Improved Faster RCNN for Traffic Sign Detection

With the increasing prevalence of autonomous driving, research on traffic sign detection (TSD) draws substantial attention recently. Existing studies usually adopt a twostep framework by first enhancing features of small Region of Interests and then for image analytic. However, the training process of this approach can be in-stable due to the lack of context information, which may restrict the quality of superresolution features. In this paper, we propose an efficient one-step learning-based solution to deal with the TSD problem. We first develop an improved Faster RCNN to detect small objects in traffic images. Then we introduce a new sampling method to optimize the proposed network by selecting highquality proposals. We also present a post-processing scheme to resample the hard false samples having significant contributions to network optimization. Moreover, we adopt Res2net as the backbone of the proposed network in order to obtain more discriminative features. We conduct extensive experiments on the Tsinghua-Tencent 100k dataset, and the results show that our method outperforms other algorithms in terms of accuracy and recall.