A Novel Probabilistic Appearance Model for Cigarette Detection Under Illumination Change

An effective appearance model is proposed for detecting cigarettes individually in a tightly packed bundle based on an intensity variance probabilistic model. Conventional image intensity threshold based segmentation is brittle and susceptible to environmental variation such as an illumination change. To mitigate this problem, a geometrical snow model is designed to describe the image intensity variances of an individual cigarette. It is then combined with a Gaussian function to generate a probabilistic model of a cigarette detector. The experimental results show that the proposed appearance model provides accurate detection performances and that it is robust to illumination change compared with other conventional methods.