Using statistical mixture models for tracking natural underwater boundaries

The paper presents signal processing and control algorithms that enable autonomous tracking of the boundaries between distinct benthic regions by an AUV equipped of a profiler sonar. By exploiting sonar scans of the region below the robot, a classical control loop is closed around the sonar data, using a feedback signal that is robust with respect to classification ''noise''. Introduction This paper presents work on the definition and implementation of a perception-based tracking behavior for underwater platforms, enabling an autonomous robot to track the boundary between distinct benthic habitats. This kind of autonomous observation behavior can be useful in a variety of applications, such as the observation of the patchy structure of some biological species, mapping of regions of contamination or of concentration of a given product, and avoidance of dangerous operational regions. While several underwater robotics groups have addressed in the past the problem of tracking perceptual features using vision sensors – mostly man-made structures, such as piplelines – we believe that the use of acoustic sensors provides a more robust approach, being able to operate in troubled waters and at larger depths, without the need to resort to artifical lightning. The ability to track natural features is, moreover, an intrinsic tool in the definition of completely autonomous navigation systems at a larger scale, in the framework of the approach proposed in [1,2]. In this work, we define an integrated perception/guidance approach for large scale navigation of autonomous robots that have no external measures of their geographical position (no GPS, nor transponders located at calibrated positions), completely based on the use of a map that is incrementally learned by the robot during its mission. This approach has been fully validated using a terrestrial mini-robot equipped of short-range distance sensors, operating in a planar environment where several mediumsized objects of arbitrarily curved shape are placed at distances which are very large compared to the range of the robot’s sensors. The robot incrementally maps the shape and the relative positions of the objects detected, and uses this map to periodically reset its position error. When executing its mission, the robot alternates between “useful” goal driven periods and periods of environment exploration. Exploration of the environment is triggered when the current representation is not rich enough to guarantee safe progression of the robot. In particular, the risk that the platform gets lost is permanently updated, and the necessary actions are undertaken to maintain its value below a prespecified threshold. The present work is a contribution to the extension of this approach to underwater environments, providing the guidance mode that enables direct autonomous acquisition of interesting detected features. In this paper, we concentrate on the definition of the perception guidance mode, presenting both the signal processing and control algorithms that support it. Future communications will address the problems of incremental map update and of map use for navigation. The paper is organised as follows. We first formulate the tracking problem. The next section is dedicated to data processing issues: how the statistical model of received data is learned, and how this model is used to compute a signal that indicates the offset of the robot with respect to the tracked boundary. The following section defines the controller structure. The subsequent section presents data acquired during real tracking experiments conducted at sea, illustrating the achieved level of performance. Finally, the last section summarizes the main results of the paper, and lists some ideas for future work. Problem formulation and approach Our goal is to define signal processing and control algorithms that enable autonomous observation of the boundary between distinct natural benthic regions (either corresponding to distinct species or to different materials) by an autonomous robot. Let L C ? ), ( be the 2D curve of length L, corresponding to the projection of the tracked boundary in an horizontal plane, parametrised by arc length . This curve is an abstraction of real boundaries between natural habitats, which are, most often, a transition region of varying width, inside which the characteristics of the sea floor, or the relative mixture of distinct species, gradually change from one type to another. A simple mathematical model of these transition regions is the dilation of ) ( C -the skeleton of the transition region – by circle of varying radius: