A novel method to detect early smoke is presented by processing the video for fire-alarming and gas leak-alarming.The whole detection procedure is mainly composed of four parts:1) off-line learning of the operation parameters of RGB color components;2) RGB color components operation combined with bit-masked color reduction to extract the smoke-like regions in real time;3) the background subtraction by background updating to remove those static disturbing regions,and then,the wavelet-analysis after the RGB color components operation to remove those dynamic disturbing regions;4) the multi-frames evidence analysis of long video sequence to check and label out the real smoke,i.e.all of the accumulated clues are combined to reach a final decision.Experimental results show that the proposed method can detect the early smoke correctly under a dynamic complex scene with the processing speed up to 16～20 fps for color image sequence of 320×240 resolutions,and the time needed for the decision-making is within 3～4 s.The novel method is robust to noises and is of high adaptability to monitor large and open spaces under a complex scene.