Fast Traffic Sign Recognition Based on Convolution Neural Network

Abstract

Due to the uncontrollable conditions and poses of traffic signs, Traffic Sign Recognition (TSR) has been challenging the robustness and efficiency of the existing algorithm. In this work, a method of fast traffic sign recognition based convolutional neural networks (FTSR-CNN) is proposed. A new architecture of convolution neural network is designed, which extracts feature using convolution kernels by sliding filter, and reduces dimension by pooling. In forward propagation, the loss of the whole network is computed. Then stochastic gradient descent method is exploited to minimize loss in step of back propagation. The parameters and activation function of the network are adjusted to optimize the performance. Finally, the full connection is utilized to classify the traffic signs. Affine transformation increases diversity of training samples, which can enhance the adaptability. The proposed method has been evaluated on German Traffic Sign Recognition Benchmark and Tsinghua-Tencent 100K datasets. Experiments demonstrate that it is sufficient to attain high classification accuracy and ensure the efficiency.