment, with errors under 10%. Its accuracy shows advantage to previous algorithms.

I. Introduction

Cross-domain unmanned systems are organic entities composed of multiple sets of unmanned systems with significant functional differences operating in different spatial domains. They can complementarily utilize various capabilities in different fields, enhancing efficiency and establishing advantages in multiple domains. Such systems have broad applications in military, counterterrorism, riot control, disaster relief, and other fields. However, it is challenging to achieve consistent representation, abstraction, and fusion of information with differences in perspectives, types, scales, and noise acquired by cross-domain platforms. Also, the large-scale environments cause difficulties in information storage and retrieval for traditional visual simultaneous localization and mapping (SLAM) system [1], [2], [3].