RGB-D SLAM based reconstruction of 3D objects

Abstract

An important part of robot design is its vision, so it is important to study and compare the methods for vision. RGB-D Simultaneous Localization And Mapping (SLAM) based methods are one way to take the color and depth data and compute make a 3D model of a scene. Most of the RGB-D SLAM based methods are designed to model a large area such as a room, but this thesis focuses on comparing RGB-D SLAM based methods on a smaller scale. The pur-pose is to study and compare the methods by reconstructing single objects. The chosen ob-jects vary in size and complexity to test the methods more broadly.

There were two methods that we were able to get working in reasonable time and thus were included in the comparison. The first was Kinect Fusion which is a part of a Point Cloud Library (PCL) and the second was Static Fusion which is its own method. RGB-D data was gathered by using an Orbbec Astra series sensor. The criteria that was used to evaluate the methods were ease of install and the quality of the 3D model.

The main result is that both methods are good but depending on what the priorities are one was superior to the other. Static Fusion was fast and easy to install and to get working, does not depend on many libraries and is simple in design. Kinect Fusion as its part of a whole brings many other neat and useful features in addition to its RGB-D Slam method, which can be a pro or a con.