6 DoF SLAM using a ToF camera: The challenge of a continuously growing number of landmarks

Localization and mapping are fundamental problems in service robotics since representations of the environment and knowledge about the own pose significantly simplify the implementation of a series of high-level applications. ToF (time-of-flight) cameras are a relatively new kind of sensors in robotics. They enable the real-time capture of the distance and the grayscale information of a scene. Due to the increase of the image resolution of ToF cameras, now highlevel computer vision algorithms for visual feature extraction (e.g. SIFT or SURF) can be applied to the captured images. These visual features combined with the corresponding distance information give a full measurement of 3D landmarks. An obvious problem to be solved is the continuously growing number of landmarks. So far, all ever seen landmarks are just accumulated irrespective of their utility and the then required resources. Rather, one should keep only really useful landmarks, e.g. such that localization quality in the whole operational area is kept above a given threshold. In fact a lifelong running SLAM approach is dependent on means to select and discard landmarks. That is even more acute in case of feature-rich sensor data as provided with high update rates by sensors like a ToF camera. We run our SLAM approach in a real-world experiment within an indoor environment. The experiment was performed on a P3DX-platform equipped with a PMD CamCube 2.0 and a Xsens IMU.