Unsupervised Data Augmentation For Improving Traffic Sign Recognition

Traffic sign recognition is a key function in driver assistant systems and autonomous vehicles. Several benchmark datasets had been proposed to test the performance of various recognition models. However, two related problems remained unsolved. First, whether the data samples are enough to evaluate the performance of the proposed recognition models? Second, whether data augmentation could be introduced to build better benchmark datasets? To solve these two problems, we show in this paper that some famous benchmark datasets can be further improved via appropriate data augmentation. Specially, we propose a feature-space data augmentation algorithm that first determines an appropriate feature space for the available data, then generates potentially useful new samples in the feature space and finally maps these new samples into original spaces to get new data samples. Numerical tests show that this algorithm helps to increase the accuracies of recognition models.