基于无人机及YOLOX视觉算法的大跨度钢结构吊装过程位移监测

Displacement monitoring in large-span steel structure hoisting process based on drone and YOLOX visual algorithm

在大跨度钢结构吊装施工过程中,节点位移及结构变形关系到吊装施工的安全和质量。对于传统接触式监测方法存在的耗时、耗力且维护费用高等问题,提出了一种以无人机为载体的非接触式监测方式。首先,针对大跨度钢结构吊装过程中无人机近距采集视角受限的问题,采用Harris图像拼接算法进行全景拼接,并与图像加权融合相结合,消除图像拼接中产生的不利光标及拼接缝,实现整体、高精度的大跨度结构图像的无缝拼接;其次,采用加入卷积块注意力机制(convolutional block attention module, CBAM)的YOLOX视觉算法解决复杂背景下不同像素面积的小目标图像识别、坐标提取和位移监测;最后,对四种不同检测模型进行对比评估,并通过对比实验室不同工况试验和实际工程验证该方法在施工环境下对大跨度钢结构测点位移监测的可行性。试验结果表明,加入CBAM注意力机制的YOLOX检测模型的平均精度及置信度均优于其他三种网络模型,且视觉识别的位移信息与Leica全站仪的误差均在亚毫米级内,满足实际工程精度的要求,实现了复杂背景下的小目标位移监测,具备较高的经济效益和广泛的应用前景。

During hoisting construction process of large-span steel structures,node displacement and structural deformation are related to the safety and quality of hoisting construction.Here,aiming at problems of time-consuming,labor-intensive and high maintenance costs associated with traditional contact monitoring methods,a non-contact monitoring method based on drones was proposed.Firstly,aiming at the problem of limited viewing angles for drone close distance collection in hoisting process of large-span steel structures,Harris image stitching algorithm was adopted for panoramic stitching,and combined with image weighted fusion to eliminate adverse cursors and stitching seams generated during image stitching,and realize seamless stitching of overall and high-precision large-span structural images.Secondly,YOLOX visual algorithm incorporating convolutional block attention module(CBAM)dual channel attention mechanism was adopted to solve recognition,coordinate extraction and displacement monitoring of small target images with different pixel areas under complex backgrounds.Finally,contrastive evaluation was performed for 4 different detection models,and the feasibility of the proposed method for monitoring displacements of large-span steel structure measurement points in construction environment was verified by comparing laboratory tests under different working conditions and actual projects.The test results showed that YOLOX detection model adding CBAM attention mechanism has better average accuracy and confidence than other 3 network models,and displacement information of visual recognition and Leica total station have errors within sub-millimeter level to meet requirements of actual project accuracy;this model can realize small target displacement monitoring under complex backgrounds with higher economic benefits and broad application prospects.