A Fast Traffic Sign Detection Algorithm Based on Modified YOLOv3. Issue 1 (April 2021)

• Record Type: Journal Article
• Authors: Dou, Huazhou; Zhang, Haiyun; Li, Bo
• Is Part Of: Journal of physics. Volume 1880:Issue 1(2021)
• Publication Date: 2021-04
• Subjects: Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/1880/1/012025
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: Traffic sign detection is an important component of both automatic driving systems and driver assistance systems. However, it is still a big challenge to detect traffic signs quickly and accurately. In this paper, in order to solve this problem, a lightweight algorithm based on YOLOv3 is proposed, which can achieve real-time and accurate traffic sign detection. Firstly, a new lightweight backbone network is proposed by modifying the darknet53 network, which is the original backbone network of YOLOv3. Secondly, in order to further speed up the detection time, depthwise separable convolution is introduced into the proposed backbone network to further reduce the amount of parameters and computations of the model. Thirdly, dense connection is adopted to maintain the accuracy. Experiments on CCTSDB dataset illustrate that the algorithm has great real-time performance and accuracy, which is better than other state-of-the-art algorithms. The algorithm can detect a 416*416 image in only 3.4 milliseconds, which is 44.3% faster than YOLOv3. In addition, the mAP and F1 value are 97.2% and 91.6% respectively, which are better than YOLOv3.