Face in Video Imagery
                  

System Overview: Face detection from video imagery
(funded by TSWG, the DOD Combating Terrorism Initiative)

Director:                       
Professor Qiang Ji (qji@ecse.rpi.edu)
Technical Support:   
Peng Wang,             (wangp2@rpi.edu)

The goal of this project is to develop improved algorithms for face detection and
segmentation from video imagery for use in facial recognition systems under current
development by the government.  Our approach consists of those major modules:

(1).  Human detection and tracking via background modeling and particle filtering
(2).  Multi-view Face detection and tracking
(3).  Eye detection and tracking
(4).  Face recognition,  and performance modeling and prediction of face recognition


The goal of human motion detection is to identify motion candidate regions in the image
possibly occupied by humans, assuming the motion in the scene is primarily caused by
human movement.  We are exploring different methods,  such as background modeling
and optical flows estimation techniques. The particle filtering is applied to track person
motion.

Given the identified image regions corresponding to the detected human, we then
propose to use appearance-based face detection method. Two specific methods are
developed. We detect multi-view faces using discriminant features selected with AdaBoost.

The similar algorithm is also applied to eye detection after face is located. Combined support

 vector machine (SVM) is used to  detect the small face (head) in cluttering environments.

กก

Detected faces are tracked through video. We combine multiple measurements for robust 

multi-view face tracking, and online learn face appearance models for individual persons. 

The face pose can be simultaneously estimated during tracking.


The face and eye detection and tracking provide good reference points for face recognition.

A face recognition method based on local features is applied in the system.  

Another research on face recognition is to model and predict system performance so that the

image data that an existing face recognition system can not correctly recognize  will be 

removed from recognition.

Some results we have achieved so far are shown in this website.

Last updated: 12/6/2005