Smoke is an important feature of early fire detection. The extraction of smoke features by traditional machine learning and two-dimensional convolutional neural network smoke detection algorithms are limited to the spatial domain, and cannot extract temporal information. The existing three-dimensional convolutional neural network detection algorithm has the problems of high calculation cost and low detection time efficiency, which leads to unsatisfactory detection accuracy and false alarm rate. To solve the above problems, a smoke video detection method based on deep learning in spatio-temporal domain is proposed. The block moving target detection method is used to extract the moving targets of the smoke video and filter the non-smoke targets. At the same time, the three-dimensional convolutional neural network is split to form a two-plus-one-dimensional spatio-temporal network module, which extracts the characteristics of the spatio-temporal domain and improves the detection time efficiency. In order to suppress irrelevant features, an attention mechanism is introduced to increase the compression and incentive network to recalibrate the weight of feature channels to improve the accuracy of smoke detection. The research results show that the average accuracy rate of the algorithm used in this paper is 97.12%, the average correct rate is 97.06%, the average false alarm rate is 2.74%, and the average detection frame rate is 10.49 frame/s. The needs of fire smoke detection is met, and the detection timeliness is improved significantly.