A detection method of traffic signs is proposed based on a modified convolutional neural network. The model is pre-trained to produce the negatives, and hard negative mining is used to add the negative samples into the network to improve the discriminating ability of the model. A feature concatenation strategy during the multi-scale training process is employed to further enhance the performance of the model. On the German traffic sign detection dataset, the effectiveness of the proposed method is simulated in the TensorFlow framework. The research results show that compared with the existing methods, the proposed method can be used to obtain a high detection rate and processing time of only 0.016 s for each image.