The invention discloses an indoor scene flame detection method and belongs to the application fields of a visible light image, video processing and the like. In the prior art, whether an initial stage flame and a small flame which may cause a fire hazard exist is difficult to accurately detect and track. By using the method of the invention, the above problem is solved. In the invention, color space conversion and segmentation of the visible light image are taken as a basis, and a morphological characteristic of the flame is combined so as to carry out distinguishing. Because a component channel of the flame image in a YCbCr color space possesses obvious gradient intensity and gradient distribution characteristics, a support vector machine is used to learn and identify an HOG characteristic, and in a dynamic video, through a transient frame difference and a center of mass drift algorithm, tracking of a flame target is realized. A lot of indoor flame image and video sample tests show that the method has advantages that an identification rate is high; a detection speed is fast; a false alarm rate is low and reliability is high compared to a traditional method. The method aims at detection and tracking of the initial stage flame, the small flame and the like.