Skip to Main Content
Tracking the pose of human face has long been an important research topic which has many important applications, and it is particularly challenging with a monocular camera because the depth information is lost due to the perspective projection. This work adopts particle filter with partitioned sampling to decompose the state space of face pose tracking into two subspaces for increasing the sampling efficiency, thus achieving satisfactory performance with fewer particles. The parameters in the first subspace describe the target on image plane, and the parameter in the second subspace is used for the estimate of the face pose in yaw angle direction. For the evaluation of each hypothesis in the second subspace, a statistical learning algorithm called relevance vector machine (RVM) is used to map a face containing image to the pose of the face. The training of RVM is tailored to each detected frontal face, and it takes less than half second, which is suitable for a real-time application. The learning based regression model also presents the insensitive ability to expression variation and unmodeled degree of freedom. The experimental results verify that the combination of particle filter and RVM can efficiently reduce the processing time and add robustness to the performance of the system, thus making this algorithm applicable to human-machine interface with low-cost webcams.