Learning OpenCV

The Condensation Algorithm

The Kalman filter models a single hypothesis. Because the underlying model of the probability distribution for that hypothesis is unimodal Gaussian, it is not possible to represent multiple hypotheses simultaneously using the Kalman filter. A somewhat more advanced technique known as the condensation algorithm [Isard98], which is based on a broader class of estimators called particle filters, will allow us to address this issue.

To understand the purpose of the condensation algorithm, consider the hypothesis that an object is moving with constant speed (as modeled by the Kalman filter). Any data measured will, in essence, be integrated into the model as if it supports this hypothesis. Consider now the case of an object moving behind an occlusion. Here we do not know what the object is doing; it might be continuing at constant speed, it might have stopped and/or reversed direction. The Kalman filter cannot represent these multiple possibilities other than by simply broadening the uncertainty associated with the (Gaussian) distribution of the object's location. The Kalman filter, since it is necessarily Gaussian, cannot represent such multimodal distributions.

As with the Kalman filter, we have two routines for (respectively) creating and destroying the data structure used to represent the condensation filter. The only difference is that in this case the creation routine cvCreateConDensation() has an extra parameter. The value entered for this parameter sets ...