Abstract :

Currently, fire detection systems based on computer vision techniques are highly appreciated for their intelligent detections at the earliest. These systems use surveillance cameras to capture high-level information from a fire that enables a system to take preventive and corrective measures before the occurrence of a fire hazard. Handling false fire detection and reducing false alarms in such developed systems are still big challenges that need to be addressed. In this research, a novel framework is proposed that uses angular and altitude information of the fire flame to measure the severity of the fire and produce an alarm based on the level of severity of the fire. The proposed system especially handles false alarms.
In this thesis, the proposed framework Unsupervised Fire Detection (USFD) records the scene through a surveillance camera, subtracts the background and identifies the fire flame as foreground region. In general, the nature of fire flame is dynamic, its shape, size, volume etc. is irregular and continually change due to external factors. The spreading of fire pixels in all directions are analysed and the dynamic features for the fire flame include: centroid (for initial foreground region), boundaries of the flame, altitudes between the centre and boundaries and the angular regions between the altitudes are computed. The whole region of the fire flame •foreground region) is divided into 12 angular regions, each of 30 degrees. The size of the altitude and the angular region is used to compute the severity of the fire flame. The predicted values of the fire and the next flame location are estimated by a combination of a Kalman filter, centroid algorithm, distance, and altitudes functions. The altitudes information of different thresholds is used to make a positive fire alarm decision.
The results achieved on different datasets of fire videos show that extracted features using the proposed framework efficiently distinguish a normal and hazardous fires. These features are also useful to estimate the size and direction of a fire flame. The altitudes and angular regions information can be fed into the support vector machine to learn a model for a fire flame that can be used to predict fire flame and its severity for a real-time fire hazard.
The experimental results proved that the proposed framework can be used to detect and measure the severity of fire for both indoor and outdoor environments. The proposed framework can be practically implemented with real-time camera monitoring systems in various environments, such as homes, forests, and markets. This USFD framework has an accuracy of 95.89% which is much better than the existing frameworks.