基于单视图稀疏点的汽车三维模型重建  

Reconstructing 3D automobile model from sparse points on a single view

作　　者：王博[1] 江祖毅 Wang Bo;Jiang Zuyi(College of Automobile and Traffic Engineering,Wuhan University of Science and Technology,Wuhan 430065,China)

机构地区：[1]武汉科技大学汽车与交通工程学院,湖北武汉430065

出　　处：《武汉科技大学学报》2023年第4期296-302,共7页Journal of Wuhan University of Science and Technology

基　　金：国家自然科学基金资助项目(51905389).

摘　　要：基于深度学习的图像识别模型训练需要大量数据,而不同角度的汽车视图数据难以在短时间内收集,为此提出一种利用单视图稀疏点的汽车三维模型重建方法,依靠少量数据也能得到精确的结果。创建了包含3000多张不同汽车品牌的多视角二维汽车图形数据集,并在TensorFlow框架下搭建了基于MobileNet V2网络和迁移学习的汽车视图角度识别系统,其结果能够进一步用于快速的模型匹配及重建;根据创建的汽车三维线框模型库以及二维关键点和三维模型的映射关系,利用带约束的最小二乘法求出模型库中不同模型对于重建的贡献量系数,直接由二维图片上稀疏的25个关键点生成三维模型。误差分析结果显示,重建的三维车身模型具有较高精度。The training of image recognition model based on deep learning requires a large amount of data,and it is difficult to collect car views at different angles in a short time.Therefore,a 3D automobile model reconstruction method based on sparse points on a single view was proposed,which can obtain accurate results even with a small amount of data.Firstly,a multi-view 2D graphic dataset containing more than 3000 images of different car brands was created,and a recognition system for the angle of car view using MobileNet V2 network and transfer learning was built in TensorFlow frame work.The recognition results can be further used for rapid model matching and reconstruction.Secondly,according to the created 3D wireframe model base and the mapping relationship between 2D key points and 3D model,the constrained least square method was used to calculate the contribution coefficients of different models in the model base to the reconstruction,and 3D model was directly generated from 25 sparse key points on the 2D image.The error analysis results show that the reconstructed 3D car body models have high precision.

关 键 词：三维模型重建 汽车造型 图像识别 单视图 稀疏点 MobileNet V2 迁移学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]