VCF presents an annual award to thebest thesis in the area of MV techonlogy research and application. This year´s winner is Mika Korhonen, who did a great job of developing new ways of improvig visual tracking, as published in his Master of Science Thesis. Congratulations to the winner of this 3000 FIM award and many thanks to all who submitted material for the competition.

Visual tracking is an interesting problem, not least because of the diversity of direct applications related to it --- intelligent surveillance and security systems, virtual reality, user interfaces, video databases and video compression provide a wide domain for practical applications.

Besides offering interesting views in the field of applications the problem of tracking does not lack challenge. The acceptable solutions to the problem range from the simple low-level tracking where objects are identified and their position is tracked over time in a given video sequence to the complex high-level tracking where the goal is to understand the relationship and interaction between objects e.g. human beings and their environment.

One of the first tasks on the way to successful tracking is the choice how to describe the environment i.e. the background scene and the foreground objects we are interested in. A straightforward approach is to use the background image itself without objects as the model. The tracked objects are then assumed to lie where the image of the sequence differs enough from the original background image. The method is called background differencing and despite its simplicity, or maybe just because of that, it performs reasonably well in many cases.

There are, however, a number of situations where the above kind of methods fail. Such cases often occur for example in outdoor tracking where the lighting is due to rapid changes caused by the combination of wind, clouds and the sun or when the camera happens to move as a result of a sudden shake. The alternative is trying to develop a more robust approach to overcome these difficulties. The approach should strive for independency of both illumination and strict position information of the background model.

One way to do that is to represent the background in terms of its color content. Various components of the background can be seen to generate a mixture of color distributions in color space. By taking an adequate amount of samples of the background color it is possible to build a statistical color distribution model of the background. Another issue is the choice of color space i.e. how the colors are represented. By choosing an intensity independent color space where the colors are represented by means of chromaticity such as logarithmic CIELAB, HSI or normalized RG the bad effects due to the changes in illumination are diminished. 

Now having created the background model the candidates for objects of interest are obtained by looking for image pixels which do not fit the model. Those pixels are then connected with neighboring ones of the same kind and labeled to form objects. In order to identify these objects and track them over time in successive frames some appropriate features are needed. In addition to the obvious feature, the position, again, the color chromaticity of the object region is used.

The tracking task then consist of simply segmenting out the regions that differ from the background model and assigning those candidates to the appropriate objects taking advantage of color that remains relatively constant from frame to frame.

So, ignoring the intensity we obtain some robustness to the changes in illumination, and moreover, since the color content representation of the background is not dependent on the strict positions of the image pixels but only color the approach provides us with the advantage that modest camera movement will have no significant effect. 

The price to pay for these desired properties is unfortunately not insignificant. Problems arise when the objects and the background are similar in color --- the objects can easily get hidden.  On the other hand intensity holds a whole lot of the image information and ignoring that sometimes means ignoring the distinctive factor between the objects and the background.

The approach represented here is certainly not a perfect one but rather forms an option to have a look at when developing better techniques for visual tracking. Eventually, there is no single choice that could solve all the problems in tracking, but a general solution has to be sought in the combinations of several various
techniques.