Autonomous 3D Modeling of Unknown Objects for Active Scene Exploration

Abstract

The thesis Autonomous 3D Modeling of Unknown Objects for Active Scene Exploration presents an approach for efficient model generation of small-scale objects applying a robot-sensor system. Active scene exploration incorporates object recognition methods for analyzing a scene of partially known objects as well as exploration approaches for autonomous modeling of unknown parts. Here, recognition, exploration, and planning methods are extended and combined in a single scene exploration system, enabling advanced techniques such as multi-view recognition from planned view positions and iterative recognition by integration of new objects from a scene. In household or industrial environments, novel and unknown objects appear regularly and need to be modeled in order for a robot to be able to recognize the object and manipulate it. Nowadays, 3D models of hand-sized objects are usually obtained by manual scanning which represents a tedious and time consuming task for the human operator. For an autonomous system to take over this task, the robot needs to autonomously obtain the model within the object scene and thereby cope with challenges such as bad incidence angle, sensor noise, reflections, collisions or occlusions. In this thesis, sensor paths denoted as Next-Best-Scan are iteratively determined by a boundary search and surface trend estimation of the acquired model. In each iteration, 3D measurements are merged into a probabilistic voxel space, which considers sensor uncertainties. It is used for scene exploration, planning collision-free paths, avoiding occlusions, and verifying the poses of the recognized objects against all previous information. In order to account for both a fast acquisition rate and a high model quality, a Next-Best-Scan is selected that maximizes a utility function integrating an exploration and a mesh-quality component. The mesh-quality component allows for the algorithm to terminate once the quality required by the application is reached. The Next-Best-Scan algorithm is verified in simulation by comparison with state-of-the-art approaches concerning processing time and final model quality and in real scenes. The versatile applicability of the method is shown by several experiments with different cultural heritage, household, and industrial objects. Modeling of single objects is evaluated on an industrial and a mobile robot. On the industrial robot, the robot moves around the object, whereas on the mobile robot, the object is moved in front of an external range sensor using the same method. For modeling of larger workspaces, the mobile platform moves around the scene. The active scene exploration approach is demonstrated using several scenes with different levels of complexity. Here, Next-Best-Scan planning is performed for improving both recognition and modeling. Concluding, the developed methods enable the robot to learn object models of unknown objects, to directly apply these models to the individual application and therefore to become more autonomous. Here, the autonomously acquired object models are successively inserted into an object database and utilized by an object recognition module.