Statistical snakes: robust tracking of benthic contours under varying background
Stefan Rolfes, Maria João Rendas · 2004
This paper presents the formalism of statistical snakes for contour estimation, and applies it to the problem of tracking a natural benthic contour using an AUV equipped with a video camera. To close the control loop that maintains the vehicle on the contour, we need to estimate its location on each frame of the video sequence. We propose a new criterion for contour estimation which is appropriate to the complexity and variability of natural environments. The criterion is based on a non-parametric statistical modeling of the regions adjacent to the tracked boundary, as mixtures of the probability distributions corresponding to the areas on each side of the contour. It is shown on the paper that minimizing the proposed criterion leads to an estimated contour such that the regions on each side have mixture coefficients close to zero and one (meaning that they are "pure" regions). Examples of application of the proposed algorithm to a real underwater image sequence are given.