【摘要】 针对地铁车底巡检机器人图像采集精度低、缺陷检测准确率不高等问题,提出一种融合视觉伺服控制与深度学习的智能巡检方法.在车底图像采集环节引入基于图像的视觉伺服控制技术,采用ORB特征点检测结合USAC精匹配算法,实现机械臂位姿自适应调整,有效抑制因车辆停靠偏差引起的视角偏离;构建融合YOLOv8模型与PaDiM异常检测算法的深度学习缺陷检测模型,实现车底部件的智能缺陷识别.将该方法应用于车底螺栓缺陷检测,试验结果表明,视觉伺服系统将目标定位偏差控制在±4像素内,显著改善示教式中目标易遮挡问题;融合检测模型能够准确识别螺栓松动、缺失等异常状态.该方法为地铁车底巡检智能化提供了可行的技术方案.更多还原

【Abstract】 To address the challenges of suboptimal image acquisition accuracy and insufficient defect detection precision in metro undercarriage inspection robots, this study proposed an intelligent inspection method integrating visual servo control with deep learning. During the undercarriage image acquisition stage, an image based visual servo control technique has been introduced: ORB feature point detection has been strategically fused with the USAC robust matching algorithm to enable adaptive adjustment of the robotic manipulator,s pose, thereby effectively mitigating the viewpoint deviations induced by vehicle parking misalignment. Furthermore, a deep learning based defect detection model has been developed by integrating the YOLOv8 object detection framework with the PaDiM anomaly detection algorithm, thereby enabling intelligent identification of undercarriage component defects. The proposed method was specifically applied to bolt defect detection on metro train undercarriages. Experimental results demonstrated that the visual servo system confined the target localization error within ± 4 pixels, thereby significantly alleviating the occlusion prone limitations of teach-and-repeat approaches. Meanwhile, the integrated detection model accurately identified anomalous states such as bolt loosening and missing bolts, thereby providing a technically viable solution for intelligent metro undercarriage inspection.更多还原