Real‐time object detection with sub‐pixel accuracy using the level set method

This paper presents a method for object detection with sub‐pixel accuracy in digital images satisfying real‐time applications. The method uses an approximation of the level‐set‐based contour evolution, which applies only integer operations in a fast two‐cycle algorithm to achieve real‐time performance. Level‐set‐based contour evolution forms a part of the geometric active contour process, which combines image data and a priori geometric knowledge of an object to be evolved over an image towards the object border. The algorithm avoids the solution of partial differential equations. Instead only integer operations are applied and the contour evolution is performed by using a two‐cycle algorithm. The proposed method for object detection with sub‐pixel accuracy refines the pixel‐accurate result of the level‐set‐based contour evolution algorithm by using a fast parabolic fitting mechanism. Sub‐pixel accuracy of the resulting object contours is achieved by adjusting every contour point towards the nearby maximum of the image gradient. Real‐time performance can be provided because the adjustment is performed only once after the contour evolution. Experimental results with images from industry, photogrammetry, remote sensing and medicine show the functionality and applicability of this method to several fields of work. In addition, the experiments are evaluated by applying quality measures for the geometric accuracy and the run time.