This Project deals with real-time face detection of a person entering room.
Abstract
This Project deals with real-time face detection of a person entering room considering the following problems:
1) Varying light conditions.
2) Changing face angle when entering room.
3) Running time has to be short.
4) Timing of picture sampling.
Working environment
– x586 computer
– Windows 98
– Fly-Video card
– Standard Video camera
Application Structure
The project contains 3 components:
CamCapture – A process that is responsible for displaying the video stream on the monitor and perform a ‘capture’ in corresponding to a request that is received from MotionDetector. The ‘capture is sent to MotionDetector via the clipboard and a Windows message is sent to MotionDetector in order to perfrom the pull of that capture.
MotionDetector – A process that is responsible for the motion processing. This processor requests an image snap from CamCapture and after capture is received it processes the motion using an algorithm that applies on this last image that arrived with comparison to the previous image that was received.
As MotionDetector recognizes motion it sends a Windows Message to FaceDetector to start detecting.
FaceDetector – A process that gets the image of the person entering and finds its face in the image. It gets a Windows message from MotionDetector when a motion at door is detected and starts working. Once face detected it’s saved to a file. The detection time is approximately 0.5 second, so another person can enter the room and be detected after a reasonable time.
The algorithm
We used a ‘mask’ of an average human face we made out of picture sampled at lab:
by ‘running’ the mask over the picture we checked the Minimum Mean Square Error (MMSE) and this gave us the area where face could be.
Then we cut the face area and in the same way looked for MMSE with an average masks of eyes and nose.
If minimums found for eyes and nose, we checked the position of the ‘eyes’ and ‘nose’ comparing to a standard face.
Results
The system finds almost every person entering in a short running time and even with a light angle of face.
There is a problem with timing the right moment of sampling the person image, but with this method (A single camera) it’s reasonable. Others kinds of motion detectors can be more useful.
Acknowledgments
We would like to thank our supervisor Johanan Erez and also Dr. Michael Elad for helping us with the algorithm.
Also we would like to thank the Ollendorff Minerva Center which supported this project.