Flame detection based on Spatio-Temporal Covariance matrix

Automatic fire flame detection is important for intelligent video surveillance. The background model and various color features are usually adopted for flame detection. In this paper, a fire flame detection method is developed by combining both background subtraction and region covariance. The color distribution method and background model with an adaptive background learning model are used to preprocess the image firstly. We then extract the space-temporal covariance matrix, which is used to fuse all the discriminative cues together. Finally we use support vector machine to classify fire scene. The proposed system is effective in detecting uncontrolled fire in complicated environment. Experiments based on several public fire video data sets are utilized to justify the effectiveness of our method.