A flame detection method based on entropy weighted Support Vector Machine(SVM) is proposed. Firstly, the three-frame difference algorithm is utilized to improve the Visual Background Extractor(VIBE) algorithm, and the Three VIBE(TH-VIBE) foreground detection algorithm is proposed to improve the accuracy and integrity of the acquisition of the suspected flame area. Secondly, entropy weighting is adopted to reduce the redundancy degree of feature data such as texture feature, area change feature, roundness feature and gray level feature, and an entropy-weighted flame recognition model is established to improve the flame recognition rate and accuracy. Finally, based on the flame data from Keimyung University in South Korea and SPG working group of Bilkent University in Turkey, the flame detection accuracy can reach 97% with high robustness.