Adaptive termination of voting in the probabilistic circular Hough transform

Reliable detection of objects using the Hough transform is often possible even if just a small random poll of edge points is used for voting. This can lead to significant computational savings. To reduce the risk of errors, it is customary to preset the poll size to a value that is much larger than necessary in average conditions. An adaptive setting of the poll size in the probabilistic Hough transform is suggested. It is experimentally demonstrated that by monitoring changes in the ranks of peaks in the parameter space, sensible decisions on voting termination can be made. Adaptive stopping leads to polls that are on average smaller than the fixed poll that leads to the same error rate. In many applications the number of objects to be detected is unknown. Finding the number of appearances of an object in a noisy image is difficult, especially with partial data. The authors present an adaptive stopping rule that terminates voting as soon as any number of objects seem to be reliably detected, even though the existence of others may not be ruled out yet