关键词: 协方差矩阵, 测地线距离, 火灾探测, 特征选择

Abstract: A flame recognition method based on video image [YCrCb] and [CMYK] space’s color, texture and other spatial-temporal blending feature is proposed, it demarcates the video frames into spatial-temporal cubic block and integrates pixel’s transformed color space’s color component, spatial texture distribution and moving characteristics into integrated covariance matrix descriptor features. It first presents a feature selecting method by evaluating corrective value distribution corresponding each two correlation coefficient of a covariance matrix. In addition, it puts forwards a feature selection and feature assemble methods by analyzing changing trend of Riemannian Manifolds distance of covariance matrix. The flame features and classification methods are analyzed and compared by covariance metrics using Riemannian Manifolds distance, logarithmic Euclidean distance and support vector machine. It is demonstrated by experiments that  the flame detection system based on covariance matrix hierarchical feature has higher accuracy and recognition efficiency.

Key words: covariance matrix, geodesic distance, fire detection, feature selection