Smoke detection using simplified descriptors of video information

Abstract

Automatic visual detection of smoke in confined or open spaces is overriding to issue early warnings that can save lives or prevent irreparable damage. While fire presents a range of characteristic colour, smoke does not present a readily apparent pattern. Changes its shape, does not contain clear edges, presents a chaotic behaviour and colour manifests from white to black, including all nuances. This paper presents an algorithm that efficiently pre-process a frame that extracts the main component of information, decreasing orders of magnitude the source size. From this new structure, algorithms based on the temporal and spatial change of subsets of the new structure are applied. Decision is based on fusion of weak classifiers. The algorithms are described and validated with experimental results of real-time detection for open and confined spaces, considering simplicity and efficiency of the proposed method suitable for embedded systems.