INTELLIGENT VISION FOR MOBILE AGENTS Contour Maps in Real Time

Real time interpretation of scenes is a critical capability for fast-moving intelligent vehicles. Generation of contour maps from stereo disparity maps is one technique which allows rapid identification of objects. Here, we describe an algorithm for generating contour maps from disparity maps produced using an implementation of the Symmetric Dynamic Programming Stereo algorithm in an FPGA. The algorithm has three steps: (a) a median filter is applied to the disparity map to remove the streaks characteristic of dynamic programming algorithms, (b) irrelevant pixels in the centre of regions are marked and (c) selected contours outlined. Results for high resolution images (similar to 1Mpixel) show that a number of critical contours can be generated in less than 30ms permitting object outlining at video frame rates. The algorithm is easily parallelized and we show that multiple core processors can be used to increase the number of contours that can be generated.