Apparatus and method for detecting a face in a video image

• US 5,715,325 A
• Filed: 08/30/1995
• Issued: 02/03/1998
• Est. Priority Date: 08/30/1995
• Status: Expired due to Term

• Current Assignee: The Trustees of Princeton University (Princeton University), Siemens Corp. (Siemens AG)
• Original Assignee: The Trustees of Princeton University (Princeton University), Siemens Corporate Research Incorporated (Siemens AG)
• Inventors: Andress, Keith M., Fang, Ming, Schwartz, Stuart C., Bang, Richard D.
• Primary Examiner(s): Boudreau, Leo
• Assistant Examiner(s): MEHTA, BHAVESH M
• Application Number:
  US08/521,175
• Time in Patent Office: 888 Days
• Field of Search: 382/115-118, 382/155-156, 382/209, 382/199, 382/190, 382/276, 382/282, 382/254, 382/274-275, 340/825.34
• US Class Current: 382/118
• CPC Class Codes: G06V 40/161 Detection; Localisation; Normalisation

Abstract

A detection system identifies a person by rapidly analyzing an image frame from a video camera for the presence of a face by reducing the image in resolution and normalizing the image to compensate for lighting changes and compares the normalized image with a background image to produce a binary interest mask. The most likely position of a face is then determined by identifying a top, bottom and sides of a possible head region to define a bounding box, normalizing the pixel data within the bounding box to further compensate for lighting variations, and resizing the box to match identification templates stored in memory. The best match from the identification templates is provided with a score indicative of the degree of the match. If the score does not exceed a first score threshold, then no face is assumed present. If the score exceeds a first and second score threshold, then a face is assumed present. If the first score threshold is exceeded but the second is not, a further eye location procedure is performed on the data to determine if a face is present.

Claims

1. An identification system for identifying a face comprising:

a video camera for viewing a scene and providing at least one image frame of said scene;

locating means for locating a target region in said image frame that is capable of containing an image of said face;

memory means for storing identification templates of faces;

comparison means for comparing said target region to said identification templates, whereby a face is identified when said image in said target region substantially matches one of said identification templates, wherein said locating means comprises means for identifying an upper boundary of said target region by comparing pixel blocks contained in rows of said image frame with a threshold number value, means for identifying side boundaries of said target region by determining a medium position of pixel blocks contained within rows of said image frame, and means for identifying a lower boundary of said target region by averaging the numbers of pixel blocks contained in each row within said target region and identifying a row with a number of pixel blocks below average.

2. The system according to claim 1, further comprising a means for distinguishing the outline of a person in said scene, wherein said target region includes the head of the person in said scene.

3. The system according to claim 1, further comprising a means for normalizing said at least one image frame for changes in lighting.

4. A method for identifying a face, comprising the steps of:

producing an image of a scene with a predetermined resolution of pixels;

distinguishing an outline of a person in said scene comprising the steps of;

reducing said predetermined resolution of said image;

normalizing said reduced resolution image with an x-gradient to produce a normalized image insensitive to illumination variations;

applying background image data to said normalized image to produce a normalized image mask;

generating a reference mask from the motion analysis of consecutive images of said object;

comparing said reference mask to said normalized image to remove noise from said normalized image and produce a final interest image indicative of said outline and insensitive to light changes;

locating a target region in said image that may contain the image of a face;

comparing said target region of said image to at least one identification template; and

determining if said target region significantly matches said at least one identification template, wherein said step of locating said target region includes identifying an upper boundary, side boundaries and a lower boundary of said target region and wherein the step of identifying said upper boundary comprises comparing pixel blocks contained in rows of said final interest image with a threshold number value and identifying the first row wherein said threshold number value is exceeded, wherein the step of identifying said side boundaries comprises determining a first median column position of pixel blocks contained within rows of said final interest image indicative of said image horizontal start position and a second median column position of pixel blocks contained within rows of said final interest image indicative of said image horizontal end position, and wherein the step of identifying said lower boundary comprises averaging the numbers of pixel blocks contained in each row within said target region and identifying a row with a number of pixel blocks below average.

5. The method according to claim 4, wherein said step of producing an image comprises viewing said scene with a video camera, whereby said image is an electronic image produced by said video camera.

6. The method according to claim 4, further comprising the steps of normalizing said reduced resolution image with a y-gradient to produce a normalized y-gradient image.

7. The method according to claim 6, further comprising the step of performing a comparison between said x-gradient image and said y-gradient image to produce said normalized image.

8. The method according to claim 4, wherein said step of comparing said target region comprises the substep of providing said at least one identification template with a score indicative of how well said at least one identification template matches said target region.

9. The method according to claim 8, wherein said step of determining if said target region significantly matches said at least one identification template comprises comparing said score to a threshold score.

10. The method according to claim 4, further comprising the step of resizing said target region to a standard size that matches the size of said at least one identification template.

11. The method according to claim 4, further comprising the step of normalizing said target region for variations in lighting to produce a trinary image indicative of said final interest image for comparison with said identification templates.

12. The method according to claim 4, further comprising the step of searching said target region for the presence of an eye in said target region if said target region matches said at least identification template by a predetermined degree.