Fast Central Catadioptric Line Extraction

Abstract

Lines are particularly important features for different tasks such as calibration, structure from motion, 3D reconstruction in computer vision. However, line detection in catadioptric images is not trivial because the projection of a 3D line is a conic eventually degenerated. If the sensor is calibrated, it has been already demonstrated that each conic can be described by two parameters. In this way, some methods based on the adaptation of conventional line detection methods have been proposed. However, most of these methods suffer from the same disadvantages than in the perspective case (computing time, accuracy, robustness, ...). In this paper, we then propose a new method for line detection in central catadioptric image comparable to the polygonal approximation approach. With this method, only two points of a chain allows to extract with a very high accuracy a catadioptric line. Moreover, this algorithm is particularly fast and is applicable in realtime. We also present experimental results with some quantitative and qualitative evaluations in order to show the quality of the results and the perspectives of this method.