Fire detection in video sequences using a machine learning system and a clustered quantitative image marker

In this paper, we proposed a real-time fire detection system used to identify fire between fire and non-fire cases in video sequences. The proposed method consisted of four steps: (1) Dividing video sequences into three-second video clips. (2) Applying a color space model to segment the fire and fire-like regions of each frame. (3) Employing a feature extraction scheme to create a feature pool of spatial and frequency domain, considering motion and time variations. (4) Reducing the size of the feature pool by Kernel-PCA. (5) Training a multi-layer neural network to classify 502 video clips into the fire and non-fire categories. The results showed that the proposed framework can provide a reliable prediction with a low false rate.