CONSISTENCY CHECKS FOR PARTICLE FILTERS WITH APPLICATION TO IMAGE STABILIZATION
Ferdinand van der Heijden · University of Twente Research Information · 2005
An ‘inconsistent’ particle filter produces – in a statistical sense – larger estimation errors than predicted by the model on which the filter is based. Inconsistent behavior of a particle filter can be detected online by checking whether the predicted measurements (derived from the particles that represent the one-step-ahead prediction pdf) comply in a statistical sense with the observed measurements. This principle is demonstrated in an image stabilization application. We consider an image sequence of a scene consisting of a dynamic foreground and a static background. The motion of the camera (slow rotations and zooming) is modeled with an 8-dim state vector describing a projective geometrical transformation that, inversely applied to the current frame, compensates the camera motion. The dynamics of the state vector is modeled as a first order AR process. The measurements of the system are corner points (detected in the first frame) that are tracked. The particle filtering estimates the state vector using the measurements. However, the filter behaves inconsistently because a few corner points belong to the foreground. Using inconsistency checks these foreground points are detected and removed from the list of measurements.