Automatic conversion of scanned sea charts into 3D models
Summary
Automatic nautical chart recognition and interpretation is a research topic that has been going on for many years. Nautical chart digitization has a variety of applications in navigation or in the development of navigational software, but also in educational applications like 3D training simulators which require a realistic representation of the seabed and its surroundings. This thesis presents a study on converting 2D scanned nautical chart images into 3D models. It is an exploration of some of the possibilities and problems occurring when designing and implementing such a system.
In order to obtain a 3D model, the scanned sea chart images have to be digitized. In digitizing a scanned sea chart, one of the major challenges is to properly separate and identify symbols on the map. The approached method first separates the background and the foreground pixels with a threshold-based segmentation method applied on the gray-scale image and then identifies individual objects in the image by searching for all connected components in the segmented binary image.
Another challenge is the classification of individual objects. The study brings a solution for the classification of different types of objects in a sea chart, focusing on the proper classification of spot soundings. Geometrical features like area, center of gravity, bounding box, density, orientation are used to build innovative decision rules that classify objects into several types of lines, characters or other symbols. The spot soundings are later recognized and interpolated to create a 3D surface of the maritime terrain. Tesseract OCR engine is used for character recognition. The spot soundings are interpolated using a method called Inverse Distance Weighting with Natural Neighbors. The interpolation method assumes that nearby points should have a greater influence than further away points. The nearby points are the vertices of the Delaunay triangle containing the interpolated point and are called natural neighbors.
The result of this research is a complete system that converts 2D scanned images into 3D simulation models. However, the performance of the algorithm is not 100% correct. Some issues remain and can be improved by further work.