Uncertainty-driven active view planning in feature-based monocular vSLAM

Abstract

Traditional feature-based monocular visual simultaneous localization and mapping (vSLAM) methods suffer from frequent tracking failure in low-texture scenes. Although tracking stability can be guaranteed by actively adjusting the camera’s orientation to track known landmarks, it can easily lead to the problem of over-exploitation. This means that instead of discovering new landmarks, the camera focuses on an area that has already been observed, which is not conducive to fully exploring unknown environments. To address this problem, an uncertainty-driven active view planning framework is proposed to actively adjust the orientation of the monocular camera equipped on a three degree of freedom (3-DoF) pan–tilt. As a result, a trade-off between exploitation, i.e., making full use of known information, and exploration, i.e., obtaining more information of unknown environments can be achieved. First, a novel landmark uncertainty model is established to represent the uncertainty of environmental information. Second, the trade-off problem is formulated as an inequality-constrained optimization mathematical model, whose objective function is related to landmark uncertainty. Last, Karush–Kuhn–Tucker (KKT) conditions are utilized to solve the optimization problem. Experimental results on a publicly available monocular dataset and in a real-world environment show that this framework reduces the rate of tracking failure by 50% on average. The localization error is also reduced by 0.07 m for translation and 0.004 rad/m for rotation on average. Meanwhile, the number of reconstructed landmarks increases by 17.86% on average, which indicates an appropriate trade-off between exploitation and exploration.