Detecting corners using the 'patchy' Hough transform

Relative vehicle motion estimation continues to be difficult. Current approaches are generally based on either optical flow or tracking feature points. In heavily carpentered environments such as institutional settings, the feature-point tracking approach requires significantly less computational effort. Algorithms which result in reduced computational effort are very attractive for the autonomous, vision-guided wheelchair project currently underway in our laboratory. We present a new method for detecting corners as feature points using a 'patchy' Hough transform. Corners are detected as intersections of straight lines extracted with the Hough transform. The image is initially segmented into patches, and only the intersections that occur within a patch due to edges also within the patch are deemed to be corners. This helps to avoid the localization problem inherent in the normal Hough transform. The use of patches also reduces the tendency of the Hough transform to ignore short edges because they are 'swamped' by longer ones. Edge pixels are initially enhanced by the Sobel operator, then detected using a histogram-derived threshold. The thresholded images are 'thinned' to produce edges of single-pixel width so as to avoid the line orientation ambiguity that arises with thick edges. The Hough transform is then computed for each patch. Peaks in the Hough accumulator array are generally poorly defined, with many adjacent accumulator 'bins' having similar values. To sharpen peaks, the Hough accumulator array is convolved with an 'annulus' kernel. The peaks are then detected as the local maxima of the convolution result. Finally, the intersections of all possible combinations of lines corresponding to the detected peaks are found. If the points of intersection lay inside the current patch, the lines are deemed to intersect. Algorithm performance is demonstrated with both synthetic and real image data.<
>