Evaluation of different techniques for foreground object classification under illumination changes

Abstract

Recently, a lot of attentions have been paid to the object-based analysis in camera surveillance for shortening the gap between high level image semantics and low level image representations. In the experiments, the main focus was on two-class foreground object classification, car and human categories. The experiments had been intensively tested on four different datasets. Various problems existed in those datasets. These problems included the illumination changes, shading, high variations of interested objects’ appearances which are caused by geometrical transformations such as scale, orientation and affine transformations. A comprehensive comparison of selected methods for foreground detection, feature transformation and classification had been conducted to evaluate their effects on the final classification accuracy. The first comparison was done on two foreground detection methods, the traditional Otsu method and the proposed method called Pixel Analysis. The other comparisons were evaluated for four combinations of two feature transformations (histogram transformation vs. binary transformation) and two classification methods (KNN vs. SVM). The results and discussions showed that the combination of binary transformation and SVM was the best for the two-class foreground object classification.