A versatile object tracking algorithm combining Particle Filter and Generalised Hough Transform

This paper introduces a new object tracking method which combines two algorithms working in parallel, and based on low-level observations (colour and gradient orientation): the Generalised Hough Transform, using a pixel-based description, and the Particle Filter, using a global description. The object model is updated by combining information from a back-projection map computed from the Generalised Hough Transform, providing for every pixel the degree to which it may belong to the object, and from the Particle Filter, providing a probability density on the global object state. The purpose of the proposed tracker is to make the most of the two algorithms, in terms of robustness to appearance variation like scaling, rotation, non-rigid deformation or illumination changes.