Geometrical Features based Visual Relocalization for Indoor Service Robot

In traditional SLAM methods, the environment map is the simply assembled points or lines, which makes it difficult to directly perform relocalization using such map. This paper presents a new implementation method for indoor environment representation and visual relocalization using RGB-D sensor. The method is developed for indoor service robots to perform relocalization using the observed point and line features. In this paper, the sparse feature map, line segment map, and dense point cloud map of an environment are learned by a random forest to regress the correspondences between visual features and 3D world locations, geometrical features and 3D world locations. Using the learned forest, landmark positions are efficiently predicted and the camera poses are then estimated in a probabilistic framework. The performance of the proposed method is demonstrated under various challenging environments using public benchmark dataset and our own dataset collected in an office environment. These conditions contain ambiguous areas, long corridor, moving people, viewpoint changes, or illumination variation. The proposed method is thoroughly evaluated against several strong state-of-the-art baselines. Experimental results prove the efficacy of our method.