Aiming at the problems that the traditional optical flow method is not suitable for gas and liquidimage detection，this paper proposes a method which uses the optimal mass transmission optical flow as a lowdimensional descriptor of the complex process for fire and smoke detection. The detection process can be abstractedinto a problem about the supervised Bayesian classification of spatio-temporal neighborhood pixels;feature vectorsare composed of the optimal mass transmission optical flow and R，G，B color channels and the single hidden layerneural network classifier are employed. Finally，we determine the pixel belongs to the flame or belongs to the smokeby the analysis the pixel probability. Experiments show that the proposed method successfully distinguishes smokeand the color-similar cloud，also distinguish between the flame and the flame color-similar background，and hasstrong robustness.