Abstract: By analyzing video data from tunnel test device, an algorithm is proposed to predict tunnel fire. The algorithm has solved the difficulty that, under tunnel environment, the change of light environment would impact the detection of fire recognition system. To analysis the fire dynamic characteristic, it also used a projection scheme to extract rapidly the suspected fire area and the features of flashing flame. Multi-feature fusion technology was employed to reflect the flame occurrence probability from video image. Finally, a flame probability model was established based on the theory of trust degree. The experimental results show that the proposed algorithm has a good adaptability and practicability against the brightness change and vehicle movement. The recognition speed is close to 3~5 s,the recognition rate can reach 97% or more.