基于计算机视觉的斗栱转动变形检测方法

Rotation deformation detection method for Dou-Gong based on computer vision

古建筑木结构斗栱节点由斗、栱等构件交错层叠而成.受自身形制、环境和外力等影响,这种“层叠式”节点易出现转动变形.针对人工及布设传感器的传统检测方法在可实施性及检测效率方面的局限性,提出了一种基于姿态估计的斗栱转动变形计算机视觉检测方法.首先定义了斗栱的关键点,并基于关键点推导出了各层斗的转动以及栌斗和阑额相对转动的计算公式;其次通过采集斗栱实景图像、实验室模型图像以及缩尺模型图像,构建了斗栱节点多样性数据集;而后搭建YOLOv8-Pose姿态估计模型,并开展了6种规模和Batch Size的23种工况对比实验;结果表明,最优性能模型目标检测的mAP50(B)达到0.94,关键点检测的mAP50(P)达到0.91.最后利用缩尺模型转动变形检测实验验证了所提方法的有效性.

Dou-Gong joints of ancient wooden structure are composed of Dou and Gong and other components.Affected by its own shape,environment and external force,such“stackable”joints are prone to rotation deformation and other types of damage.In view of the limitations of traditional detection methods of manual and deployed sensors in the implementation and detection efficiency,a computer vision detection method of Dougong rotation deformation based on pose estimation was proposed.Firstly,the key points of the Dou-Gong were defined,and based on the key points,the calculation formula of the rotation of each layer of the Dou-Gong and the relative rotation of the Lu-Dou and the Lan-E were deduced.Secondly,a diverse dataset of Dou-Gong joints was constructed by collecting Dougong real scene image,laboratory model image and scaled model image.Then,the YOLOv8-Pose pose estimation model was built,and the comparative experiments were conducted under 23 conditions with 6 varying sizes and Batch Sizes.The results showed that the target detection part mAP50(B)of the optimal performance model reached 0.94,and the key point detection part mAP50(P)reached 0.91.The effectiveness of the proposed method was verified by the rotation deformation detection experiment of the scale model.