Image segmentation by unsupervised adaptive clustering in the distribution space for AUV guidance along sea-bed boundaries using vision

Abstract

We address the problem of autonomous underwater vehicle guidance along the boundaries of different benthic species using video information. This form of guidance provides a robust navigation behavior enabling observation of the occupancy of the sea bed independently of the presence of external position references. The major innovation of the work presented concerns the image segmentation algorithm. It is an unsupervised algorithm, which identifies clusters in the space of gray level probability distributions of image neighborhoods. The metric used to compare gray level distributions is the Kullback-Leibler directed divergence, which is related to the probability of confusing members of distinct clusters. The algorithm is self-tuned, in the sense that the number of clusters is automatically determined. It works adaptively, updating the classes' representations for each new frame, accommodating gradual lighting and texture variations within the same region. The visual controller, a simple integral law with saturation, controls heading rate to minimize the distance between the contour and the image center, while keeping a constant forward speed along the body axis. A separate controller (classic PI) keeps the robot at constant altitude from the sea bottom. The performance of the algorithm proposed is validated by real experiments conducted with the robot Phantom 500 XTL.