Detection and Tracking of Moving

This paper presents a method for detection and tracking of moving cast shadows on a dominating scene background in a monocular video sequence. The method assumes moving shadows on a dominant smooth shaped background. The shadow causing light sources are assumed to be strong enough to cause visible temporal frame diierences by moving cast shadows. These diierences are detected and classiied into regions covered and regions uncovered by moving shadows. Entire moving cast shadows are detected and tracked by temporal integration of the covered background regions while subtracting the uncovered background regions. Results obtained with various test sequences show that those shadows are detected and tracked successfully that are also easily visible to a human observer. A moving cast shadow in a video sequence is generated by an object occluding a light source with respect to another object. It's generation depends on the scene geometry. This includes position and shape of the objects but also position and kind of light sources. The light sources, the objects and the generated cast shadows in a scene are linked by light rays. If knowledge on the objects is available, these light rays can be used to detect cast shadows in video images 7]]1]. If cast shadows can be detected by other methods, the light rays can be used to determine the 3D object shape 3]. Furthermore, cast shadows play a role in object segmentation. For video editing, surveillance, remote data bases and other typical MPEG-4 applications 2], moving objects need to be segmented. As cast shadows move together with the objects, they are likely to be segmented erroneously as objects. Without knowing any 3D geometry, image regions changed by moving cast shadows can be iden-tiied using assumptions on the scene and on the illumination 12]]11]]10]]6]]15]. In 13], all assumptions have been recently combined to detect image regions changed by moving cast shadows. By a combination with the Analysis Model of the Euro-pean project COST 211 4]]9]]5], it has been shown that hereby shadow artifacts of object segmenta-tion can be reduced drastically. In this paper, the detection and tracking of entire cast shadows is addressed. Therefore, a nonlinear ltering approach will be proposed that processes the image regions that are changed by moving cast shadows. First, these image regions will be detected as in 13]. Then, they will be classiied into two classes. Regions of the rst class are changed, because they have been …