基于通专结合的铁路异物侵限及遗留检测

Detection of Railway Foreign Object Intrusion Limit and Legacy based on Combination of General and Special Technology

摘要:

针对铁路异物侵限检测中传统方法泛化能力差以及基于深度学习的检测模型存在漏检率和误检率较高的问题,提出了一种基于通专结合的铁路异物检测方法；通过解耦处理不同光照条件(白天/夜晚)与摄像头模态(可见光/红外)下的检测任务,结合YOLOv7检测模型与BLIP多模态大模型的语义理解能力,构建了双阈值动态判定策略；采用YOLOv8分割模型精准提取铁轨区域以减少背景干扰；训练适用于不同模态和光照条件的YOLOv7检测模型,并引入低光增强与噪声抑制技术优化夜间检测性能；利用BLIP模型对图像进行语义分析,根据其输出动态调整YOLOv7的检测阈值以平衡漏检率与误检率；经实验测试,在自建铁路异物检测数据集上该方法的mAP达到88.9%,相比基线模型提升0.5%,在真实场景的测试集上误检率和漏检率分别低至1.09%和0.22%；该方法具备良好的实时性与鲁棒性,满足复杂环境下的工程应用需求。

Abstract:

To address the issues of poor generalization ability of traditional methods and high false detection and missed detection rates of deep learning-based detection models in railway foreign object intrusion detection, a railway foreign object detection method based on the combination of generalization and specialization is proposed. By decoupling the detection tasks under different lighting conditions (day/night) and camera modalities (visible light/infrared), and combining the YOLOv7 detection model with the semantic understanding ability of the BLIP multimodal large model, a dual-threshold dynamic determination strategy is constructed. The YOLOv8 segmentation model is used to precisely extract the railway track area to reduce background interference. YOLOv7 detection models suitable for different modalities and lighting conditions are trained, and low-light enhancement and noise suppression techniques are introduced to optimize the detection performance at night. The BLIP model is used to perform semantic analysis on the images, and the detection threshold of YOLOv7 is dynamically adjusted based on its output to balance the missed detection rate and false detection rate. Experimental tests show that the mAP of this method on the self-built railway foreign object detection dataset reaches 88.9%, an improvement of 0.5% compared to the baseline model. On the real scene test set, the false detection rate and missed detection rate are as low as 1.09% and 0.22%, respectively. This method has good real-time performance and robustness, meeting the engineering application requirements in complex environments.