Analytical least squares Hough transform with an implementation on a transputer network

Computer vision for real-world imagery normally consists of three stages: acquisition/low-level feature extraction (e.g. capture followed by edge detection), medium-level feature extraction (typically into a geometric and/or topological space) and task-oriented scene understanding (e.g. aggregation of geometric features to characterise objects of interest). The Hough transform (HT) is an efficient medium level method to extract geometric features from an image which works fairly well for images that contain noise and occlusion. However, its performance decreases with image and parameter space quantisation noise. This paper describes two HT variants based on an analytical least squares refinement procedure that helps overcome some of these difficulties. A parallel implementation on a transputer based system is also discussed and evaluated