Face Detection and Recognition

Face Detection and Recognition

Specifications – Face Detection, Capture & Recognition

• No Retraining: When a new face is introduced into the face database, the training should be done on the newly introduced face only and already existing facial images in the database should not be retrained for computational efficiency and ease of use.
• Fast Induction of New Face: If a new person is introduced to the face database, it should not take more than 0.01 sec to train that face using a single core processor.
• Independent Feature Database: Once trained, the facial feature database should be separate from the face image database in order to protect it from potential corruption (tamper) of the images into the facial database.
• High Accuracy: The accuracy of the face recognition system should be at least 90% with respect to the FERET database.
• Resolution Independent: The face recognition algorithm should not be dependent on high resolution images, the video can be of any resolution, 640×480 or higher.
• Robust: The face recognition system should be robust to variations in facial expressions, luminance, pose, etc.
• Face Changes: Beard, moustache and other changeable face features influence face recognition quality and if frequent face changes are typical for some individual, face database should contain them e.g. face with beard and cleanly shaved face enrolled with identical ID.
• Lighting Condition: The photos should be captured in a good lighting for optimum performance, i.e. lighting equally distributed on each side of the face and from top to bottom with no significant direction of the light or visible shadows.
  • Eye-glasses: Glasses have to be of clear glass and transparent so the eyes and irises are clearly visible. Heavily tinted glasses and lighting reflections affect the recognition. However, face recognition typically works best when matching people with eyeglasses against themselves are wearing the same eyeglasses.
  • Camera Height and Angle: Video camera is directed toward face for optimum performance. Usually average height of a person is 5-6 ft. Video camera should be installed at the height of 6-8 ft. Distance of the person from the camera should be 10-20 feet as shown in the figure below.

• Camera Mounting Arrangement: The mounting base has to be strong and stable enough to avoid any vibration, movement, shaking, jerking, tilting or tampering of the camera.
• Permissible Horizontal Angular Deviation of Face: Camera should be focused straight to the human face so that both of the left and right parts of the face are equally visible. The maximum left-right shift is allowed not more than 1/20th of the distance in between the captured position and camera.

• Pose: The near frontal pose (full-face) must be used. Rotation of the head should be less than +/- 25 degrees from frontal in every direction – up/down, rotated left/right, and tilted left/right.
• Focus of the camera: The camera should be well focused so that the face is sharp with high contrast, clearly visible and distinguishable with naked eyes.
• Size of the Face: Size of the face should not be less than 1/5 height of the entire FOV for a D1 resolution (640 x 480) video or at least 100×100 pixels for higher resolution video.
• Face Database:
  • All the face images in the database should be sharp with high contrast, clearly visible and distinguishable with naked eyes.
  • The Face database design should be scalable and should be able to accommodate faces of more than 1000 persons.
• Face Image in Database: JPEG, JPEG 2000, RAW RGB, YUV, PNG, BMP, PPM and PGM.
• Input Image: Captured faces from the live video stream in any of the video formats - MJPEG, MPEG4, H.264, etc.
• Video Sources: The Face Recognition system should be camera agnostic and not particularly optimized for a particular brand, make and model of the camera. It should work with any IP camera with at least D1 resolution size of video. For analog camera, the system should be able to easily integrate with an encoder to do the face recognition in the server.
• Integrated with an Intelligent VMS:
  • Unified Architecture: The face recognition module should be integrated with a VMS (Video Management System) and the Face Capture routine should be in-built with the VMS. The captured faces from the VMS should be used for the recognition of faces.
  • Video Analytics: The VMS, Video Analytics, and Face Recognition should be designed into the same platform as a unified architecture. The face recognition should be integrated with video analytic applications. The video streams should be analyzed automatically, detect and localize face regions and send the cropped face to the face recognition system.