Abstract: The traditional neural network only uses the end-layer feature and needs massive and time-consuming computation in the traffic sign recognition, thereby resulting in an inaccurate and non-real-time classification. To solve the problem, a traffic sign recognition (TSR) method based on multi-layer feature expression and extreme learning machine (ELM) is proposed. Firstly, the multi-layer features of traffic signs are extracted using the convolutional neural network (CNN). Then, the multi-scale pooling operation is used to combine the extracted feature vectors of each layer to form a multi-scale multi-attribute traffic sign feature vector. Finally, the extreme learning machine (ELM) classifier is used to realize the classification of traffic signs. Experimental results show that the proposed method can effectively improve the accuracy and it has strong generalization ability and real-time performance in TSR.