改进Faster-RCNN的交通标志检测  

Improved Faste-Rcnn for traffic sign detection

作　　者：黄洛庆 冯寿廷[1] HUANG Luoqing;FENG Shouting(School of Physics and Telecommunication Engineering,South China Normal University,Guangzhou 510006,China)

机构地区：[1]华南师范大学物理与电信工程学院,广州510006

出　　处：《激光杂志》2020年第4期57-60,共4页Laser Journal

基　　金：国家自然科学基金重点项目(No.U1301251);广东省科技计划项目(No.2016A010101021)。

摘　　要：准确且快速地检测道路交通标志是自动驾驶的一个难题,为此提出一种改进Faster-RCNN的交通标志检测算法。首先使用残差网络代替传统VGG16网络提取图像特征,再利用区域建议网络筛选特征图中的目标交通标志并进行初步边框回归。利用ROI池化层将建议区域输出为大小固定的建议框,最后由全连接网络执行目标交通标志的分类与边框的精确回归。算法在德国交通数据集上进行了实验,结果表明,算法能够取得97.7%的平均精度和每张图片0.076 s的检测速度。与同类算法相比,在精度不减的情况下,检测速度更快,具有较大的应用潜力。Accurately and rapidly detecting road traffic signs is a difficult problem for autonomous driving tasks.To this end,a improved Faster-RCNN for traffic signs detection is proposed.Firstly,the residual network is utilized to replace the traditional VGG16 network to extract the feature maps,and then the regional suggestion network is used to filter the target traffic signs in the feature maps and perform preliminary border regression.The ROI pooling layer is applied to output the suggested area as a suggestion box with a fixed size.Finally,the fully connected network executes the accurate classification and the border of the target traffic signs.The algorithm is validated on the German traffic dataset.The experimental results show that the algorithm can achieve an average precision of 97.7%and a detection speed of 0.076 s per picture.The detection speed can faster than the similar algorithm with an acceptable detection accuracy,which has great application potential.

关 键 词：目标检测 交通标志 Faster-RCNN 残差网络 ROI池化 

分 类 号：TN27[电子电信—物理电子学]