基于轻量化智能的多机协同 SLAM 系统

视觉多机协同即时定位与地图构建(SLAM)主要以相机作为传感器,并通过多机器人合作实现定位与建图。 然而,在 面对复杂环境时前端计算量过大,易导致整体系统精度不理想。 启发于 REVO 和 SVO 算法的轻量化特点,提出一种基于轻量 化智能的多机协同 SLAM 系统,旨在降低前端计算资源的同时提升系统可扩展性。 提出改进 REVO 算法—L-REVO,通过轻量 化改进实现前端实时运行;将 L-REVO 融合 CCMSLAM 系统后端,提出一种完整的多机协同 SLAM 架构;调整前端传感器和算 法,分别验证前端为同构或异构时对系统性能的影响。 在公开数据集 TUM 上,相比 CCMSLAM 系统,该系统两种模式下定位精 度分别提高了 59. 4% 和 31. 6% ,能效比提升了 8 倍。 最后,将该系统用于室内场景实验,前端功耗仅 1. 43 W,验证了所提系统 的可行性和有效性。

Lightweight intelligence-based multi-machine collaborative SLAM system

Vision multi-robot cooperative SLAM mainly uses cameras as sensors and achieves localization and map building through multi-robot cooperation. However, the front-end computation is too large in the face of complex environments, which tends to lead to unsatisfactory overall system accuracy. Inspired by the lightweight features of REVO and SVO algorithms, this article proposes a multirobot cooperative SLAM system based on lightweight intelligence, aiming to reduce front-end computational resources while improving system scalability. This article proposes an improved REVO algorithm-L-REVO to realize the front-end real-time operation through lightweight improvement; fusing L-REVO with the back-end of CCMSLAM system to propose a complete multi-machine collaborative SLAM architecture; adjusting the front-end sensors and algorithms to verify the impact on system performance when the front-end is homogeneous or heterogeneous, respectively. On the public dataset TUM, the system improves the localization accuracy by 59. 4% and 31. 6% in both modes, respectively, and the energy efficiency ratio by 8 times compared with the CCMSLAM system. Finally, the system is used for indoor scenario experiments with a front-end power consumption of only 1. 43 W, which verifies the feasibility and effectiveness of the proposed system.