Tracking dynamic boundaries by evolving curves
Carlo Tomasi, Tingting Jiang · 2007
Boundary tracking is the problem of representing dynamic boundaries in space and using observations to track the evolution of those boundaries over time. Boundary tracking has a wide range of applications. An example in image processing is the analysis of image sequences taken by a camera to track a moving object, such as a pedestrian. In the medical field, one can use ultrasound images, computed tomography (CT) images or magnetic resonance imaging (MRI) to track the contours of human organs, such as the heart. In forestry, a swarm of airplanes equipped with temperature sensors might track the boundary of a forest fire. Other applications include tracking spills of oil or poisonous gas, or clouds and weather patterns. In this dissertation, the author describes a new tracking framework which can give probabilistic estimates of the boundaries with topological changes by distributed sensing and strengthens the framework by prior information. The first part introduces the basic tracking framework called Level-Set Curve Particles which combines the known Level Set Methods and particle filters. Our main contribution is in controlling the potentially high expense of multiplying the cost of a level set representation of boundaries by the number of particles needed. The second part studies the observer dispatch problem in boundary tracking with distributed sensing. When the observers are mobile and controllable, it is desirable to have an optimal strategy which considers the change of the boundary, the time and energy cost of reallocations, and the distribution of the observers to maximize the benefit of their measurements. We formulate the observer dispatch problem as a utility optimization problem for which we propose a decision-theoretic solution. The third part focuses on the shape prior model and its application in boundary tracking problems. When measurements are noisy, it is important to refer to the prior information to keep the right track of the boundaries. We propose a new shape prior model, which is invariant to translation, rotation and scaling, and apply it to boundary tracking. The experimental results show that the shape prior model can make the tracking framework robust with respect to noisy measurements.