Abstract

Efficient curve detection and feature extraction is a very important step in many videorelated applications, such as video content analysis and representation, surveillance systems, medical diagnoses, etc. For example, in video surveillance systems, curve tracking and feature extraction can be used in detecting moving targets from a video, allowing potential interesting events to be identified and analyzed for surveillance purposes. Curve detection usually includes edge detection and post processing procedures such as thinning, curve fitting or edge following, etc. Curve detection can significantly reduce less important data in a video frame while preserving structural information. Perceptual features can be extracted from curves for analysis or recognition purpose. However, Conventional edge detectors provide only an output of edge pixels. It is difficult to extract perceptual features directly from the edge detection results. Post-processing is then needed to remove noise, fill gaps, and fit edge pixels into curves. Unfortunately, most post-processing is too timeconsuming for use in real-time applications (Fan et al., 2001). Most edge detection techniques fall into two categories, gradient based methods and second order methods. Gradient-based methods detect edges based on the first derivative of the intensity. Examples include the Sobel, Prewitt, Roberts, and Canny operators, in which the Canny operator (Canny 1986) is the one of most commonly used edge detector. The second order methods find edges by searching for zero crossings in the second derivative of the intensity. Examples of the second order methods include the Laplacian, Marr-Hildreth operators, etc. In color images, the color information also can be used to determine discontinuities in the color space (Cheng et al. 2001). Perez and Kock claimed in (Perez & Koch, 1994) that hue in HSI is more robust to certain types of highlights, shading, and shadows than the components in RGB, normalized RGB, or CIE color spaces. The edges with small hue change are removed from the Canny detector output in (Perez & Koch, 1994). A compass operator is proposed in (Ruzon & Tomasi, 1999), which considers distribution of pixel colors during edge detection. A 2D edge detection functional is used in (Qian & Huang, 1996), which is guided by the zerocrossing contours of the Laplacian-of-Gaussian (LOG) to find the edge locations. In curve feature extraction, the Hough Transform is a well known technique for detecting curve features. It transforms the image space into the parameter space to find possible

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call