Abstract
A random scanning algorithm for tracking curves in image sequences is proposed. First it verifies the presence of a curve by votes, whose values indicate the number of the points on the curve, thus being able to robustly detect curves against outlier and occlusion. Furthermore, a procedure for preventing redundant verification is introduced by determining equivalence curves in the digital space to reduce the time complexity. Second it propagates the distribution which represents the presence of the curve to the successive image of a given sequence. This temporal propagation enables to focus on the potential region where the curves detected at time t-1 are likely to appear at time t. As a result, the time complexity does not depend on the dimension of the curve to be detected. To evaluate the performance, we use three noisy image sequences, consisting of 90 frames with 320 x 240 pixels. The results show that the algorithm successfully tracks the target even in noisy or cluttered binary images.