A robust registration algorithm based on salient object detection

Runzhao Yao, Shaoyi Du, Teng Wan, Wenting Cui

Abstract: Point cloud registration plays an important role in 3D computer vision. A challenge in this field is the presence of small salient objects with huge flat backgrounds in point clouds, which may result in poor registration. Despite substantial approaches for point cloud registration have been proposed, few attempts have been made to address this problem. In this paper, we present an approach which fully leverages not only geometric information but also texture information presented by RGB images to tackle with this problem. To mitigate the influence of background, we introduce a superior 2D salient object detection method to highlight the role of the salient objects. The color supported generalized iterative closest points algorithm is the state-of-the-art approach in iterative closest points (ICP) variations, which can efficiently exploit the color information. However, it cannot deal with the mentioned problem. On this basis, we further propose a joint objective to align both salient color points and background points. The registration and reconstruction experiments demonstrate the robustness and accuracy of our method.