Publication

Abstract : With the increase in number of fire accidents, the need for the fire detection system is growing every year. Detecting the fire at early stages can prevent both material loss and loss of human lives. Sensor based fire detection systems are commonly used for detecting the fire. But they have drawbacks like time delay and close proximity. Vision based fire detectors are cost efficient and can potentially detect fire in its early stages. Here we propose a lightweight pixel based fire detection model to extract frames from videos and identify frames with fire in it. We use matrix multiplication to transform our input frame to a new color space in which separation of fire pixels from non-fire pixels is easier. The optimal value for the matrix to be multiplied is obtained using fuzzy-c-means clustering and particle swarm optimization. Otsu thresholding is applied on transformed image in new color space to classify the fire pixels in the frame. Our result shows high accuracy on forest fire videos with very less inference time.