自动驾驶场景下小且密集的交通标志检测

摘要

在自动驾驶场景中,交通标志的检测和识别对行车周围环境的理解至关重要.行车过程中拍摄的图片中存在许多较小的交通标志,它们很难被现有的物体检测方法检测到.为了能够精确地检测到这部分小的交通标志,我们提出了用浅层VGG16网络作为物体检测框架R-FCN的主体网络,并改进VGG16网络,主要有两个改进点:1)减小特征图缩放倍数,去掉VGG16网络卷积conv4_3后面的特征图,使用RPN网络在浅层卷积conv4_3上提取候选框;2)特征拼层,将尺度相同的卷积conv4_1、conv4_2、conv4_3层的特征拼接起来形成组合特征(aggregated feature).改进后的物体检测框架能够检测到更多的小物体,在驭势科技提供的交通标志数据集上取得了很好的性能,检测的准确率mAP达到了65％.%In self-driving scenarios,the detection and recognition of traffic signs is critical to understanding the driving environment.The plethora of small traffic signs are hard to detect by the existing object detection technology.To detect these small traffic signs accurately,we propose the use of the shallow network VGG 16 as the R-FCN's backbone and the modification of the VGG 16 network.There are mainly two improvements in the VGG 16 network.First,we reduce the multiple zooming of feature maps,remove the feature maps behind the VGG16 network convolution conv4_3,and use the RPN network to extract the region proposal in the shallow convolution conv4_3 layer.We then concatenate the feature maps.The features of the layers of the convolutions conv4_1,conv4_2,and conv4_3 are adjoined to form an aggregated feature.The improved object detection framework can detect more small objects.We use a dataset of traffic signs to test the performance and mAP accuracy.