Distributed Dynamic Processing for Edge Detection

This paper discusses a dynamic method of edge detection which works from a sequence of frames. Most edge detection algorithms process image information statically, regardless of whether the application is static i.e. whether the input is a singular, unique image or a dynamic sequence of frames used for tracking or optic flow extraction. As many applications are dynamic, such as robotics, autonomous vehicle control and satellite tracking, it makes sense for edge detection processes to exploit dynamic phenomena. Employing dynamic processing offers a number of advantages, the main one being noise reduction. This paper discusses a dynamic edge detection process implemented as a network of simple processing units. In addition to edge detection, the network simultaneously determines optic flow and areas of occlusion and disocclusion. These areas provide information on how the image view is changing, which is useful for such applications as autonomous vehicle guidance.The integration of these two processes helps overcome problems such as feature matching. This paper describes mainly the edge detection process. Details of the optic flow processing have been described elsewhere ([3, 2, 4]). Following a description of the dynamic processing network, the results of this method are compared to the static edge detection scheme of Canny [1]. The network proves to be an efficient means of obtaining moving edges when a sequence of frames are available. Finally a recent extension for determining edges in images with multiple motions present is outlined. The Advantages of Dynamic Processing Noise Reduction The dynamic method used in this network is very robust to noise distortion. Conventional static methods determine edges at all intensity gradients with no notion as to whether a given 'feature' has permanance. The network described here extracts edges by exploiting the fact that noise does not normally correlate between frames i.e. noise at one point in an image is unlikely to be present at the same point in the next image (or at a neighbouring point when considering motion). Whilst most correlation methods correlate statically-obtained edge information in the spatial domain, this network uses temporal correlation, both to reduce noise and to achieve determine motion. Movement (through occlusion) helps to guide edge detection and vice versa. If a certain percentage of a given image frame is statistically identified as BMVC 1991 doi:10.5244/C.5.3