Real-time image segmentation for action recognition

Abstract

In order to recognize actions of a human properly, an algorithm must be able to detect a person accurately in any given image. This data will be used to ascertain what action the user is performing. Many methods have been proposed that use a difference picture technique in which the image of the user is subtracted from a previously known image with the same background, without the user. These methods, however, are plagued with problems of noise and "ghosting", the undesirable introduction of image fragments as a result of changes in light intensity and moving objects in the background. The proposed algorithm combines several image processing methods in order to produce a clean difference image, while being far more robust to changes in light intensities and in the background scene.