Video smoke detection has many advantages such as fast response and non-contact. Due to large variance of smoke shape, color and texture, it's difficult for existing methods to achieve satisfactory results. This paper proposes a robust feature extraction method by using support vector machine (SVM) for classification. First, an edge orientation histogram (EOH) is extracted. Then, circular shift technique is used to transform the maximum value bin of EOH to the fixed bin of EOH, thus eliminating the influence of rotation. To further enhance the robustness of features, Hu invariant moments, mean, deviation, skewness, and kurtosis are extracted from both illuminance and saturation component images converted from original RGB images. Finally, all the features are combined together to form a 38-dimentional feature vector, and SVM is used to train and classify smoke images. Experiments show that the features have good discrimination capabilities, and the method can achieve about 98% and 85% detection rates on selected large training and testing data sets.