Particle filters are a powerful and widely used visual tracking technology. Their strength lies in their ability to represent multi-modal probability distributions that capture and maintain multiple hypotheses about target properties. A potential weakness, however, is that the particle set can become diffused, dispersing across the image plane rather than clustering around the target. A number of solutions to this problem have been proposed, including the use of the more recently developed Kernel Mean Shift tracker to guide particles towards a local mode. While this hybrid Condensation/Mean Shift tracker is effective, in most cases the Condensation component is an unnecessary overhead: Kernel Mean Shift is a competent tracker that only needs the particle filter to deal with more ambiguous situations in which errors might be made. We therefore propose an alternative hybrid approach in which Kernel Mean Shift is the dominant tracking technology, with a small number of particles being generated, in a structured fashion, to ex plore further and so resist errors when confidence in the Mean Shift algorithm is low. The proposed algorithm, which we term the Structured Octal Kernel (SOK) filter, has been implemented and is compared with Condensation, Kernel Mean Shift and Hybrid trackers. The SOK filter provides the most robust results, with comparable accuracy, at the lowest computational cost.