基于无人机机器视觉的轨道异物入侵检测方法

Orbital foreign body intrusion detection method based on UAV machine vision

针对传统轨道入侵异物检测方法效率和时效性的不足,文章提出一种基于无人机航拍视频的轨道异物入侵检测算法。首先,通过LSD算法定位铁轨位置并划定铁路安全边界;然后,以YOLOv5s网络为基础,针对其小目标识别率低的问题,改进网络原有的检测层尺度;最后,再引入双向特征金字塔网络,并结合CBAM注意力机制重构模型的Neck部分,进而提出改进后的模型。实验结果表明,改进后模型在VisDrone数据集和自制轨道异物数据集的平均精度值分别提高了9.7%和5.9%,模型体积缩小32%,检测速率达到50.5 frames·s-1,可以为铁路异物入侵智能化检测的研究提供参考。

Aiming at the deficiency of efficiency and timeliness of traditional orbital intrusion detection methods,an algorithm for orbital intrusion detection based on UAV aerial video was proposed.Firstly,LSD algorithm is used to locate the railway track and delimit the railway safety boundary;Then,based on the YOLOv5s network,the original detection layer scale of the network is improved to address the problem of low small target recognition rate;Finally,a bidirectional feature pyramid network is introduced and combined with the CBAM attention mechanism to reconstruct the Neck part of the model,thereby proposing an improved model.The experimental results show that the average accuracy of the improved model in VisDrone data set and self-made track foreign body data set is increased by 9.7%and 5.9%respectively,the model volume is reduced by 32%,and the detection rate reaches 50.5 frames·s-1,which can provide reference for the research of intelligent detection of railway foreign body intrusion.