【作者】 曾伟；

【导师】 管涛；

【作者基本信息】 华中科技大学 ， 计算机应用技术， 2022， 硕士

【摘要】 近年来,三维建模技术被广泛应用于航姿测量、安全监控、虚拟飞行舱之类的非接触测量和数据可视化领域。在航空工业空气动力学试验中,高清图像采集设备在一次风洞试验中可以获取大量二维弱纹理图像序列,但是基于多视图几何的传统三维建模方法无法在这些图像上获取足量的匹配特征点,致使建模结果准确度差强人意,难以满足非接触式测量对三维建模精度的要求。因此,有必要开展弱纹理对象三维建模方法的研究。针对上述问题,在单视图三维建模方法的基础上,改进了损失函数和特征提取网络,同时搭配改进的模型粗配准与多视图模型融合方法,设计并实现了一种面向弱纹理对象的三维建模方法。该方法采用了弱纹理金属网格模型拟合调整的方式,根据预分类选取调整基础模型,并根据图像特征信息对基础模型进行顶点调整,来完成弱纹理模型重建获得多视角的重建子模型。在此基础上,对传统粗配准与精配准算法进行了改进,完成多视图子模型配准,并对配准后的网格模型设计了顶点融合和三角面片索引方法,解决了单目图像遮蔽部分建模精度下降的问题,为生成高质量的三维模型提供了保障。实验结果表明,整体方法相对传统三维建模方法对弱纹理物体的处理能力更强,建模成功率得到了一定提升;相比目前基于深度学习的单目建模算法,可以将多个单视图模型融合为精度更高的三维网格模型,重建精度提升了约8%,为非接触测量提供了更精确的参考数据。更多还原

【Abstract】 In recent years,3D modeling techniques have been widely used in non-contact measurement and data visualization fields such as aerial attitude measurement,safety monitoring,and virtual flight cabin.In the aerodynamic test of aviation industry,highdefinition image acquisition equipment can acquire a large number of 2D weakly textured image sequences in one wind tunnel test,but the traditional 3D modeling method based on multi-view geometry cannot obtain sufficient matching feature points on these images,resulting in poor accuracy of modeling results,which is difficult to meet the requirements of non-contact measurement for 3D modeling accuracy.Therefore,it is necessary to carry out research on 3D modeling methods for weakly textured objects.Based on the single-view 3D modeling method,a 3D modeling method for weakly textured objects is designed and implemented by improving the loss function and feature extraction process,together with the improved point cloud alignment and multi-view fusion method.The method adopts a weakly textured metal mesh model fitting adjustment method,and adjusts the base model according to the pre-classification selection,and adjusts the vertices of the base model according to the image feature information to complete the reconstruction of the weakly textured model to obtain the reconstructed sub-model with multiple views.On this basis,the traditional coarse alignment and fine alignment algorithms are improved to complete the multi-view sub-model alignment,and the vertex fusion and triangular surface slice indexing methods are designed for the aligned mesh model to solve the problem of decreasing modeling accuracy of the masked part of the monocular image,which provides a guarantee for generating high-quality 3D models.The experimental results show that this method is more capable of handling weakly textured objects compared with traditional 3D modeling methods,and the modeling success rate is greatly improved;compared with the current deep learning-based monocular modeling algorithm,multiple monocular view models can be fused into a 3D mesh model with higher accuracy,and the reconstruction accuracy is improved by about 8%,which provides more accurate reference data for non-contact measurements.更多还原