Introduction

A fundamental problem in the design of autonomous mobile cognitive systems is to perceive the environment. A basic part of perception is to learn, detect and recognize objects, which must be done with respect to the limited resources of a mobile robot and the limited choice of available kinds of sensors. The performance of a mobile robot crucially depends on the accuracy, duration and reliability of its perceptions and the involved interpretation process. This paper describes a real time capable, color and scale invariant object learning and detection scheme. The arrangement of Haar-like features of objects is learned. To calculate the features, the computationally efficient representation of integral images is applied. The Gentle Ada Boost learning technique is used to learn a selection of Classification and Regression Trees (CARTs) with four splits. Several selections are combined to a cascade of classifiers. To ensure the color-invariance of the input images, they are first preprocessed by applying an edge detection Sobel filter. By passing the filtered images through a threshold, all color information is omitted.

The motivation of our research was triggered by our interest in the Robot World Cup Soccer Games and Conferences (RoboCup) [5], which was created as a standard scenario where technologies could be integrated and developed, thereby encouraging innovative work in the fields of robotics and computer vision and promoting the public understanding of science.

Our experiments were carried out by the autonomous mobile robot Kurt3D, originally constructed to digitalize environments in 3D [13,15,14]. Kurt3D also has other applications, such as educational robotics, or in our case, soccer robotics [5].

The most common techniques for object detection, i.e., ball detection, in the RoboCup context rely on color information. In the last few years, fast color segmentation algorithms have been developed to detect and track objects in this scenario [1,7]. The community agreed that in the near future, visual cues like color will be removed to come to a more realistic setup with robots playing with a ``normal'' soccer ball [16].

Some research groups have already started to develop algorithms for color invariant ball detection. One is described by Coath and Musumeci, who presented an edge-based ball detection system [2]. They developed an adaptive arc identification and location system that processes image data containing edge information.

General Object detection and classification, i.e., not within the RoboCup context, has intensely been researched in computer vision [10,11,17]. Common approaches use neural networks or support vector machines (SVM), for example, to detect and classify objects. Rowley et al. detect faces using a small set of simple features and neural networks [11] and Papageorgiou et al. recognize pedestrians with simple vertical, horizontal and diagonal features and SVMs [10]. Recently, Viola and Jones have proposed a boosted cascade of simple classifiers for fast face detection [17].

Independent from our work Treptow et al. used Viola and Jones' algorithm to track objects without color information. In contrast to their object detection system we preprocess the images to enhance the simple vertical and horizontal features. In addition to this, diagonal features and rotated integral images are used. To recognize different balls we learned Classification and Regression Trees.

The rest of the paper is structured as follows: First we describe the robot platform Kurt3D. Section III presents the learning algorithm. Results are given in section IV and section V concludes the paper.