An Improved Feature Extraction and Description Method For Flame Alarm Systems

Abstract

Recently, while the machine learning based approach has been growing rapidly, the automated visual surveillance camera systems has also become one of the sought-after area in computer vision. Fire detection is one of the things which can be applicable to the visual surveillance system. Thus, there have been several researches for vision based fire detection, though, they still have a high false positive ratio as many fire-like colored moving objects exist in the world. For this reason, An Improved Feature Extraction and Description Method For Flame Alarm Systems is proposed. In this algorithm, fire candidates are detected by the defined chroma-intensity map based on the color property of the flame. Next, the Brownian correlation descriptors are extracted from the candidates through multiple adjacent frames in fire and non-fire video. After extracting them, the Brownian motion classifier is trained and tested by the Support Vector Machine to discriminate the dynamic property between the fire and fire-like objects. The proposed method is evaluated by comparing with 2D classical correlation method and one of the recent algorithm using multiple positive and negative videos.