Towards a robust visual SLAM approach: Addressing the challenge of life-long operation

Localization and mapping are fundamental problems in service robotics. Knowledge about the own pose and representations of the environment are needed for a series of high level applications. Service robots should be designed for life-long and robust operation in dynamic environments. The contribution of this paper is twofold. First, an approach to address the ever growing number of landmarks in life-long operation is presented. Typically, SLAM approaches just accumulate features over time and do not discard them anymore. Therefore, the required resources in terms of memory and processing power are growing over time. In our approach, the absolute number of landmarks can be restricted by an upper bound since we introduce a method to specifically select and replace landmarks once the upper bound has been reached. The second contribution is related to improving the robustness of the landmark assignment problem in case of image based features as needed with natural landmarks. The approach has been successfully evaluated in a real world experiment on a Pioneer-3DX platform within a complex unmodified indoor environment.