Browse Prior Art Database

Method for fast calculation of effective features for rapid object detection

IP.com Disclosure Number: IPCOM000008510D
Publication Date: 2002-Jun-18
Document File: 8 page(s) / 2M

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed is a method for fast calculation of effective features for rapid object detection. Benefits include improved performance.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 50% of the total text.

Method for fast calculation of effective features for rapid object detection

Disclosed is a method for fast calculation of effective features for rapid object detection. Benefits include improved performance.

General description

              The disclosed method includes a very fast calculation scheme of features for learning object-detection tasks such as face and people detection (see Figure 1).

Features

              The basic unit for testing for the presence of an object is a window of  WxH pixels. Assume that a very fast way of computing the sum of pixels of any upright and 45-degree rotated rectangle inside the window is available (see Figure 2).

              A rectangle is specified by the tuple

 
with the following:

              A rectangle’s pixel sum is denoted by:

              The raw feature set is the set of all possible features of the form:

where the following are arbitrarily chosen:

ri

              This raw feature set is infinitely large. For practical reasons, it is reduced as follows:

§         Only weighted combinations of pixel sums of two rectan­gles are considered (N=2).

§         The weights must have opposite signs and compensate for the difference in area size, that is,

              Without restrictions we can set
and get

§         The features should mimic HAAR-like features as well as early features of the human visual pathway.

              These restrictions lead us to the 14 feature prototypes (see Figure 3):

§         Four edge features

§         Eight line features

§         Two center-surround features

              These prototypes scale independently in verti­cal and horizontal direction. The line features can be calculated by two rectangles only. Hereto it assumed that the first rectangle r0 encompasses the black and white rect­angle and the second rectangle, r1...