结合三维边缘特征约束的网格模型优化方法
张 昊(1997—),男,四川成都人,硕士,主要研究方向为网格模型构建与优化。E-mail: 921989874@qq.com |
收稿日期: 2023-10-26
修回日期: 2024-02-26
网络出版日期: 2024-05-21
基金资助
国家自然科学基金面上项目(41871379)
辽宁省兴辽英才计划项目(XLYC2007026)
辽宁省应用基础研究计划项目(2022JH2/101300273)
辽宁省应用基础研究计划项目(2022JH2/101300257)
Optimization of Mesh Model with 3D Edge Feature Constraints
Received date: 2023-10-26
Revised date: 2024-02-26
Online published: 2024-05-21
Supported by
National Natural Science Foundation of China(41871379)
Liaoning Revitalization Talents Program(XLYC2007026)
Fundamental Applied Research Foundation of Liaoning Province(2022JH2/101300273)
Fundamental Applied Research Foundation of Liaoning Province(2022JH2/101300257)
对于具有复杂几何结构特征的建筑物,三维网格模型重建结果易存在表面扭曲、边缘特征平滑的问题,无法较好地反映重建目标的真实信息。针对上述问题,本文提出一种结合三维边缘约束的网格模型优化方法。该方法以OpenMVS算法得到的初始网格为基础数据,利用变分原理构建能量函数,将网格模型优化问题转换为能量函数最小化问题。首先,从多视图像中提取三维边缘点,以三维边缘点构成的边缘轮廓来定位网格模型边缘区域;接着,利用灰度一致性测度构建数据项,利用顶点自身曲率构建平滑项,利用三维边缘轮廓约束构建附加约束项,将3个约束项构建为一个总体能量函数;最后,采用梯度下降法迭代求解总体能量函数最小值,将梯度变化量分配到网格模型的顶点上来驱动网格形变,以此优化模型。选取Strecha数据集中两个真实室外场景和ETH3D数据集中一个真实室内场景进行优化实验,使用ETH3D评估框架对实验结果进行评估。结果表明,经本文算法优化后的网格模型完整度和精度最高为89.76%、94.45%,本文算法在解决复杂几何结构建筑物模型边缘优化问题的基础上,提高了建筑物模型的准确程度。
张昊 , 王竞雪 , 谢潇 . 结合三维边缘特征约束的网格模型优化方法[J]. 地球信息科学学报, 2024 , 26(5) : 1138 -1150 . DOI: 10.12082/dqxxkx.2024.230633
The dense point cloud of the urban scene reconstructed by Multi-View Stereo reconstruction technology (MVS) often contains noise, resulting in surface distortion of the generated model and loss of some edge features, which cannot well reflect the real information of the reconstructed target. To solve these problems, a variational method combining 3D edge constraints is proposed to optimize the mesh model. Based on the initial grid data obtained by MVS algorithm, the energy function is constructed by the variational principle, and the grid model optimization problem is transformed into an energy function minimization problem. Firstly, the initial grid model is reconstructed from the dense point cloud. Then, the energy function is constructed by using the luminosity consistency measure, using the vertex curvature as the smooth term, and using the three-dimensional edge point constraint as the additional constraint term. Finally, the gradient descent method is used to solve the minimum energy function iteratively, and the grid deformation is driven by discretizing the gradient change to the vertex of the triangle to optimize the model. In order to construct 3D edge constraints, 3D edges must be extracted first. In this paper, 2D edges are extracted from multi-view images first, and the 2D edges are represented as multi-segment lines according to the polar constraints. Then, the 2D multi-segment line nodes are restored as 3D edge points according to the polar constraints, and the 3D edge points of the recovery points are a series of 3D multi-segment lines representing the edge outline. Finally, the edge region of the mesh model is located by taking the vertex of the mesh model closest to the 3D edge point as the neighborhood point. In this way, 3D edge features are constructed. In order to verify the effectiveness of the proposed algorithm, two real outdoor scenes from the Strecha dataset and one real indoor scene from the ETH3D dataset are selected to evaluate the reconstruction results of the proposed algorithm. In addition, the efficiency of this algorithm is analyzed by comparisons with other algorithms. Experimental results show that the proposed algorithm can effectively improve the accuracy and integrity of the grid model and retain the edge features of the target better on the grid model.
表1 三维边缘点点位误差Tab. 1 3D edge points position error |
数据 | 点集 | MAE/cm | σ | 占比/% | 三维边缘点数量/个 |
---|---|---|---|---|---|
Fountain-P11 | 初始网格顶点 | 2.48 | 17.18 | 96.7 | |
三维边缘点 | 0.66 | 9.41 | 97.9 | 16 567 | |
Herz-Jesu-P8 | 初始网格顶点 | 0.78 | 4.86 | 97.4 | |
三维边缘点 | 0.39 | 2.41 | 99.1 | 27 390 | |
Relief | 初始网格顶点 | 1.78 | 11.57 | 95.8 | |
三维边缘点 | 0.91 | 6.12 | 98.6 | 38 245 |
注:加粗数值为误差较小值。 |
表2 不同方法在不同评估阈值下的评估结果Tab. 2 Evaluation results of different methods under different evaluation thresholds |
数据 | 方法 | 阈值(Te) = 2 cm | 阈值(Te)= 10 cm | ||||
---|---|---|---|---|---|---|---|
完整度/% | 精度/% | F1 | 完整度/% | 精度/% | F1 | ||
Fountain-P11 | 初始网格 | 57.32 | 89.36 | 70.57 | 69.54 | 94.88 | 80.26 |
OpenMVS | 58.48 | 89.99 | 70.89 | 70.04 | 97.50 | 81.52 | |
文献[19]算法 | 58.53 | 90.11 | 70.96 | 70.51 | 96.82 | 81.59 | |
本文算法 | 57.69 | 92.72 | 71.12 | 69.95 | 98.21 | 81.71 | |
Herz-Jezu-P8 | 初始网格 | 36.15 | 65.68 | 46.63 | 63.15 | 93.61 | 75.42 |
OpenMVS | 41.38 | 68.63 | 51.63 | 63.18 | 94.09 | 75.60 | |
文献[19]算法 | 41.94 | 72.16 | 53.05 | 63.68 | 94.31 | 76.02 | |
本文算法 | 42.39 | 74.29 | 53.98 | 63.80 | 95.20 | 76.39 | |
Relief | 初始网格 | 89.17 | 94.89 | 91.94 | 93.46 | 97.47 | 95.43 |
OpenMVS | 89.77 | 94.43 | 92.04 | 93.46 | 97.82 | 95.59 | |
文献[19]算法 | 89.56 | 94.07 | 91.75 | 93.39 | 97.79 | 95.53 | |
本文算法 | 89.76 | 95.45 | 92.22 | 93.48 | 97.85 | 95.62 |
注:加粗数值为评估结果的最优值。 |
[1] |
|
[2] |
|
[3] |
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
|
[9] |
|
[10] |
|
[11] |
|
[12] |
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
俞思帆, 闫利, 谢洪, 等. 附加多视线特征约束的城市场景网格模型优化[J]. 测绘科学, 2022, 47(2):102-109.
[
|
[19] |
邓非, 陈欣, 颜青松, 等. 带线约束的摄影测量网格变分精化算法[J]. 测绘学报, 2020, 49(4):469-479.
[
|
[20] |
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
|
[27] |
|
[28] |
|
[29] |
|
/
〈 |
|
〉 |