Traditional video smoke detection algorithm only uses brightness values as image information when building a multi-dimensional image block, and has higher computation cost owing to the dense sampling.In order to improve the reliability of dynamic texture feature analysis and reduce the cost of computation, this paper puts forward a smoke detection algorithm based on multi-dimensional dynamic texture analysis.In the preprocessing stage, the preliminary separation of smoke foreground is performed by using ICA model, and then the smoke foreground region is found by extracting bottom features with multiple channels and multiple scales through GBVS, thus to improve the segmentation accuracy.In the phase of smoke feature extraction, a smoke feature extraction and detection method based on multi-dimensional feature analysis is proposed.First, the smoke candidate region is obtained through smoke color and background subtraction processing;then the RGB and HOG features are introduced in the four-dimensional image block;finally, the dynamic characteristics of smoke video are analyzed on the basis of the higher order decomposition of multidimensional image data.Owing to the sliding time window, the exact position and the specific time of smoke occurrence can be determined, therefore, the stability criterion of the smoke feature extraction phase and the reliability analysis of dynamic texture features are improved.Experiments show that the recognition rate of the proposed algorithm is higher than that of LDS and h-LDS/GRB.