STREAMING CLUSTERING ALGORITHMS FOR FOREGROUND

DETECTION IN COLOR VIDEOS

Zoran Duric, Wallace E. Lawson and Dana Richards

Department of Computer Science, George Mason University, Fairfax, Virginia, USA

Keywords:

Streaming algorithms, clustering, backgroundmaintenance, video surveillance.

Abstract:

A new method is given for locating foreground objects in color videos.This is an essential task in many appli-

cations such as surveillance. The algorithm uses clustering techniques to permit ﬂexibility and adaptability in

the description of the background. The approach is an example of the streaming data paradigm of algorithms

design, which only permits limited information to be retained about previous video frames. Experimental

results show that it is an effective and robust technique.

1 INTRODUCTION

Many authors have developed methods of detect-

ing people in images (Haritaoglu et al., 1998; Wren

et al., 1997); a comprehensive survey (Moeslund and

Granum, 2001) reviews most of the relevant refer-

ences. Most of this work has been based on back-

ground subtraction using color or luminance informa-

tion. Recently, edge information has been used for

background subtraction (Jabri et al., 2000; McKenna

et al., 2000) These methods usually use a number of

frames to ”learn” a model of the background scene

which is later used to classify pixels in new images as

either a background or a foreground. These methods

assume that the camera does not move from frame to

frame since any movement of the camera or the back-

ground objects could cause static parts of the scene to

be classiﬁed as a moving foreground. The results fre-

quently suffer from false positives/negatives and re-

quire additional post-processing to remove false ob-

jects and/or holes. In this paper, we present a novel

moving object detection and tracking method.

In this paper we explore another technique for rep-

resenting the background image and we use it suc-

cessfully to do foreground detection. A primary mot-

vation for our technique is that representing the back-

ground by the mean image is not effective in many

applications. A ﬁxed camera will typically experi-

ence vibrations. The background, even if “ﬁxed”, will

move sightly: leaves will ﬂutter, waves will shimmer,

and distant objects will move a little. Another motiva-

tion is that a ﬁxed camera, such as is used by surveil-

lence, will over time see a change in the “ﬁxed” back-

ground; lights will go on and off, the sun’s shadows

will move, and the color palette will drift.

To address these concerns in a natural way has led

us to use clustering methods. The general idea is to

represent each small patch of the frame by a small

set of exemplars, which are regarded as the centers of

clusters. Each disjoint cluster represents an equiva-

lence class of very similar patches; small variations

will be recognized as members of the same equiva-

lence class. Two disjoint clusters can represent dif-

ferent states of the same background; a ﬂag may furl

and unfurl. And, gradual changes can be addressed

by permitting the location of the cluster centers to be

self-adjusting.

2 BACKGROUND

This algorithm brings together two ideas, not previ-

ously used for foreground detection. The ﬁrst is the

use of clustering techniques. Clustering is a well-

studied problem with an enormous literature (see, for

example, (Duda et al., 2000)). Clustering can be re-

garded as a paradigm for unsupervised learning; the

486

Duric Z., E. Lawson W. and Richards D. (2007).

STREAMING CLUSTERING ALGORITHMS FOR FOREGROUND DETECTION IN COLOR VIDEOS.

In Proceedings of the Second International Conference on Computer Vision Theor y and Applications - IU/MTSV, pages 486-491

 SciTePress