This study proposes a video smoke detection method combining the Gaussian mixture model (GMM) with a convolutional neural network (CNN) to ensure real-time and accurate video smoke detection in complex scenarios. First, background subtraction based on GMM and morphological methods are used to extract motion objects from video images. Second, a CNN model for video smoke detection is designed, taking into account the limitations of smoke detection efficiency and overfitting of the network model. Finally, the designed CNN model is trained and tested by using positive and negative smoke sample images. The output probability threshold of the network model of motion objects is set reasonably, which can effectively remove the non-smoke interference items that are not covered in the training samples. The false alarm rate can thereby be reduced. Experimental results prove the validity and feasibility of the method. The accuracy of video smoke detection reaches 97.5%, and the average response time of the smoke alarm is 4.58 s, satisfying the real-time demand of smoke detection in complex scenarios.