基于Faster R-CNN的铁路接触网鸟巢检测

Detection of Bird’s Nest in Overhead Caten Ary System Images for Railway Based on Faster R-CNN

为了解决铁路接触上筑巢对铁路安全运行造成的潜在危害,提出一种基于Faster R-CNN的铁路接触网鸟巢检测方法。首先通过改进卷积神经网络VGG16对目标进行特征提取;然后参考RPN网络利用3x3的滑动窗口分别在不同分辨率的卷积特征图上获取目标初始建议区域,最后选择在分辨率较高的Conv4卷积特征图上增加一个反卷积操作对该层特征图的分辨率进行进一步提升,并作为建议窗口的特征映射层传入目标检测子网络中。通过对实际高速铁路行进中拍摄的含有鸟巢的图像进行试验验证。试验结果表明:文中提出的方法在检测精度与速度上,均优于目前主流的Faster R-CNN算法。该方法为实现铁路沿线接触上鸟巢的自动检测提供了可靠依据。

In order to solve the potential hazards of the bird’s nest in OCS to the safe operation of railway,a method of detecting bird’s nest in OCS based on Faster R-CNN is proposed in this paper.Firstly,the object feature is extracted by improved convolution neural network VGG16;then,the region proposals of the object are obtained by using 3×3 sliding window of RPN network on convolutional feature maps of different resolutions,and finally,a deconvolution operation is added to Conv4 convolutional feature map with higher resolution to further enhance the resolution of this layer,and as the feature mapping layer of the region proposals is introduced into the object detection sub-network.The image with bird’s nest taken during the actual high-speed railway is verified by experiments.The experimental results show that the proposed method is superior to the current mainstream Faster R-CNN algorithm in detection precision and speed.This method provides a reliable basis for the automatic detection of bird’s nests in OCS along the railway lines.