In order to solve the problems that traditional research methods often have limitations in dealing with high-dimensional and complex data features in complex environments, and it is difficult to achieve high-precision and robust prediction, a traffic accidents severity intelligent prediction method based on convolutional neural network and attention mechanism is proposed. A multi-scale feature extraction model with attention fusion, which is called channel and multi-head attention network (CMHANet), is constructed to make full use of the advantages of convolution and attention mechanism. The convolution layer is used to effectively extract spatial features in the data, while the channel attention mechanism can weight and enhance important features, suppress unimportant features, and emphasize the focus on key data points. In addition, in order to capture the complex dependencies between different features, a multi-head attention mechanism is also introduced. Finally, experiments are conducted on the US-Accidents dataset. The experimental results show that the prediction framework with this model as the backbone achieves improvements in F1-score, precision, recall and accuracy. While improving the effect of feature extraction and association modeling for high-dimensional and complex data, the proposed model also provides a new idea for intelligent prediction of traffic accidents.