Algorithm 1. Procedure to detect the smoke object in a video.

Input: A video. Initialize: k = 1, N = 100. While obtaining a frame from the input video do 1. using bilinear interpolation method to scale the current frame image size to 224 × 224, denoted as $Y_k^{}{U}_k^{}{V}_k^{}$. 2. if (k > = N) then, 3. calculating the change-cumulative image of the current frame image according to formulas (1) to (4), denoted as $Y_k^{(\mathrm{a})}{U}_k^{(\mathrm{a})}{V}_k^{(\mathrm{a})}$; 4. calculating the classification score s of the cumulative image after fusion deep network; 5. If (s > 0.5), then 6. outputting the alarm signal of smoke object; 7. End if 8. End if 9. caching images $Y_k^{}{U}_k^{}{V}_k^{}$ and $Y_k^{(\mathrm{a})}{U}_k^{(\mathrm{a})}{V}_k^{(\mathrm{a})}$; 10. k = k + 1; end while