Verification of Smoke Detection in Video Sequences Based on Spatio-temporal Local Binary Patterns

Abstract

The early smoke detection in outdoor scenes using video sequences is one of the crucial tasks of modern surveillance systems. Real scenes may include objects that are similar to smoke with dynamic behavior due to low resolution cameras, blurring, or weather conditions. Therefore, verification of smoke detection is a necessary stage in such systems. Verification confirms the true smoke regions, when the regions similar to smoke are already detected in a video sequence. The contributions are two-fold. First, many types of Local Binary Patterns (LBPs) in 2D and 3D variants were investigated during experiments according to changing properties of smoke during fire gain. Second, map of brightness differences, edge map, and Laplacian map were studied in Spatio-Temporal LBP (STLBP) specification. The descriptors are based on histograms, and a classification into three classes such as dense smoke, transparent smoke, and non-smoke was implemented using Kullback-Leibler divergence. The recognition results achieved 96–99% and 86–94% of accuracy for dense smoke in dependence of various types of LPBs and shooting artifacts including noise.