LS-ATR: Autonomous Target 3-D Reconstruction System Based on Fusion of Low-Cost Sensors

Target 3-D reconstruction is a common requirement for monitoring and exploration tasks, where the low-cost small unmanned vehicles have been popular platforms for civil applications. However, the existing methods rarely deal with the online target 3-D reconstruction constrained by the limited resource and payload of the small unmanned vehicles, and thus fail to provide a well-designed tradeoff between the accuracy performance and the cost. In this article, a novel autonomous target 3-D reconstruction system is developed based on fusion of low-cost sensors, a monocular camera, and a 2-D laser scanner (LS-ATR). First, a segmentation-based Gaussian process regression method is proposed to reconstruct a dense point cloud of an interested target segmented from the background by fusing the 2-D image and a sparse point cloud, which in the meantime provides an uncertainty evaluation model of the current reconstructed dense point cloud. Second, to improve the reconstruction performance online, a next-best-scan selection method is proposed by maximizing the uncertainty reduction via the rotation control of a gimbal connected with the scanner. Finally, a low-cost 3-D reconstruction prototype system is realized, and the reconstruction of targets in both the public dataset and our own dataset is carried out to validate the effectiveness and superiority of the proposed methods.