Online Learning Based Long-Term Feature Existence State Prediction for Visual Topological Localization

Visual localization of autonomous robots in dynamic scenes is a critical problem to be solved. Unlike the existing methods that regard the dynamic changes as outliers, we explore the hidden regularities of the long-term dynamic changes, and propose a new visual topological localization system. According to the regular pattern of feature existence changing with time, its feature existence matrix is constructed incrementally, which is applied for modeling the time-varying states of each feature. In particular, a new online learning-based modeling method is proposed, whose parameters are online updated by constrained Newton step adaptively. The feature sets with the largest existence possibilities are predicted for accurate topological localization. Further, extensive experiments are performed on both simulated and real measured datasets, the results verify that our method outperforms the state-of-the-art methods in the prediction and localization accuracy, and it achieves competitive real-time performance.