Shadow detection based on adaboost classifiers in a co-training framework

The problem of shadow detection is a challenging assignment in video surveillance systems. There are plentiful research achievements about shadow detection but they are not intellective owning to abundant manual input. In this paper, we describe a semi-supervised ensemble technique based on adaboost classifiers in a co-training framework. In this way to detect shadows just demand a fraction of labled datas, and then apply unlabled datas to enhance categorical performance. In the co-training framework, the two detectors are trained synchronously form independent viewpoints. Afterwards the unlabled datas with high confidence which are trained by one classifier are labled and appended to the training pool of the other one. These datas are extracted the information about color, edge, and luminance from RGB color space. Contrary to most of other methods, we increase the illumination assessment to forecast the probability of shadows existence. The experimental results which are operated on the standard roadway and indoor video sequences are ideal and comparable.