基于改进Faster R-CNN的铁路信号灯与停留车检测方法

在铁路编组站调车作业中，因调车头机车结构特点，司机难以时刻观察地面信号，常因主观误判导致调车头误闯信号灯挤坏道岔和冲撞停留车的事故发生。文章针对铁路信号灯与停留车目标大小悬殊的多尺度目标检测问题，改进Faster R-CNN目标检测算法，采用深浅层特征融合方法和多尺度训练策略，较好地兼顾了对二者的高质量检测。此外，文章采集车载铁路视频图像，标注信号灯和停留车，构建了较大型的目标检测数据集。实验结果表明，改进的Faster R-CNN在所构建数据集中，信号灯检测精确率达到96.6%，停留车检测精确率达到98.9%，检测速度约10帧/秒，能够满足铁路编组站低速调车作业应用场景的实时性要求。

Detection Method of Railway Signal Lights and Parking Carriages Based onImproved Faster R-CNN

In railway shunting operation, the locomotives break the railroad switch and collide with parking carriage. Most of these accidents are caused by the misjudgment of drivers who cannot always observe the traffic lights on ground and parking carriages, owing to the special structure of the shunting locomotive. Aiming to settle the multi-scale object detection problem of railway signals and parking carriages, this paper improves the Faster R-CNN by using a fusion strategy of deep and shallow features and muti-scale training skills. Moreover, a large-scale dataset with annotated railway signals and parking carriages is constructed based on locomotive on-board videos. The experimental results demonstrate that on the built dataset the improved Faster R-CNN detector can achieve the detection accuracies of 96.6% and 98.9% on railway signal lights and parking carriages, respectively, and the detection efficiency reaches about 10 frames per second, which can meet the real-time requirements of low-speed shunting operation application scenarios in railway marshalling stations.