Many algorithms for tracking moving objects are discussed in the literature [2,10]. Especially, for the increasing interest in building autonomous mobile systems, the detection and tracking of moving objects in natural scenes is a very important and challenging task. In the context of mobile systems, such algorithms cannot be limited to static cameras. In addition, real--time constraints must be satisfied. The task must be solved in a closed loop between image acquisition and reaction (for example, camera movement to pursuit the moving object).
In the literature algorithms for tracking of moving objects can be divided into three classes: optical flow , correlation , and feature based approaches [8,18]. Among the feature based approaches model based  and data driven algorithms  can be found. A widely used data driven approach is based on the so called active contour model (snake) . The localization and extraction of an object's contour is done within one energy minimization step. This means, that no motion computation for the whole image is needed, followed by a segmentation step to identify moving objects. Also, the image is only processed locally, namely nearby the contour points. Both facts in combination with the data driven formulation make active contour models well suited for real--time application. This has been proven in the past .
Existing realizations of object tracking systems using active contours are mostly based on extraction of sharp contours in front of a homogeneous background [4,13]. The image gradient is used to identify the contour of an object. In natural scenes with heterogeneous background, problems arise when strong background edges are found near the moving object, in the case of partial occlusions or for weak object contours. Some work exists, which introduces sophisticated energies, which take such cases into account [21,14]. They are well suited for segmentation tasks, for example in medical imaging. Up to now, such extensions have not proven to meet real--time constraints. Thus, for the mentioned area of application, i.e. mobile systems in natural scenes (for example, service robots), an extension to textural energies which can be applied in real--time is an important requirement. Then, the promising results in laboratory scenes can be transferred to real world problems.
This introduction describes a general framework for localization, extraction and tracking of contours in the image plane. The approach is based on the ideas of active contours. The localization and extraction of a contour is formulated as an optimization problem. However, a new radial representation of the contour points is described, instead of a formulation as a parametric function in the 2D image plane. The localization of contour points is then reduced to a 1D search problem on 1D signals, the so called active rays. These 1D signals are the gray values, which are sampled along a straight line from a reference point inside the contour in certain directions (see 2). An internal energy is responsible for the connection of contour points on neighboring active rays. For the localization of contour points on an active ray, a judgement function is introduced. Besides judgement functions, which correspond to the image gradient, we present more specialized functions, which allow for the detection of boundaries between textured objects. The nature of the formalism, i.e. contour point search on 1D signals instead of 2D one, results in an efficient calculation of textural energies and can be applied to real--time applications.
First, we briefly discuss related work. A short presentation of the theory of active contours is followed by a formal introduction of active rays. We define the contour extraction as an energy minimization problem, and show how this approach can be applied to real--time object tracking. Then, we focus on special energies, which are suitable to detect boundaries between textured objects. Experiments show, that this approach is accurate and robust. This introduction concludes with a summary and a discussion of the presented approach.