交通运输系统工程与信息

交通运输系统工程与信息 ›› 2025, Vol. 25 ›› Issue (4): 162-174.DOI: 10.16097/j.cnki.1009-6744.2025.04.016

• 智能交通系统与信息技术 • 上一篇    下一篇

基于计算机视觉的地铁车站内乘客异常行为检测模型

吴剑凡a ，谢征宇*a,b,c ，秦勇b,d ，王力a ，王佳丽a   

  1. 北京交通大学，a.交通运输学院；b.运营主动安全保障与风险防控铁路行业重点实验室； c. 北京市城市交通信息智能感知与服务工程技术研究中心；d.先进轨道交通自主运行全国重点实验室，北京100044
  • 收稿日期:2025-04-21 修回日期:2025-05-21 接受日期:2025-06-20 出版日期:2025-08-25 发布日期:2025-08-25
  • 作者简介:吴剑凡(1995—)，男，福建福清人，博士生。
  • 基金资助:
    国家重点研发计划项目 (2022YFB4301305)。

Computer Vision-based Model for Detecting Abnormal Passenger Behavior in Metro Stations

WU Jianfana, XIE Zhengyu*a,b,c, QIN Yongb,d, WANG Lia, WANG Jialia   

  1. a. School of Traffic and Transportation; b. Key Laboratory of Railway Industry for Operational Active Safety Assurance and Risk Prevention and Control; c. Beijing Engineering Research Center of Intelligent Perception and Service for Urban Traffic Information; d. State Key Laboratory of Advanced Rail Autonomous Operation, Beijing Jiaotong University, Beijing 100044, China
  • Received:2025-04-21 Revised:2025-05-21 Accepted:2025-06-20 Online:2025-08-25 Published:2025-08-25
  • Supported by:
    National Key Research and Development Program of China (2022YFB4301305)。

PDF

摘要： 为及时有效地应对地铁车站内乘客异常行为事件，本文提出一种基于计算机视觉的两阶段融合模型BiFuseNet(Bi-Fusion Network)，该模型通过融合轻量级检测网络LMD(LCAB, MCAB, DyHead)-YOLO和基于EfficientformerV2的高效分类网络，实现高效且精准的异常行为检测。在模型第1阶段，通过引入轻量卷积聚合块(LCAB)、混合卷积聚合块(MCAB)和动态检测头(DyHead)，有效减少模型的规模，同时，提升对小目标和遮挡目标的检测能力；在第2阶段，采用多层次加权融合策略优化检测和分类结果，进一步增强模型的鲁棒性。实验结果表明，BiFuseNet在自建的MetroAB数据集上取得了89.3%的准确率，较传统模型提高了6.1%，且实现了43.7 frame·s-1的检测速度(FPS)；在PASCAL VOC(Pattern Analysis, Statistical Modelling and Computational Learning Visual Object Classes)和VisDrone(Visual Detection of Drones)公开数据集上，分别提高了10.1%和2.7%的准确率，进一步验证了模型在小目标和遮挡目标检测方面的优势，以及其优异的泛化能力。通过以上设计，BiFuseNet显著提升了地铁车站内乘客异常行为检测的效率和精度。

关键词: 智能交通, 异常行为检测, 两阶段融合, 地铁车站, 计算机视觉

Abstract: To address abnormal passenger behavior incidents in metro stations timely and effectively, this paper proposes a two stage fusion model BiFuseNet based on computer vision. The model integrates the lightweight detection network LMD-YOLO and the efficient classification network based on EfficientformerV2 to achieve efficient and accurate abnormal behavior detection. In the first stage, the model incorporates lightweight convolution aggregation blocks (LCAB), the mixed convolution aggregation blocks (MCAB), and the dynamic detection heads (DyHead), which not only reduce the size of model but also enhance the detection capability for small and occluded objects. In the second stage, a multi-level weighted fusion strategy is employed to optimize detection and classification results, further enhancing the robustness of the model. The experimental results show that BiFuseNet achieves an accuracy of 89.3% on the self-built MetroAB dataset, which is 6.1% higher than that in traditional models, and realizes a detection speed of 43.7 frames per second (FPS). On the PASCAL VOC and VisDrone public datasets, the model improves accuracy by 10.1% and 2.7%, respectively, further verifying the advantages of the model in detecting small and occluded targets, as well as its excellent generalization ability. Through these innovative designs, BiFuseNet enhances the efficiency and accuracy of abnormal passenger behavior detection in metro stations significantly.

Key words: intelligent transportation, abnormal behavior detection, two-stage fusion, metro station, computer vision

中图分类号: