Abs t ract Due to industrial deployment and expansion of urban zones, early warning systems play critical role in giving emergency response to unexpected events. Fire detection systems are very important in comparison with other warning systems due to fire huge damage and its destructive power. Nowadays many companies try to present efficient methods for early fire detection which prevents further disasters. Usually manpower is used for monitoring in traditional methods of fire detection; but they are costly and do not have suitable efficiency in some applications. Fire sensors are modern methods which are used in fire and smoke detection. These methods cannot be used in outdoor applications, since an enormous number of sensors are needed in order to have an adequate density and completely cover the area. Hence, an important disadvantage of point detectors is that they are distance-limited and produce high rate of false alarms in outdoor environments. Due to rapid developments in digital camera technology and improvement in computing power and memory, there is a major trend to replace conventional fire detection methods with vision-based systems. These systems can solve the shortcomings of previous methods. Modern methods try to automatically distinguish fire in their coverage area. These methods are economical and reduce human errors. Camera is an important part in such systems which covers vast areas and collects suitable information about size, expansion velocity and fire locality. In order to distinguish fire using a camera, we need to process video sequences to find out the presence or absence of fire in them. The main goal in fire detection systems is designing the systems with high detection rate, low error rate and reasonable detection time. One of the main requirements of vision-based fire detection systems is features extraction with high separation capability. Therefore, different methods are presented for extracting vision-based information and fire region detection. The goal of this thesis is to introduce an efficient method for fire detection in outdoor application using video sequences. We propose two models. In the first one, we use motion and color features due to dynamism and color range of fire. A more general method is proposed in second model which can be used in both fixed and moving camera scenarios. We use saliency map as a new feature for fire detection in this later method which distinguishes important regions in an image. Texture descriptor in time domain for fire texture extraction provides another feature for this method which improves accuracy. Keywords: Video fire detection, vision-base information, color space, motion detection, saliency detection, 3D texture descriptor.
