Multi-pose face tracking using multiple appearance models

1. A computer implemented method of tracking a face moving through multiple frames of a video sequence, the method comprising:

receiving into a computer processor a predicted position of a face in a video frame;

performing a search using the computer processor to determine color model similarity values around the predicted position of the face in the video frame;

performing a search using the computer processor to determine edge model similarity values around the predicted position of the face in the video frame; and

combining the color model similarity values with the edge model similarity values using the computer processor to determine a best match to track a face in the video;

wherein similarity values represent correlation values between two data blocks; and

wherein combining the color model similarity values with the edge model similarity values comprises multiplying the color model similarity values with the edge model similarity values.

2. The method of claim 1 wherein the correlation values are normalized prior to combining.

3. The method of claim 2 , wherein the normalization is done using predefined mean and variance, or computed minimum and maximum, or computed mean and variance of the correlation values.

4. The method of claim 1 wherein the edge model comprises a difference of Gaussians (DOG) model, or comprises of Laplacian of Gaussians (LOG) model.

5. The method of claim 4 wherein the DOG or LOG models comprise a DOG or LOG filters that has a high response around edges of an image.

6. The method of claim 1 wherein a YCbCr color space is used to represent faces in the color model.

7. The method of claim 1 wherein a particle filter is used to obtain a predicted position of a face in a video frame.

8. The method of claim 1 and further comprising tracking the face backwards, starting from the frame where the face was detected, until the face moves out of the frame, so as to construct a full track of the face.

9. A tracker for tracking faces in surveillance video frames, the tracker comprising:

a position predictor that provides a predicted position of a face in a video frame;

a color model that provides color model similarity values around the predicted position of the face;

an edge model that provides edge model similarity values around the predicted position of the face; and

means for combining the color model similarity values with the edge model similarity values to determine a best match around the predicted position to track a face in the video;

wherein the means for combining the color model similarity values with the edge model similarity values multiplies the color model similarity values with the edge model correlation values.

10. The tracker of claim 9 wherein the similarity values are normalized prior to multiplying.

11. The tracker of claim 9 wherein the edge model comprises a difference of Gaussians (DOG) model.

12. The tracker of claim 11 wherein the DOG model comprises a DOG filter that has a high response around edges of an image.

13. The tracker of claim 9 wherein a YCbCr color space is used to represent faces in the color model.

14. The tracker of claim 9 wherein the face is scaled for the color model as the face moves through frames of video.

15. The tracker of claim 9 wherein a particle filter is used to obtain a predicted position of a face in a video frame.

16. The tracker of claim 15 wherein the particle filter is updated using a uniform distribution of random numbers.