Improving Temporal–Spatial Features Extraction of Forest Flame Video

An effective and prompt features extraction method plays a significant role in a forest fire detection system. However, exhaustively detecting all images in a long monitoring video is a serious time-consuming and simply extracting moving feature between two adjacent frames will greatly degrade the recognition accuracy. In this paper, a novel flame feature extraction method based on temporal-spatial domain is presented. A suspected flame video detected by color feature and moving feature is divided into a series of video clips based on the sliding time window. Color covariance descriptors and intensity descriptors are defined for representing color feature and texture feature of forest flame video. The width variation descriptor and the height variation descriptor are also defined for distinguishing flame videos from interfering videos such as red flowers video and red leaves video. Finally, the feature vector of a suspected flame video is extracted by computing co-covariance matrix composed of the above descriptors. The experimental results on flame detection demonstrate the effectiveness and usefulness of the proposed scheme.