【Objective】In order to solve the difficulties such as small target feature extraction, a transformerbased traffic sign detection model was proposed.【Method】Through fully utilizing the advantages of convolution and Transformer, a multi-scale feature extraction backbone model was established with attention fusion, which could enable the backbone network to selectively enhance the features of useful information and suppress the unimportant ones with the support of global context information. In addition, pooling-like connection are incorporated in order to prevent network degradation while enhancing feature fusion. Finally, experiments were conducted on the TT100K dataset.【Result】The experimental results show that the meta-architecture with this model as the backbone achieves the highest mAP of 84%, and the maximum improvement of mAP is about 7% compared with the baseline model.【Conclusion】The model provides a new idea for traffic sign detection while improving feature extraction.