Introduction

The Intelligent Computer Interface (ICI) system (figure 1) is a long term research project aimed at combining computer vision (face recognition), animation, speech recognition and speech synthesis in an AI framework. The overall goal is the develop a computer program which behaves "intelligently", and demonstrates some of the possiblities with AI technology. The objectives are mostly teaching purposes, in the form of student projects and case studies in Gjøvik College' AI-courses¹. The project was initiated in January 1998, and this report describes the first approach taken to the face detection part of the system.

Face detection is a specific problem in the more general class of computer vision problems known as object detection and tracking. The task is to detect, localize and track a human face real time from a video sequence which is obtained from a digital video camera connected to the computer. Several approaches have been suggested to solve this task, a comprehensive review is presented in [2]. Most of these approaches can be divided into two groups:

Exhaustive scanning

Each single image-frame is scanned at almost every subframe size and a classifier determines at each subframe if a face is present or not.

Motion and low level image analysis

From motion analysis of a video sequence (e.g. frame-difference) and image analysis techniques such as edge or gradient operators a face is deemed present after applying certain heuristics (higher-level image analysis), perhaps together with color information.

In this paper we propose a simple technique which is a combination of motion analysis and scanning. Since our goal is to have a system performing real time, we cannot apply exhaustive scanning, thus we use motion analysis and heuristics to get an initial guess at the location of a face. We then scan the detected location at as many resolutions as our computer allows us to do in real time, and use the reconstruction error from a eigenface basis to determine the presence of a face.

The following sections describes the system in further detail. Section two explains motion analysis, section three lists the heuristics we apply (the "AI part" of the system) and section four describes the pattern recognition algorithm used for classfying a subframe as a face or non-face.

  

Figure 1: A screenshot of the ICI system.