A robust technique for detecting custom patterns of roundish features

Abstract

A fundamental task in computer vision is extracting low-level features from the image. Since this is one of the first tasks executed by most vision-based systems, imprecisions and errors committed during its execution are propagated to the next stages thus affecting the system overall performance. Therefore, robust and precise feature extractors are mandatory in computer vision. In the literature, two kinds of low-level features are commonly used: natural features, and artificial patterns of features. Natural features are extractable only from scenarios rich in textured elements. On the other hand, artificial patterns of features can be easily crafted by using commodity printers, which permits its application in a diversity of scenarios. Moreover, since the real dimensions of the pattern are known beforehand, the usage of artificial patterns allows the construction of metric systems. This thesis presents a new detection technique for patterns formed by roundish features. The new technique is composed of two stages: the extraction of candidates for features of the pattern; and the searching for the elements (among the candidates) that actually constitute the pattern. Differently from the techniques found in the related literature, the proposed one does not restrict the patterns to be rectangular grids of regularly-spaced features, but it allows the creation of a variety of patterns through the use of graphs (the pattern template). Experimental results collected from two case studies evidence that the new technique is robust to uneven and low-lighting conditions.