Abstract: In the intelligent transportation system, traffic sign recognition is required to have good robustness and real- time performance. In the actual traffic environment may be due to road signs fuzzy, light intensity, scale size, complex background and so on, resulting in traffic recognition accuracy is very low. In this paper, to solve these problems, having improve uses depth learning method to design convolutional neural networks, through the multi-level processing of convolution and pool sampling, combining the RPN network structure of the target detection method, the candidate regions of the image are extracted, so that the candidate regions are extracted, finally using connection network to return to the characteristics of figure, access to the location of the object detection and recognition. The experimental results show that the method can effectively improve the detection accuracy and computational efficiency, reduces the error rate, for traffic sign detection under adverse factors such as illumination, rotation has good stability and accuracy, effectively improves the efficiency of traffic sign recognition, has good generalization ability and adaptability, and meets the requirement of real-time.

Key words: traffic sign detection, real-time, Region Proposal Network（RPN）, intelligent transportation

摘要： 在智能交通系统中要求交通标志识别具有良好的鲁棒性、实时性，并且实际交通环境中可能因路标模糊、光照强弱、尺度大小、复杂背景等因素的问题，导致交通标志识别准确率很低。针对上述问题，提出了利用深度学习方法设计卷积神经网络，并通过卷积和池采样的多层处理，结合目标检测方法中的RPN网络结构，以提取图像的候选区域，从而对候选区域进行特征提取，最后利用全连接网络实现对特征图进行回归处理，获取检测目标的位置及识别。实验结果表明，该方法能有效地提高检测精度和计算效率，降低错误率，对于光照、旋转等不良因素下交通标志检测具有较好的稳定性和准确性，有效地提高了交通标志识别效率，具有良好的泛化能力和适应性，且满足一定的实时性的要求。

关键词: 交通标志检测, 实时, 区域生成网络（RPN）, 智能交通