摘要
针对机器视觉对驾驶员视角的道路交通标识进行识别时无法同时满足准确率与实时性要求的问题,提出采用深度学习YOLO v2模型的方法。该方法将目标检测看作回归问题,首先将图像划分为网格,然后预测每个网格区域上的边界框和交通标识类别概率,最后通过非极大值抑制获得交通标识的类别和位置,将特征提取与分类集成在同一个神经网络运算过程中以提升实时性和鲁棒性。实验中,YOLO v2模型在交通标识检测数据集上的准确率为80. 1%,检测速度达到40帧/s,相比Faster RCNN,准确率提高5%,检测用时缩短65%;相比SSD算法,准确率提高1%,检测用时缩短20%。结果表明,YOLO v2模型可以满足交通路况的实时性要求。
Focusing on the problem that traffic sign detection based on machine vision from the driver s perspective,can not satisfy both accuracy and real-time performance requirement,a method based on deep learning YOLO v2(You Only Look Once version 2)model was proposed.This method treated the object detection as a regression problem.Firstly,an image was divided into regions.Then,the probablities of the bounding box and the traffic sign category in each region were predicted.Finally,the category and location of the traffic sign were obtained by non-maximum suppression,so that feature extraction and classification were integrated in a same neural network operation process to improve real-time performance and robustness.Experimental results show that YOLO v2 model achieves the precision of 80.1%with the efficiency of 40 frames per second on traffic sign dataset.Compared to the Faster R-CNN algorithm,the precision improves by 5%,and the detection time decreases by 65%;compared to the SSD algorithm,the precision improves by 1%,and the detection time decreases by 20%.YOLO v2 model basically meets the real-time requirements of traffic sign detection.
作者
王超
付子昂
WANG Chao;FU Zi ang(Computer School,Beijing Information Science and Technology University,Beijing 100101,China)
出处
《计算机应用》
CSCD
北大核心
2018年第A02期276-278,共3页
journal of Computer Applications
基金
北京市自然科学基金资助项目(4174091)
北京市自然科学基金重点研究专题项目(Z16002)
北京市教委面上项目(KM201711232013)