A phase coded disk approach to thick curvilinear line detection

This paper examines the well-known problem of line detection, but where the lines are wider than one pixel. The motivation behind the paper is the extraction of road information from high resolution photogrammetry and Light Detection and Ranging (LIDAR) data. Wide lines cause varying problems during detection. The HOUGH or RADON transform approaches do not find the road centrelines accurately; diagonals of the thick lines are found instead whilst other methods also tend to be error prone. Our approach convolves a raw, pixelated, binary road classification with a complex-valued disk. The technique provides three separate pieces of information about the road or thick line: the centreline, the direction and the width of the road at any point along the centreline. The road centreline can be detected from the position of the peak of the magnitude image resulting from the complex convolution. Road width can also be estimated from the magnitude peak whilst the direction of the road may be obtained from the phase image.