Automated 3D Reconstruction of Unknown Objects by Planning the Next-Best-View based on a Spatial Data Structure

Kurzfassung

Autonomous 3D object reconstruction is a special field of robotics that has come up in the last decades. Since then, scientists have not been able to find and algorithm that fully meets the requirements of autonomous 3D object reconstruction. For more than 20 years many people all over the world have been taking part in the search for the perfect scan algorithm. 3D-modelling and 3D-reconstruction are established in industrial, medical, architectural and cultural areas. Many scanning systems are manually operated.Thereby a human operator moves the system along the object’s surface, but this process is very time consuming and monotonous. Different approaches and mathematical algorithms have been developed in order to scan unknown objects automatically, so money and time can be saved. The industrial market offers a variety of applications in the field of product quality control. Other areas like robotics and autonomous systems want to develop a system that localizes and orients itself in order to identify or grasp known or unknown objects. The objective of these applications is to develop a scanning algorithm that is able to reconstruct unknown objects in order to get a complete 3D model. The big challenge is to make the entire process run automatically. First, the robot’s workspace needs to be explored in order to locate the object. During the scanning process, the robot should calculate different scan positions to get a good viewpoint to the object and to scan the object entirely. This problem is known as the “view planning problem” and has been investigated by several researchers since the 1980s. The scanning process of an object is not difficult for a human operator, if a real-time visual feedback of the previously scanned areas is available. Therefore the quality of scanning depends on the experience of the operator and in most cases more time is needed and the precision is lower as compared to a robot. From a robot’s perspective the view planning process is a very complex problem. Based on the scanned data and the defined parameters the system has to decide where to position the sensor next in order to scan as many new areas of the object as possible. This scenario is also known as “next-best-view (NBV)”, with the goal to get a complete model of the object by the smallest number of views. This work focuses on finding a new way of object scanning and 3D object reconstruction based on different related approaches. Therefore, already existing and well working approaches are analysed in order to identify their advantages and disadvantages. In a second step, we use these findings to develop a new approch, consisting of a combination of several approaches, that allows to improve the end result. Some combinations will be presented in the scope of this work.