3D EKF-SLAM Using Cylindrical and Planar Features With Multilevel Mapping Representation

This work presents an expansion of the traditional landmark-based EKF-SLAM for 3D indoor environments. It uses a mobile robot with an RGBD camera to decompose a scene and update its 3D location. A point-cloud map is reconstructed by initially simplifying the scene into cylinders and planes. Then, the captured points are segmented and associated to their specific EKF feature. This enables the output of a map using different levels of abstraction, which is helpful when working in applications with limited memory requirements. In our feature extraction method, the RANSAC algorithm is used for planar segmentation. The non-planar points are clustered using DBSCAN and parameterized as cylinders after passing through validation gates. The classical EKF framework propagates and updates the features’ states in each step. This also triggers the point-cloud growth and increments the point-cloud of the observed feature. Our simulation results show that this method successfully reconstructs maps, reducing the IAE criteria of trajectory error to 15% of that compared to the odometry. Finally, we successfully validate our proposal in a real-world environment using a ground robot with an RGBD Intel Realsense camera.