Rapid Non-Template Edge Recognition in Range Image

Traditional edge detection methods often employ self-defining templates to recognize edge points in 2D images. Recently, this technique has extended to develop many 3D edge operators used to understand 3D range images in computer vision community. Existing template-based approaches are mostly variant to rotation or scaling transformation and sensitive to noise in raw data. In this paper, a non-template edge recognition method for 2D scan lines in range images is proposed. Based on point's distribution analysis in a single scan line, some new coefficients are defined to characterize features of large noise and edge points. Subsequently, corresponding discriminant criteria are stipulated for denoising and recognition. The method performs an extremely rapid speed for data processing based on simple calculations on few points, and reaches equilibrium of strong robustness and high adaptability. Because all the coefficients are established on the basis of distances between the points or their ratio, the proposed method is inherently invariant to translation and rotation transformation, and is suitable for any sensor viewpoints. Computational complexity and time expenditure are evaluated during a real data experiment.