John MacCormick and Michael Isard
Partitioned sampling is a technique which was introduced in (MacCormick and Blake, 1999) for avoiding the high cost of particle filters when tracking more than one object. In fact this technique can reduce the curse of dimensionality in other situations too. This paper describes how to use partitioned sampling on articulated objects, obtaining results that would be impossible with standard sampling methods. A new concept relating to particle filters, termed the survival rate is introduced, which sheds light on the efficacy of partitioned sampling. The domain of articulated objects also highlights two important features of partitioned sampling which are discussed here for the first time: firstly, that the number of particles allocated to each partition can be varied to obtain the maximum benefit from a fixed computational resource; and secondly, that the number of likelihood evaluations (the most expensive operation in vision-based particle filters) required can be halved by taking advantage of the way the likelihood function factorises for an articulated object.
Another important contribution of the paper is the presentation of a vision-based “interface-quality” hand tracker: a self-initialising, real-time, robust and accurate system of sufficient quality to be used for complex interactive tasks such as drawing packages. The tracker models the hand as an articulated object and partitioned sampling is the crucial component in achieving these favourable properties. The system tracks a user's hand on an arbitrary background using a standard colour camera, in such a way that the hand can be employed as a 4-dimensional mouse (planar translation and orientation of thumb and index finger).
|Published in||Proc. European Conf. Computer Vision|