Abstract: Followed the idea of visual saliency in computer vision and the model of visual attention with top-down, the model of flame detection in video sequences was proposed based on the temporal-spatial visual attention.The Vcomponent of the HSV color space was used to describe the saliency of flame brightness.The relation of R and B in the RGB color space was used to indicate the saliency of flame color.The distance between the feature vectors which are the combination of features with three Local Binary Patterns was applied to express the saliency of flame texture.The dimensions of LBP feature vectors were reduced before by the Principal Component Analysis to lower the computational complexity.Then the saliency of motion was generated by the improved accumulative difference based on the flame color.Finally, the integral saliency map of current frame was formed by weighted linear combination of the static saliency maps and the motion saliency map.Experiments were done on 13fire videos in the fire video dataset from Bilkent University and 2non-fire videos from Internet.The experimental results show that the proposed model achieves better performance on flame detection than other state-of-the-art models.