Component fusion for face detection

1. A method for detecting an object in an object detection device, the method comprising:

receiving a plurality of training images, at least one of the training images including a plurality of components of an object;

forming a statistical model from a mean and a covariance matrix of the components of each training image;

detecting a plurality of components in an input image using a component detector that is trained with the training images;

determining a covariance matrix for each of the detected components; and

matching a probabilistic observation comprising the covariance matrixes of the detected components with a probabilistic model comprising the mean and covariance matrixes of the statistical model to detect the object.

2. A method as defined in claim 1 , wherein the object is a human face.

3. A method as defined in claim 2 wherein said components are indicative of a left eye, a right eye, and a lower face, respectively.

4. A method as defined in claim 1 wherein said input image is a real-time image.

5. A method as defined in claim 1 , further comprising calibrating said statistical model for heteroscedastic noise.

6. A method as defined in claim 1 wherein the components are rectangular areas.

7. A method as defined in claim 1 wherein the training images include examples of the object at a plurality of different poses.

8. A method as defined in claim 1 , wherein the component detector is generated using AdaBoosting.

9. A system for object detection, the system comprising:

observation means for receiving a plurality of training images, at least one of the training images including a plurality of components of an object;

modeling means for forming a statistical model from a mean and a covariance matrix of the components of each training image;

a component detection unit for detecting components in an input image, wherein the component detection unit is trained using the training images;

a component fusion unit in signal communication with the component detection unit for fusing the components; and

a CPU in signal communication with said detection and fusion units for detecting the object by matching a probabilistic observation comprising the covariance matrixes of the detected components with a probabilistic model comprising the mean and covariance matrix of the statistical model.

10. A system as defined in claim 9 , further comprising:

receiving means for receiving the input image.

11. A system as defined in claim 9 , wherein the object is a human face.

12. A system as defined in claim 9 wherein said components are indicative of a left eye, a right eye, and a lower face, respectively.

13. A system as defined in claim 9 wherein said input image is a real-time image.

14. A system as defined in claim 9 , further comprising calibration means for calibrating said statistical model for heteroscedastic noise.

15. A system as defined in claim 9 , wherein the components are each rectangular areas.

16. A system as defined in claim 9 , wherein the fused components are aligned to a predetermined size.

17. A system as defined in claim 9 wherein the training images include examples of the object at a plurality of different poses.

18. A system as defined in claim 9 , wherein the component detection unit includes a classifier generated using AdaBoosting to detect the components.

19. A program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for object detection, the method steps comprising:

forming a statistical model from a mean M and a covariance matrix C of components of known example images;

scanning locations in an input image using the statistical model to determine component confidence maps;

identifying a local maximum from the confidence maps, wherein the local maximum has a location U;

estimating a covariance matrix Q in an area around the local maximum; and

matching a probabilistic observation comprising the location U and covariance matrix Q with a probabilistic model comprising the mean M and the covariance matrix C to determine whether the input image includes the object.

20. A program storage device as defined in claim 19 , wherein the object is a human face.

21. A program storage device as defined in claim 19 wherein said components are indicative of a left eye, a right eye, and a lower face, respectively.

22. A program storage device as defined in claim 19 wherein said input image is a real-time image.

23. A program storage device as defined in claim 19 , further comprising a program step for calibrating said statistical model for heteroscedastic noise.

24. A program storage device as defined in claim 19 , wherein the components are each rectangular areas.

25. A program storage device as defined in claim 19 wherein the training images include examples of the object at a plurality of different poses.

26. A program storage device as defined in claim 19 , wherein the component detector is generated using AdaBoosting.

27. A method as defined in claim 1 , wherein the probabilistic observation includes a location and the covariance matrix of the probabilistic observation is determined by:

scanning locations in the input image using the statistical model to determine component confidence maps;

identifying a local maximum from the confidence maps, wherein the local maximum includes the location; and

estimating the covariance matrix of the probabilistic observation in an area around the local maximum.