铁道标准设计2025，Vol.69Issue(6)：196-204,226,10.DOI:10.13238/j.issn.1004-2954.202308030003
基于深度学习的城际高铁轨道交通站台异常检测研究
Research on Anomaly Detection of Intercity High-speed Railway Platform Based on Deep Learning
摘要
Abstract
The risk space between platform screen doors and trains in intercity railway systems is characterized by extreme length,large size,numerous visual blind spots,and highly complex environments.These features make it more susceptible to safety incidents that affect train operations,such as left-behind objects,abnormal passenger behaviors,and boundary violations.Existing anomaly detection methods suffer from large blind areas,high false detection rates,and an inability to effectively identify anomalies,making them unsuitable for detection tasks in intercity railway platforms where the gap between the platform edge and screen doors exceeds 1.2 meters or involves even larger risk zones.To address this issue,based on an analysis of the characteristics of the intercity railway risk space,the study explored the advantages and potential of top-mounted visual sensors over conventional detection methods in anomaly detection tasks.It examined the specific requirements of anomaly detection in this risk space and analyzed the high adaptability of deep learning methods to such tasks.Finally,several deep learning-based algorithms were introduced for detecting abnormal objects and identifying anomalous passenger behaviors within the intercity railway risk space.These were compared against existing technologies in terms of performance and applicability.For the foreign object detection task,the study proposed an image inpainting-based anomaly detection network specifically designed for intercity railway platforms.By leveraging an image inpainting autoencoder,a global reconstruction error,and a local anomaly enhancement module,the network effectively highlighted discrepancies between reconstructed normal images and input images containing anomalies,achieving a detection accuracy of 99.3%.For abnormal behavior detection,the study proposed a skeleton-based recognition framework.This approach utilized a pose estimation network to extract individual skeletal data and employed a graph convolutional neural network(GCN)to classify skeletal sequences.The framework achieved an average recognition accuracy of 91.7%across four types of behaviors:falling,squatting,bending,and walking.关键词
高速铁路/城际轨道交通/深度学习/异物检测/异常行为检测Key words
high-speed railway/intercity railway/deep learning/foreign object detection/abnormal behavior detection分类
交通运输引用本文复制引用
毛良,邱启盛,刘瑞康,段梦飞,董佳勋,刘伟铭..基于深度学习的城际高铁轨道交通站台异常检测研究[J].铁道标准设计,2025,69(6):196-204,226,10.基金项目
国家重点研发计划项目(2016YFB1200402) （2016YFB1200402）
