Face Detection in Meeting Room Using Omni-directional View

omni-directional image, transformation, face detection, recognition, computer vision

Gabor wavelet networks are used for face detection and recognition in omni-directional images, which are captured in meeting room. This meeting room is equiped with hyperbolic mirror and standard video camera for capturing meeting events.

Monitoring meeting rooms usually requires several video cameras so that the entire space is "covered". Position and gaze of faces are mostly the main source of information for monitoring and action detection. An alternative method is to use one video camera system with special optics - "omni-directional view" system. Such system can consist, e.g., of digital camera and mirror. The cost of such system is usually much lower compared to a multiple standard camera setup. Moreover, an advantage of such setup is portability and easy installation in the room. On the other side, it achieves lower effective resolution of videos and suffers adverse image distortion. These disadvantages can be partially solved by using high definition (HD) cameras with larger resolution than standard DV cameras and by applyng advanced geometrical transformations.
The poster will present description and evaluation the efficiency and behaviour of the tracking and face detection algorithms based on colour detection and Gabor wavellet networks [GWN] on the omni-directional images that have different features than the "standard" captured images. The difference from the "standard" images is mainly the resolution in the radius direction given by the mirror geometry and in the illumination. Additionally, distortions of the captured objects caused by different distances to the mirror can be observed in the images. Methods for geometrical corrections of the images when they need to be presented human are also presented.
Faces and hands detection methods for the omni-directional view are based on skin colour detection with automatic colour statistics that produce the input for the GWN recognizer that determines presence of face (or even a specific person). The GWN output can also be used for specification of the head orientation. These techniques are being tested on the IDIAP meeting data and on the data acquired by our omni-directional system. Results will serve for evaluation and comparing of the usability these techniques together with omni-directional system.