Active SLAM With Dynamic Viewpoint Optimization for Robust Visual Navigation

Robust and precise visual localization is critical for autonomous systems, as it mimics human perception while offering cost-effective, scalable, and versatile advantages. However, existing fixed-mounted visual sensors (passive perception methods) often exhibit performance degradation or instability under conditions such as insufficient or missing texture in real-world environments, dynamic target disturbances, and environmental changes. In this article, we present an adaptive camera view adjustment technique for developing an active simultaneous localization and mapping (SLAM) system. This method assumes that the vision feature projection process can be represented as a 2-D Gaussian distribution, addressing the complex task of assessing the efficacy of view adjustments and simplifying camera view optimization using a standard gradient descent framework. Unlike passive visual localization methods using fixed cameras, this approach adopts the observation pose based on the distribution of environmental features to establish and maintain more robust feature associations, significantly improving both the robustness and accuracy of visual localization. Furthermore, the active SLAM framework can seamlessly integrate additional constraints, enhancing its capability in visual navigation within complex environments that demand heightened safety and efficiency. Experiments in simulated and real-world scenarios demonstrate significant improvements in localization accuracy and robustness, a 26.9% average improvement in positioning availability, highlighting its value for enhancing the autonomous localization capability of mobile robots.