Currently, deep learning-based methods for detecting black smoke emitted by vehicles mainly rely on two-dimensional convolutional neural networks to extract features from images. However, since the spatial features of black smoke are extracted from a single image, the detection ability of the model is limited, resulting in a low detection rate of black smoke. In real-world scenarios, road surveillance videos are commonly used as the data source for black smoke detection. To enhance detection accuracy, this paper proposes a Resnet50-based network model that integrates 3D convolution and a self-attention mechanism. The model first introduces the input video sequence and the down-sampled video sequence through the ResNet50 network of 3D convolution. This step is used to obtain short-term motion features and local detail features between adjacent frames of continuous black smoke images, and to fully integrate the space–time information of black smoke. The model employs a self-attention mechanism module to capture long-term information by exploring the global time clue. Subsequently, the feature fusion module is used to fuse the fragment-level features. Experimental results demonstrate that this model significantly enhances the accuracy of vehicle black smoke detection.