A probabilistic framework for tracking small objects in infra-red images

The tracking of objects in images is often performed solely on the trajectory of the object in the image plane, based on the use of the linear Kalman filter or nonlinear extended Kalman filter. Both rely on the availability a continuous stream of observations of the objects being tracked. The observations are assumed to have zero mean independent identically distributed (IID) noise, the variance of which is known or can be approximated. Difficulties arise if objects are detected only intermittently, false objects are detected or the trajectories of the objects intersect. Data association techniques have been devised to deal with such problems, but in them too, the association is usually performed primarily or totally on the basis of object trajectories. This approach may be suitable when considering the tracking of point features, but if extended objects such as blobs or line segments are being tracked, then there is additional radiometric information that should be incorporated into the tracking and data association process. In this paper we shall consider methods of including radiometric information in the tracking process using a joint probabilistic model for the two separate sources of information (intensity and trajectory).