基于机器视觉的铁路限界入侵检测方法  

Railway boundary foreign object intrusion detection method based on machine vision

作　　者：杨文 胡昊 李凌志 冯爽 吴浩楠 YANG Wen;HU Hao;LI Lingzhi;FENG Shuang;WU Haonan(Institute of Computing Technologies,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China;China Railway Chengdu Group Corporation Limited,Chengdu 610081,China;Postgraduate Department,China Academy of Railway Sciences,Beijing 100081,China)

机构地区：[1]中国铁道科学研究院集团有限公司电子计算技术研究所,北京100081 [2]中国铁路成都局集团有限公司,四川成都610081 [3]中国铁道科学研究院研究生部,北京100081

出　　处：《铁道科学与工程学报》2025年第3期1328-1343,共16页Journal of Railway Science and Engineering

基　　金：中国国家铁路集团有限公司重大课题(K2023T003);国家重点研发计划项目(2022YFC3005202);中国铁道科学研究院集团有限公司院基金课题(2023YJ088);国家自然科学基金联合基金资助项目(U2268217)。

摘　　要：当异物侵入铁路界限内,会极大地威胁铁路运营安全及乘客生命财产安全,常见入侵目标有闲杂人员与落石、树枝等,但在铁路复杂开放环境中小尺度与小样本入侵目标识别较难。针对以上问题,提出一种基于特征聚焦扩散网络的铁路限界异物入侵检测方法。首先,针对边端计算设备的算力制约问题,对基准模型的主干网络结构进行了轻量化改进,降低了计算消耗,同时维持了相近的检测精度;其次,提出了特征聚焦扩散金字塔网络,改进基准模型的颈部网络结构,强化了不同层级特征间的有效交互,提升了不同尺度目标识别能力;然后,通过动态检测头对基准模型进行改进,改善了在深度网络中目标细粒度特征信息丢失的情况;最后,通过损失函数的改进,让模型更加注重小样本、难识别类型的目标特征信息,并有效提升在此类情况下的识别能力。针对铁路异物入侵真实样本少的问题,模拟采集大量不同场景的异物入侵数据,构建了数据集。实验结果显示,通过增加改进模块,本文所提方法的识别准确率持续上升,最终改进模型的平均准确率达到94.9%,相比基准模型提高了3.7个百分点。对比多种主流目标检测方法,在小目标识别能力提升最为显著,识别率到达了最高的91.3%。研究结果表明本文改进模型在实际复杂铁路环境下能有效识别入侵目标,具有较好的应用价值。When foreign objects invade railway boundaries,they pose a great threat to railway operation safety and passenger life and property safety.Common invasion objects include idle personnel,falling rocks,tree branches,etc.However,it is difficult to identify small-scale and small sample invasion objects in complex open railway environments.A railway boundary foreign object intrusion detection method based on feature focused diffusion network was proposed to address the above issues.Firstly,in response to the computing power constraints of edge computing devices,the backbone network structure of the benchmark model had been improved to reduce computational consumption while maintaining similar detection accuracy.Secondly,a feature focused diffusion pyramid network was proposed to improve the neck network structure of the benchmark model,enhance the effective interaction between features at different levels,and improve the ability to recognize objects at different scales.Then,by using dynamic detection heads to improve the benchmark model,the situation of losing fine-grained feature information of objects in deep networks was improved.Finally,by improving the loss function,the model pays more attention to the object feature information of small samples and difficult to recognize types,and effectively enhances its recognition ability in such situations.In response to the problem of limited real samples of foreign object intrusion in railways,a dataset was constructed by simulating and collecting a large amount of foreign object intrusion data from different scenarios.The experimental results show that by adding improvement modules,the recognition accuracy of the method proposed in this paper continues to improve,and the average accuracy of the improved model reaches 94.9%,which is 3.7 percentage points higher than that of the baseline model.Compared with various mainstream object detection methods,the improvement in small object recognition ability is the most significant,with a recognition rate of the highest 91.3%.The

关 键 词：铁路运输 限界入侵 特征聚焦扩散金字塔网络 动态检测头 损失函数 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]