Robust 2D location of interest points by accumulation

Various interest point and corner definitions were proposed in the past with associated detection algorithms. We propose an intuitive and novel detection algorithm for finding the location of such features in an image. The detection is based on the edges in the original image. Interest points are detected as accumulation points where several edge tangent lines in a neighborhood are crossing. Edge connectivity is not used and thus detected interest points are robust to partial edges, outliers and edge extraction failures at junctions. One advantage of the approach is that detected interest points are not shifted in location when the original image is smoothed compared with other approaches. Experiments performed on Oxford and Cambridge reference databases allow us to show that the proposed detection algorithm performs better than 9 existing interest point detectors in terms of repeatability from multiple camera views.