Research On Video-based Fire Smoke Detecting Algorithms

With the growth of society and economy, there are more and more large space sites, such as supermarkets, underground tunnels, large warehouses,etc. Traditional fire detection techniques give fire alarms only when the combustion particles reach the sensors fixed on the ceiling, whereas in large space sites, the movement of the combustion particles is so slow that it makes a significant delay to trigger the alarm. Therefore, traditional fire detection techniques can not meet the requirement of early fire detection in large space sites. With the application of image processing and pattern recognition techniques in video surveillance domain, video-based fire detection techniques become appealing. Moreover, these techniques rely on existing video surveillance system, which makes it possible to apply and promote these techniques in future.Video-based fire detection techniques include video flame detection, video smoke detection and video flame & smoke detection. In this paper, we choose smoke as the early sign of fire because smoke appears before flame and is difficult to hide. There are two parts included in video-based smoke detection. One is the detection of suspected smoke regions and the other is the extraction of effective smoke features. Proved by many experiments, normal motion detection approaches are not able to segment the suspected smoke regions with continuous edge and complete inner. Here we present a novel method based on optical flow residual increment, which can not only segment the suspected smoke regions completely, but also distinguish smoke from other dynamic texture interferences. Our method outperforms traditional motion detecting approaches. Based on a survey of the dynamic features and static features of smokes, we have also found five most prominent features of smoke, i.e. color feature, chrominance decrease, edge energy decrease, optical flow orientation diffusion trait and shape complexity. We get the significance of each feature through a large amount of tests, and set different weight to each feature accordingly. In the end, the five features are combined through a feature fusion method to give the final decision of the suspected region.Experimental results show that the proposed approach can detect smokes in their early stage and is robust to most kinds of interferences, while there is still more research to be done to improve the generality of our approach.