Fine-grained Traffic Sign Detection and Matching Algorithm

Abstract

With the development of intelligent driving, the autonomous recognition of traffic signals plays an important role in providing traffic information for the autonomous driving systems. In this paper, we propose a mechanism to detect the traffic signs from an image of the traffic scene and to match the same sign from the pictures shot in different time and weather conditions. The dataset we use, provided by Baidu, contains 19 fine-grained objects belonging to 3 coarse categories, most of these objects are very small and different categories of traffic signs have various frequencies of appearance. To accurately accomplish the fine-grained detection task with such an imbalance dataset, we divide the process of traffic signs detection into two parts: detection and fine classification, and use different methods of data augmentation on different categories of data to alleviate the imbalance issue. Except for the sign detection, we also applied a matching mechanism to match the same target signs under different road conditions with metric learning algorithms. As a result, our model achieves results comparable to the top results of related contest.