融合注意力机制的半监督物体检测在轨道交通上的应用  

Application of semi-supervised object detection technology incorporating attention mechanism in rail transit

作　　者：王俊宇 陈哲 孙俊勇[2] WANG Junyu;CHEN Zhe;SUN Junyong(Locomotive Branch,Baoshen Railway Group Co.,Ltd.,CHN ENERGY,Yulin,Shaanxi 719000,China;State Key Laboratory of Heavy-duty and Express High-power Electric Locomotive,CRRC Zhuzhou Locomotive Co.,Ltd.,Zhuzhou,Hunan 412001,China;State Key Laboratory of Rail Transit Vehicle System,Southwest Jiaotong University,Chengdu,Sichuan 610031,China)

机构地区：[1]国能包神铁路集团有限责任公司机务分公司,陕西榆林719000 [2]中车株洲电力机车有限公司重载快捷大功率电力机车全国重点实验室,湖南株洲412001 [3]西南交通大学轨道交通运载系统全国重点实验室,四川成都610031

出　　处：《机车电传动》2025年第1期35-41,共7页Electric Drive for Locomotives

基　　金：国家能源集团重大项目(168204800040)。

摘　　要：轨道交通的运营和维护需要高精度的物体检测和跟踪技术,以确保乘客安全和系统正常运行。近年来,随着Vision in Transformer的提出,融合注意力机制的大模型在物体检测领域取得了广泛的应用,但是在实现高精度与高鲁棒性的模型训练的过程中,注意力机制对数据的需求量是巨大的。在处理图像数据的过程中,往往会伴随着大量人力、物力的消耗。为降低数据处理成本,文章提出了一种融入了注意力机制的半监督物体检测策略,以提高模型的鲁棒性。研究结果表明,在只处理了10%数据的前提下,采用Grounding DINO和YOLO-World等检测器作为算法的主干,然后在算法的head层采用CBAM、CoTAttention、SEAttention等融合注意力机制,在数据集上达到了0.70±0.04的mAP精度,相比于传统的半监督物体检测,可以得到5.14%的mAP增益,为后续轨道交通物体检测大模型的研究提供参考。The operation and maintenance of rail transit systems require high-accuracy object detection and tracking technology to ensure the safety of passengers and the normal operation of traffic systems.In recent years,with the emergence of Vision in Transformer(VIT),large models incorporating attention mechanisms have garnered widespread application in the field of object detection.However,attention mechanisms require substantial demands of data in the process of model training to achieve high precision and robustness.Moreover,the process of image data processing is often accompanied by significant consumption of manpower and material resources.This paper proposes a semi-supervised object detection strategy that incorporates an attention mechanism,to reduce the cost of data processing and improve the robustness of the models.Detectors such as Grounding DINO and YOLO-World are employed as the backbone of the algorithm,and attention mechanisms such as CBAM,CoTAttention,and SEAttention are applied at the algorithm’s HEAD layer.The study results show that,with only 10%of the data processed,a mean Average Precision(mAP)of 0.70±0.04 is achieved on the dataset,corresponding to a mAP gain of 5.14%compared to traditional semi-supervised object detection techniques.The study findings provide a theoretical reference for future research on object detection large models in the field of rail transit.

关 键 词：半监督 注意力机制 物体检测 轨道交通 

分 类 号：U268.4[机械工程—车辆工程] U298.1[交通运输工程—载运工具运用工程]