Monocular-Based SLAM for Mobile Robots: Filtering-Optimization Hybrid Approach

The SLAM or Simultaneous Localization and Mapping still remains one of the most important problems to be fully addressed in the path to building fully autonomous mobile robots. This work aims to contribute to the above objective by presenting a novel hybrid architecture for implementing monocular-based SLAM systems for mobile robots. The key idea is to take advantage in a complementary manner of the inherent properties of the two main methodologies available today for implementing visual-based SLAM systems: the filter-based methods, and the optimization-based methods. The proposed method differs from previous approaches among other aspects because filter-based and optimization-based functionalities run concurrently in separate processes, improving the modularity and robustness of the system due to a degree of redundancy. Also, due to its modularity, the proposed approach can be easily applied to different mobile robot platforms using a monocular camera as its main sensor and different setups of additional sensors. In this sense, experiments with real data obtained from a differential robot and a quadrotor are presented to show that the proposed SLAM system can perform well in out-of-the-lab environments by using low-cost sensors. The experiments also show that the SLAM system exhibits real-time and stable computational performance. Moreover, a ROS-2 C++ open-source implementation of the proposed system is provided.