Mutual information-based hierarchical NBV decision for active semantic visual SLAM under dynamic environments

Active Simultaneous Localization and Mapping (A-SLAM) technology enables a robot to autonomously plan its movements to build a comprehensive and accurate map of its surroundings. However, most existing SLAM algorithms are not robust in dynamic environments, as moving objects can negatively impact mapping and localization accuracy, making it difficult for the robot to keep tracking and fully understand its environment. While some semantic SLAM methods can identify and exclude dynamic objects, in active SLAM, excluding features without proper path planning carries significant risks of losing track. In this work, we propose a real-time mutual information-based active SLAM approach designed to enhance robustness in dynamic environments. The proposed method not only excludes dynamic objects from the mapping process but also integrates two Next-Best-View (NBV) decision modules to improve path planning and maintain robustness. This feature allows for improved mapping efficiency and robustness to avoid losing tracking in dynamic environments. Experiments conducted in two simulated environments and one real-world scenario demonstrate that the proposed active SLAM algorithm maintains its robustness and efficiency in dynamic environments, and is deployable in real applications.