Real-time trust region ground plane segmentation for monocular mobile robots

Abstract

Ground plane segmentation is quite a challenging fundamental problem for monocular mobile robot navigation due to the dynamic unknown environments and the initialization of coordinate system which induces outliers to the bottom region of interest. Current geometric-based methods are mostly limited to deal with multiple plane segmentation in stationary known scene from depth sensor. In this paper, we propose a robust realtime trust region ground plane segmentation method to handle the unknown environments with a single camera. The proposed method utilizes Radius Outlier Removal filter to exclude the outliers of candidate points generated by the state-of-the-art method, Direct Sparse Odometry (DSO), then candidate points in the trust region are provided to fit the ground plane. The coefficients of fitted plane will be used to remove the outliers and to compensate omissive points. Therefore the ground plane segmentation is refined iteratively. Comprehensive experiments on the TUM monoVO dataset demonstrate that our method outperforms the random sample consensus (RANSAC) methods on time consumption and robustness in the unknown scenes, even when the initial coordinate system is pitched and rolled.