Pushing the "speed limit": High-accuracy US traffic sign recognition with convolutional neural networks

This paper presents a novel convolutional neural network (CNN)-based traffic sign recognition system and investigates pre- and post-processing methods for enhancing performance. We focus on speed limit signs, the most difficult superclass in the US traffic sign set. The Cuda-convnet is chosen as a suitable model for the traffic sign recognition task with low-resolution training images and limited dataset size. We test on the world's largest public dataset of US traffic signs, the LISA-TS extension, and testing dataset. Compared with the current state-of-the-art aggregated channel features detector that has achieved near-perfect detection accuracy except for US speed limit signs, our approach improves the area under precision-recall curve (AUC) of the speed limit sign detection by more than 5%. We also discuss potential improvements of the CNN-based traffic sign recognition method.