Spatial-temporal features for smoke detections on video images

Three temporal-space features are proposed in this paper on video images for smoke detections on the issue of dynamic textures, motion patterns and contour variations. Considering that the previous methods for smoke detection provides global descriptions on dynamic features, but its illustrations do not well indicate the temporal correlation among frames, the novel spatial temporal descriptors among neighboring frames are proposed. Our method combines local and global descriptions in the temporal-space domains, where concatenated histograms of LBP (local binary pattern) are utilized for dynamic texture descriptors, the 2D mutual information for motion disorders and the information entropy for contour variations. The experimental results show that the proposed spatial-temporal features could effectively discriminate smoke and the non-smoke objects.