Direct Range Image Processing (DRIP)

Edge detection or, more commonly, feature extraction can be readily performed on intensity images where the pixels are regularly placed like squares on a chessboard. More recently, the world of imaging and computer vision has moved towards the use of range data, obtained using a range camera or sensor. This range data is not regularly spaced, but instead is slightly randomly spaced and some of the required information may be missing. In order to perform any type of image processing, such as feature extraction or segmentation, on such irregularly spaced data, the data must be re-aligned mathematically on to a regular lattice, and in some cases the missing data is reconstructed. It not only takes time to perform these calculations, but such calculations can introduce approximation errors and data mis-representations. To avoid this unnecessary computation and error introduction, this project proposes a technique based on the use of finite element methods (FEMs) that will enable feature extraction operators to be generated that can be applied directly to the range image without any such pre-processing, thus proving to be more appropiate for real-time vision with the application of moving robots. This project will initially develop and implement the operators for use on irregular data and evaluate them in comparison to other existing techniques available. The project will then address the issue of finding out what type of feature has been found in the range image. The reason for this is that range image contain various types of edges: roof, jump, crease and smooth edge. Each of these features has different characteristics that must be found in order to determine the type of feature in the image. This is an important aspect of this project as most existing research focuses only on finding crease and jump edges in range images and not roof or smooth edges.Many object recognition systems used today are based on segmentation algorithms. Rather than a robot being able to determine precisely all the detail of a scene, it segments the scene into recognisable regions or objects and tries to match them with objects that it has seen before. On perfecting the finite element based feature extraction techniques and evaluating their ability to accurately characterise the features found in the range images, this technique will be used for segmentation of range data. The technique, combined with a simple edge-linking algorithm, should provide enclosed regions and reduce over or under segmentation.In order for this research to be appropriate for real-time imaging and hence useful for developing robot vision systems, it is required that the programs are coded in the C++ programming language, where the programmer has control of garbage collection and the time that it takes the program to run can be easily measured.Overall, this project aims to provide feature extraction operators that can be applied directly to range data for range image processing without the pre-processing steps that are currently essential to other techniques. This will reduce the mathematical computation required and thus enable improved real-time vision that can be useful for developing robot vision systems.