基于改进R-FCN的交通标志检测

Traffic Sign Detection Based on Improved R-FCN

为在交通标志检测过程中同时满足精度和速度的需求,建立一种改进的基于区域全卷积网络(R-FCN)的交通标志检测模型。通过K-means聚类算法对数据集进行分析,选择合适的锚点框。对特征提取网络ResNet101进行结构简化,只使用前25层来提取特征,以缩短检测时间。在模型中引入可变形卷积和可变形位置敏感RoI池化层,以提高模型对交通标志的感应能力。模型训练过程中使用在线困难样本挖掘策略从而减少简单样本数量。在交通标志检测数据集GTSDB上的实验结果表明,该模型对交通标志位置信息较敏感,AP50和AP75指标分别达到97.8%和94.7%,检测时间缩至48 ms,检测精度与速度优于Faster R-CNN、R-FCN等模型。

To meet both the precision and speed requirements in the process of traffic sign detection,this paper proposes an improved traffic sign detection model based on improved Region-based Fully Convolutional Network(R-FCN).By analyzing the data set through the K-means clustering algorithm,the appropriate anchor box is selected.Then the structure of the feature extraction network,ResNet101,is simplified,and only the first 25 layers are used for feature extraction to reduce the detection time.The deformable convolution and the deformable position-sensitive RoI pooling layer are introduced into the model to improve the ability of the model to sense traffic signs.In the training process,the online hard example mining strategy is used to reduce the number of simple samples.The experimental results on the GTSDB dataset for traffic sign detection show that the improved model is more sensitive to traffic sign location information.Its AP50 reaches 97.8%,and the AP75 reaches 94.7%.The model also reduces the detection time to 48 ms,displaying a higher accuracy and speed than Faster R-CNN,R-FCN and other models.