基于红外热像的YOLOV8n-TOD列车障碍物检测算法

YOLOV8n-TOD Algorithm for Train Obstacle Detection Based on Infrared Thermal Imaging

摘要: 热成像可在低光环境下检测障碍物。针对列车颠簸影响图像质量的问题，基于ORB特征提取算法与Farneback、Lucas-Kanade光流法加权平均设计一种EIS算法，对采集的数据进行EIS及CLAHE预处理。同时，针对红外图像低分辨率、高噪声敏感性的问题，提出一种列车障碍物检测算法YOLOV8n-TOD，该算法从3个方面进行改进：在YOLOV8n算法中使用MobileNetV3网络替换原主干，通过轻量级结构和深度可分离卷积操作提高算法的计算效率和特征提取能力；在颈部网络中使用FasterBlock网络重构C2f模块，优化特征融合及增强信息传递，提高算法的稳定性与检测精度；优化CIOU损失函数，提高算法的泛化能力。测试结果显示：经预处理后YOLOV8n算法的mAP提高了2.4%；采用YOLOV8n-TOD算法后mAP又提升了7.2%，显著增强了障碍物检测能力。

Abstract: Thermal imaging can detect obstacles in low-light environments. To address the issue of image quality degradation caused by train vibrations, an EIS algorithm has been designed based on ORB feature extraction and weighted averaging using Farneback and Lucas-Kanade optical flow methods. The collected data undergoes EIS and preprocessing with CLAHE. To address the low resolution and high noise sensitivity of infrared images, a train obstacle detection algorithm, YOLOv8n-TOD, is proposed. The algorithm enhances YOLOv8n in three ways: replacing the original backbone with MobileNetV3 for efficient feature extraction using its lightweight structure and depthwise separable convolutions; by using FasterBlock networks to reconstruct the C2f module in the neck network, optimizing feature fusion and enhancing information transfer to improve model stability and detection Accuracy; and by refining the CIOU loss function to boost model generalization capability. Experimental results show that after preprocessing, the mAP of the YOLOV8n algorithm increased by 2.4%; with the YOLOV8n-TOD model, the mAP further improved by 7.2%, significantly enhancing obstacle detection performance.