Scene Representation, Registration and ObjectDetection in a Truncated Signed Distance FunctionRepresentation of 3D Space

This thesis presents a study of the signed distance function as a three-dimensional implicit surface representation and provides a detailed overview of its different properties. A method for generating such a representation using the depth-image output from a Kinect camera is reviewed in detail. In order to improve the quality of the implicit function that can be obtained, registration of multiple sensor views is proposed and formulated as a camera pose-estimation problem. To solve this problem, we first propose to minimize an objective function, based on the signed distance function itself. We then linearise this objective and reformulate the pose-estimation problem as a sequence of convex optimization problems. This allows us to combine multiple depth measurements into a single distance function and perform tracking using the resulting surface representation. Having these components well defined and implemented in a multi-threaded fashion, we tackle the problem of object detection. This is done by applying the same pose-estimation procedure to a 3D object template, at several locations, in an environment reconstructed using the aforementioned surface representation. We then present results for localization, mapping and object detection. Experiments on a well-known benchmark indicate that our method for localization performs very well, and is comparable both in terms of speed and error to similar algorithms that are widely used today. The quality of our surface reconstruction is close to the state of the art. Furthermore, we show an experimental set-up, in which the location of a known object is successfully determined within an environment, by means of registration.