Smoke detection based on condensed image

Abstract Smoke is difficult to detect whether based on static characteristic such as smoke color, texture, etc., or on dynamic characteristics such as frequency, shape or fluttering, etc. Smoke detection are always subject to some significant exceptions, e.g. swinging bags, driving cars in night, fog, and even moving persons. To efficiently detect smoke we condense video and find that smoke trajectories have some special characteristics, such as right-leaning line, smooth streamline, low-frequency, fixed source and vertical–horizontal ratio. Based on these characteristics it is very accurate and fast with little cost to find smoke even in a complex environment. To demonstrate the effectiveness, the proposed method is conducted on various videos, e. g. videos shot by ourselves, videos from Korea CVPR Lab, videos from YouTube, videos from Bilkent University. Experiments show that similar objects often mistaken by other algorithms are differentiated rightly and the ratio of correct detection is promoted to a higher level of 83.0%.