Forest fire detection is one of the most important researches in the forestry applications. It is essential for the safety and protection of forest if the detection of forest fire is timely and accurate. The existing forest fire detection methods using video analysis technology are almost based on the single band video such as the visible light video or infrared video. However, the detection results are not very robust depending on a single band video in the complex forest environment. At the present stage, the study of forest fire detection based on the several bands video is not deep. Using the features of forest fire extracted from the infrared video and visible light video, a new forest fire detection algorithm based on fractional-calculus theory was proposed. The fused video was achieved from the infrared video and the visible light video using the fractional-calculus operator. In the fused video, the background subtraction method was used to determine if there was an abnormal phenomenon occurred. This abnormal case is an indication of the forest fire possibility. Then, a threshold segmentation method was applied into the current frame and the difference image between the current frame and background frame if the fused video was abnormal. Finally, the forest fire region will be detected with these two segmented images. The proposed algorithm was compared with the regional energy fusion algorithm, regional variance fusion algorithm and HSI fusion algorithm using four different measures. Those measures include the spatial frequency, the average gradient, the detection accuracy rate and the detection time rate. Experimental results demonstrate that the proposed algorithm is efficient and accurate and it provides a favorable new approach for the forest fire detection.