Abstract: The contact net is a special form of transmission line that is erected along the railway line and supplies power to electric locomotive. The foreign objects such as the bird’s nest attached to the net cause safety hazards to train operation. At present, the contact net foreign matter is detected and removed by manual inspection. This method is not only costly and inefficient, but also cannot eliminate safety hazards in time. In order to timely and effectively detect foreign matter in the contact net and reduce labor cost, a real-time detection algorithm of railway intrusion foreign objects is proposed by using computer vision and deep learning techniques for the fixed structural features of high-speed rail operating environment. Firstly, based on the LSD straight line segment detection algorithm, the Region of Interest（ROI） that may appear in the nest is obtained. Secondly, the YOLOv3 network is used on ImageNet to obtain a pre-training weight, and the manually labeled data set is adopted to continue training the network until the network converges. Finally, the trained nest is applied to detect the nests in the ROI. The experimental results show that the average detection accuracy is 0.89 and the average detection speed is 38 f/s, which can achieve accurate real-time detection of foreign objects.

Key words: foreign object intrusion detection, high-speed rail contact net, line detection, deep learning, YOLOv3

摘要： 接触网是沿铁路线上空架设的为电力机车供电的输电线路，接触网上附着的鸟巢等异物将对列车运行造成安全隐患。目前主要是通过人工检查的方式对接触网异物进行检测并清除，这种方式不仅成本高，效率低，往往不能及时排除安全隐患。为了对接触网异物进行及时有效的检测，同时降低人力成本，针对高铁运行环境的固定结构化特征，综合运用计算机视觉、深度学习等技术对铁路入侵异物进行实时检测。首先基于LSD直线段检测算法获取鸟巢可能出现的感兴趣区域；其次利用YOLOv3网络在ImageNet上进行训练得到一个预训练权重，并使用人工标注的数据集继续训练网络直到网络收敛；最后使用训练好的网络对感兴趣区域存在的鸟巢进行检测。实验结果表明，最终得到的平均检测精度为0.89，平均检测速度为38?f/s，可以实现对异物目标的准确实时检测。

关键词: 异物入侵检测, 高铁接触网, 直线检测, 深度学习, YOLOv3