基于Kinect相机的油麦菜自动化三维点云重建

Automated 3D Reconstruction of Leaf Lettuce Based on Kinect Camera

摘要:

为了解决传统三维点云重建过程中人工调参费时、费力，且精度得不到保障等问题，提出了一种三维点云自动化配准算法，并应用于油麦菜三维重建。使用Kinect相机采集油麦菜不同视角下的点云数据，通过配准实验分析配准参数的变化规律，继而建立了配准评价体系，实现了两片点云的自动化配准，并通过最小化匹配误差积累将多幅点云变换到同一基准坐标系下，实现了油麦菜三维重建。对随机选取的12株油麦菜进行自动化三维重建，结果表明，在两片点云重叠率不低于30%的前提下，本文算法可获得最优参数组合，自动全局配准平均距离误差为0.65cm，平均耗时为44.05s，具有较高的配准精确度和稳定性。本文算法能有效减少配准误差积累、构建较高精度的完整结构，可为其他作物三维重建提供参考。

Abstract:

In order to solve the problems such as manual adjustment of parameters during traditional three-dimensional (3D) point cloud reconstruction process which is time-consuming and laborious, and the registration accuracy was not guaranteed, a 3D point cloud automatic registration algorithm was proposed and applied to the 3D reconstruction research of leaf lettuce. Firstly, a Kinect camera was used to collect point cloud data from different perspectives of the leaf lettuce. Secondly, the changing patterns of parameters during the registration process were investigated through a large number of registration experiments, and accordingly each parameter’s initial value was determined due to its most positive impact to the result. Thirdly, a registration evaluation system was established, which included the inner point overlap rate, point dispersion degree and initial registration distance error, so that the automatic registration algorithm of two point clouds were implemented. Finally, based on point cloud automatic registration algorithm, a leaf lettuce point cloud 3D reconstruction was achieved because the accumulation errors were minimized through two adjacent point clouds’ automatic registration. And then the obtained point clouds were converted to the same target coordinate system therefore the leaf lettuce 3D point cloud was reconstructed. The automatic three-dimensional reconstruction experiment was carried out on 12 lettuce plants, and the results showed that under the premise, the overlap of two point clouds was not less than 30%, the automatic registration algorithm can get the optimal parameter combination by applying the registration evaluation system;the average registration error of global registration was 0.65cm, the average registration efficiency was 44.05s, and the algorithm greatly improved the accuracy and stability of registration;the leaf lettuce point cloud 3D reconstruction algorithm can effectively reduce the registration error accumulation, and provide complete structural and morphological data for further measurement of plant phenotypic parameters, and it can be used in other plants’ 3D reconstruction and phenotype researches.