Abstract: Aiming at the characteristics of variable signs and difficult to detect, on the basis of SSD (single shot multi-box detector) model, a real-time smoke detection model is proposed based on multi-feature fusion and progressive pooling technology, which can be used for real-time smoke detection and ultimately realize early warning of fire. MobileNet is used as the basic network to extract the smoke image features layer by layer. Then, the feature model is compressed by progressive pooling technology, and the key features are fused forward by deconvolution operation to avoid the loss of key features. Finally, after convolution of 1×1, the smoke image features of different types are fused, and SSD model is used to combine detection boxes with features of different scales, so as to unify the model target box predictor, enhance the ability to judge the positive and negative samples of the model, and realize the accurate judgment of the category and location of the target. Experimental results show that the improved model can not only accurately detect smoke in normal environment, but also achieve good results in smoke image detection with different illumination and scale.