For over fifty years, simple smoke and heat sensors have been the primary means of automated fire detection. We are now at the point where computer processing power is cheap enough and machine vision technology is sophisticated enough for a new generation of automated fire detection systems: video-based fire detection (VBFD). While current smoke and fire detection technology has proven to be reliable and effective, VBFD technology promises to go where existing systems can’t and to detect fires faster than its venerable predecessors ever could.

This thesis explores a few methods for achieving real-time video-based fire detection in low quality data. Assuming a stationary source camera, we describe an algorithm that uses a support vector machine to classify short, targeted video sequences as fire/non-fire. The algorithm achieves a classification rate of 96.0% on a holdout set of real world data. Furthermore, the system is robust with respect to the distance from the fire source, works day or night, and only requires the processing power of a common desktop computer.