Samuel Janzon presenterar sitt examensarbete

Vehicle Tracking Using Bayesian Form Models
	
Abstract

Automatic tracking of vehicles in video sequences can be useful for
safety evaluations of intersections. In this report a model for
vehicle tracking in an intersection is suggested.  Data consist of
video sequences from an intersection; the videos have previously been
segmented into foreground and background by an online
EM-algorithm. Due to noise and imperfections in the images it is
important to utilize all available information to facilitate
tracking. Therefore a Bayesian hierarchical model is used to improve
tracking.  The data is modeled with truncated Gaussian distributions
and independence between pixels are assumed. A dynamic model handles
prior belief of appearance and how the tracked object is supposed to
move. The outline of the object is modeled using snakes, with speed
and center of the object modeled by a GaussianMarkov chain. The
dynamic model and data model are combined into a posterior
distribution and the maximum a posteriori estimates are used to
estimate model parameters. The resulting tracking is good as long as
disturbances are not to large. A drawback is that the model does not
handle multiple objects, since the tracking breaks down if there is a
cluttering of objects.