Video-based smoke detection using local binary patterns

Smoke detection is particularly important for early warning systems because smoke usually rises before flames arise. Video surveillance systems are widely applied in a variety of fields such as urban scenes and forest scenes. Hence, video-based smoke detection is regarded as an effective and inexpensive way for fire detection in an open space. The existing methods can be classified as histogram-based detection, methods of temporal analysis, smoke detection based on heuristic rules, and hybrid methods. This paper presents an automatic smoke detection method using computer vision and pattern recognition techniques. The method involves texture analysis with rotation and illumination invariant local binary pattern, local ternary pattern, and extended local binary pattern. The novel Local Binary Patterns (LBPs) called as Temporal LBPs were developed. Temporal LBPs are built as 3D structure based on neighbor frames for analysis of dynamic textures. For smoke verification, two different classes of histogram are computed. As a measure of the differences for smoke and non-smoke histograms, Kullback-Leibler Divergence was used. Experiments on the Dyntex database illustrate the effectiveness of the proposed method. Numerical results were obtained by using various types of known LBP for semi-transparent and opaque smoke. The set of all samples was divided in training set (80 %) and testing set (20 %o). Experiments show the advantages of 3D Temporal LBPs against classical 2D LBPs for dynamic fast changed textures. Experimental results show that the proposed method is feasible and effective for videobased smoke classification at interactive frame rates.