<p>Distracted driving behavior has great influence on the driving safety of electric buses. To address the problem of low accuracy of the driving behavior recognition model caused by ignoring the shallow features and only adopting the deep features extracted from the final convolution layer of convolutional neural network (CNN), a novel distracted driving behavior recognition method based on improved CNN with cross-layer multi-scale features fusion approach and multi-head attention mechanism (CMCNN-AM) is proposed. First, using the cross-layer multi-scale features fusion approach to fuse the shallow features and the deep features extracted by convolutional layers. Second, the multi-head attention mechanism is applied to give more attention to the key features of the distracted driving behavior. After that, the model is trained and verified with the distracted driving behavior dataset of electric bus. Finally, the mapping relationship between the features and the categories of distracted driving behavior is established and the CMCNN-AM model is constructed. Besides, the generalizability of the proposed model is validated based on the State Farm Distracted Driver Detection dataset. The experimental results show that the accuracy of the model reaches 95.56%, which outperforms VGG-16, ResNet-50 and CNN-AM, and the model has good generalization. The proposed method is significant to prevent the occurrence of distracted driving safety accidents in electric buses and improve the management level of transport safety.</p>