A real-time and high-precision method for small traffic-signs recognition
A real-time and high-precision method for small traffic-signs recognition
复制标题
DOI：
10.1007/s00521-021-06526-1
复制
发表时间：
2021-09-25
影响因子：
6
通讯作者：
Zhang, Ronghui
中科院分区：
文献类型：
--
作者：
Chen, Junzhou;Jia, Kunkun;Zhang, Ronghui
交通标志作为交通系统的基本组成部分，通过向驾驶员、行人等提供有关道路状况的基本信息来降低事故风险。随着计算机视觉和人工智能的快速发展，交通标志识别系统已被应用于先进的驾驶辅助系统和自动驾驶系统，以帮助驾驶员和自动驾驶车辆准确地捕获重要的道路信息。然而，在真实的应用中，小型交通标志的识别仍然具有挑战性。在这篇文章中，我们提出了一种有效的方法来识别小尺寸的交通标志，命名为交通标志识别小感知，灵感来自最先进的对象检测框架YOLOv 4和YOLOv 5。总体而言，我们的工作有四个方面的贡献：（1）对于模型的主干部分，我们引入了高层次的特征来构造更好的检测头;（2）对于模型的颈部部分，我们利用感受野块交叉来捕获特征图的上下文信息;（3）对于模型的头部部分，我们细化了检测头网格，以实现更准确的小交通标志检测;（4）在输入方面，提出了一种随机擦除-注意的数据增强方法，增加了困难样本，增强了模型的鲁棒性。在具有挑战性的数据集TT 100 K上的真实的实验表明，与现有方法相比，该方法的性能有了显著的提高，并且是一种实时的方法，在高级驾驶辅助系统和自动驾驶系统中具有巨大的应用潜力。
As a fundamental element of the traffic system, traffic signs reduce the risk of accidents by providing essential information about the road condition to drivers, pedestrians, etc. With the rapid progress of computer vision and artificial intelligence, traffic-signs recognition systems have been applied for the advanced driver assistance system and auto driving system, to help drivers and self-driving vehicles capture the important road information precisely. However, in real applications, small traffic-signs recognition is still challenging. In this article, we propose an efficient method for small-size traffic-signs recognition, named traffic-signs recognition small-aware, with the inspiration of the state-of-the-art object detection framework YOLOv4 and YOLOv5. In general, there are four contributions in our work: (1) for the Backbone of the model, we introduce high-level features to construct a better detector head; (2) for the Neck of the model, receptive field block-cross is utilized for capturing the contextual information of feature map; (3) for the Head of the model, we refine the detector head grid to achieve more accurate detection of small traffic signs; (4) for the input, we propose a data augmentation method named Random Erasing-Attention, which can increase difficult samples and enhance the robustness of the model. Real experiments on the challenging dataset TT100K demonstrate that our method can achieve significant performance improvement compared with the state of the art. Moreover, it is a real-time method and shows huge potential applications in advanced driver assistance system and auto driving system.