A Divide-and-Conquer Approach to Contour Analysis and Extraction of Invariant Arc Features

This paper presents a newl divide-and-conquer method to analyze geometric shapes, contours and trajectories extracted as a discrete sequence of points (or pixels) from images or spatial sensors such as GPS or transponders. The method converts contour point sequences into Circular Augmented Rotational Trajectories (CART) and uses a scale invariant analysis to extract invariant arc features. The arc feature is a generalization of scale invariant corners used in many object recognition and image matching methods. The method considers detection of corner like features in the presence of slow curvature and sharp noise and discretization (spatial quantization). The resulting feature vectors can be used for stable and robust object feature analysis and object detection. The divide and conquer CART method presented in this paper is order of magnitude faster than previously presented CART methods and offer more robust and precise representation of shapes. The presented method is found to be capable of ignoring local sharp noise and detecting globally prevailing sharp features. Experimental analysis confirms the efficiency and robustness of this method using several difficult shapes with considerable noise and ambiguity. The method allows not only stable feature detection but also general shape analysis such as convexity, linearity and curvature.