In the smoke foreground separation process, in view of a lot of noise still produced in the traditional Gaussian mixture model to separate the foreground this paper applied independent component analysis (ICA) to separate the smoke foreground and improved traditional smoke separation algorithm. The algorithm with the ICA to eliminate the second order and higher order related between foreground and background of smoke, and reduce the interference of non-smoke composition. The smoke foreground area above was reduced through the graph based visual saliency (GBVS) to be a pure smoke area. The experimental results show that compared with cigarette-smoke detection based on Gaussian mixture model and frame difference, the smoke area extracted in this paper is smaller and more concentrated. Subjective visual evaluation and objective indicators both shows better recognition effect of the algorithm in this paper.