Shadow detection for improved object tracking in surveillance cameras

Object tracking algorithms and motion triggered alarms are often dis- turbed by shadows. It is challenging to separate between moving objects and shadows since they have similar movement patterns and image prop- erties. In this thesis, three different approaches to detect shadows are developed and evaluated. The identified shadows determine what parts of the image not to track and what alarms to ignore. The first approach utilizes a mathematical model to estimate the intensity attenuation of a shadowed region. The second approach applies thresh- olding to identify shadows based on information about the attenuation, color change and texture preservation. The third approach makes use of probability distributions describing shadows, background and objects. An energy minimization method using discrete optimization is then used in order to classify the pixels as shadow, object or background. All three approaches were evaluated using several different image sequences with corresponding ground truth. Deriving a shadow detection algorithm that is independent of environ- ment and type of objects in the scene turned out to be the major chal- lenge of this thesis. The best result, a true positive rate of 65.5% and a false positive rate of 2.2%, was achieved with the second approach apply- ing intensity, chromaticity and texture. However, there is still a trade-off between the shadow detection and object discrimination. To further im- prove the performance, more features and a more extensive data set could be useful