Abstract:

Aiming at the problems of high maintenance cost of equipment and long application landing period, and according to the advantages of Building Information Modeling（BIM）, which is consistent with the real scene space, flexible and adjustable scene perspective, and can simulate various lighting conditions, a method of equipment switch state detection is proposed based on LSD（Line Segment Detector） and deep learning. By detecting the line segment information of the image and selecting the line segment based on the edge features of the switch box, the position of the switch box in the image is framed to generate the image data set, which is then input to YOLOv3（You Only Look Once version3） network training to generate deep learning model. Finally, the deep learning network framework is deployed to the edge devices, and the operation state of the switch box in the real environment is detected on the edge side. Experimental results show that the proposed method can recognize the working state of the switch box in BIM environment in a short time, which has good adaptability to the real environment.

Key words: image processing, switch status, target detection, line detection

摘要：

针对工业复杂环境下设备维保成本高、视觉检测落地周期长等问题，并根据建筑信息模型（Building Information Modeling，BIM）具有与现实场景空间一致，场景视角灵活可调以及可以模拟各类光照条件等优点，提出一种在BIM环境下融合LSD（Line Segment Detector）直线检测与深度学习的设备开关状态检测方法。通过检测图像直线段信息，并基于开关盒边沿特征对直线段进行筛选，实现在图像中框定开关盒位置生成图像数据集，进而输入到YOLOv3（You Only Look Once version3）网络训练生成深度学习模型。将深度学习网络框架部署到边缘设备，在边缘侧对真实环境下开关盒工作状态进行检测。实验结果表明，该方法能够在短时间内实现BIM环境下识别检测机柜设备上的开关盒工作状态，并对真实环境下开关盒工作状态检测具有良好适应性。

关键词: 图像处理, 开关状态, 目标检测, 直线检测