基于Libra R-CNN方法的高铁接触网腕臂管帽检测

Cap Detection of the Cantilever Bracket in the High-speed Train Overhead Contact Line Based on Libra R-CNN

由于列车长期高速运行,接触网腕臂会发生管帽脱落的现象。没有管帽的保护,腕臂容易进入杂物,从而引起腕臂的锈蚀、结构失稳等问题,影响铁路的安全运行。目前,铁路局主要采用图像监测结合人工查看的方法检查管帽脱落与否。该方法不仅效率较低,而且工作量巨大。针对管帽对象较小、管帽样本不均衡和管帽数据集较小等问题,依据计算机视觉理论,使用基于目标检测模型的方法,并结合迁移学习技术,进行管帽脱落的识别,可以显著提高管帽脱落的检测效率。通过试验调整R-CNN架构的结构,选择合适的预训练模型,改变主干网络的训练模式,选择最优的检测模型。最终,结合COCO预训练模型的Libra R-CNN方法在测试集上取得最好的效果,mAP@0.7达到了98.2%,具有实际工程意义和应用前景。

Cantilever bracket’s caps would lose because of the long-term and high-speed movement of the railway.The cantilever bracket was easy to enter the sundries without the protection of caps,it caused the corrosion,structural instability and other problems of the cantilever brackets,which affected the safety of the high-speed railway.At present,the railway bureau mainly used the method of image monitoring combined with manual inspection to check whether caps lose or not.This method was not only inefficient,but also with huge workload.Aimed at the problem of the cap being small,cap samples being unbalanced,the cap data set being small,and according to the theory of computer vision,we used the method based on the object detection and combining with the transfer learning,which could significantly improve the efficiency of cap detection.Through experiments adjusting the structure of R-CNN,we selected the appropriate pre-training model and changed the training mode of the backbone network,which selected the optimal detection model.Finally,the Libra R-CNN with the COCO pre-training model was the best in the test set,the mAP@0.7 was 98.2%.The method had practical engineering significance and application prospect.