Segmenting foreground from similarly colored background
Xiang Zhang, Jie Yang, Zhi Liu, Xiangyang Wang · Optical Engineering · 2008
Color similarity between foreground and background causes many foreground segmentation algorithms to fail. In this paper, a new algorithm is presented to segment foreground from similarly colored background. First, model precision and model recall are presented to quantify the model accuracy of various foreground models. Model accuracy tests show that the more accurate the foreground model is, the more accurate the segmentation is. Second, a new foreground model, which is more accurate than the general foreground model, is the developed by blending in different historical segmentations. Finally, the foreground is segmented using the new foreground model combined with a likelihood modification technique. Experimental results on typical sequences show that many foreground pixels misclassified by previous algorithms can be correctly classified by the new algorithm