BibTeX entry

@PHDTHESIS{201007Charles_Bibby,
  AUTHOR={Charles Bibby},
  TITLE={Probabilistic Methods for Enhanced Marine Situational Awareness},
  SCHOOL={Oxford University},
  MONTH=Jul,
  YEAR=2010,
  URL={http://www.bmva.org/theses/2010/2010-bibby.pdf},
}

Abstract

We present a system that uses probabilistic methods for enhanced situational awareness in marine environments. More specifically, we present significant contributions in the areas of Simultaneous Localisation and Mapping (SLAM) from the robotics community, and visual tracking from the computer vision community. We demonstrate how these theoretical contributions can be applied to the practical problem of marine surveillance. Traditionally, SLAM has been used to produce static maps of the environment and dynamic objects have either been filtered out and ignored, or detected and tracked separately. In contrast, we show how dynamic objects can be included directly in the SLAM estimate. We propose a hybrid representation that combines point features, occupancy grids and cubic splines. The point features are used to represent small stationary objects, the occupancy grid is used to represent landmasses and cubic splines are used to represent the trajectories of dynamic objects. We also present a new region-based, level-set framework, for visual tracking and segmentation. In contrast to all previous methods, we use the pixel-wise posterior when computing the foreground/background pixel membership, as opposed to the traditional pixel-wise likelihood. We show that this approach produces better behaved objective functions and hence provides more resilient visual tracking. We are able to track twelve or more objects in real-time, simultaneously estimating their position, rotation, scale, depth-ordering, figure-figure and figure-ground segmentations. Finally, we describe a prototype system that combines our work on SLAM and visual tracking, enabling a mobile vehicle to be used for marine surveillance. This system produces a hybrid map of the environment that contains both metric and visual information. The metric information represents the position, speed, shape and size of objects. The visual information represents the appearance of objects and is acquired using a high-performance pan-tilt device, which we have designed, built and tested. We use the pan-tilt device to automatically make visual contact with all objects within sensor range. The prototype system has been demonstrated protecting a river in the Thames Estuary against potential security threats.