Abstract：Phenotyping is important for understanding of the relationships between plant genotypes and environment. Developing efficient and low-cost phenotyping technologies is a typical demand in many fields such asprecision agriculture. As a representative RGB-D device, Kinect has been used for plant phenotyping, but its technical potential has not been fully explored. To address this gap, this study compared the three mainstream principles of Kinect characterizing three-dimensional structures, i.e., point clouds generated from depth images (DI), from the color images using the method of Structure from Motion (SfM), and from the data by merging the DI- and SfM-derived data (MD). The performance of the three methods was evaluated based on the reference data, which was measured by a FAROX330 laser scanner. The results after the analyses in the case of Hosta plantaginea showed that DI made the most accurate estimationsin terms of leaf areas, MD out performed DI and SfM when regarding the predictions of leaf circularities and eccentricities, and SfM had the best performance on the retrievals of leaf inclinations. The difference between the results of the three methods stems from their distinctive performance for different structures. For leaf area estimation,SfMcan characterize plant leaves in a relatively incomplete way, while the edges of the MD-recon structed leaves are not smooth, resulting in the lowness of accuracy for these two methods. For the geometric characteristics of leaves, MD point clouds generated by merging the related DI and SfM data can achieve the effect of information enhancement, making its performance better than DI and SFM point clouds. The leaf inclination angle is more sensitive to the accuracy of depth measurement. Due to Kinect depth measurement often with the errors, the accuracies of the DI and MD point cloud-based leaf inclination retrievals may be low. TheSfM point cloudsare only generated from the color images, and so this method canpresent the best performance on retrieval of leaf inclination angles. Performance comparison sindicated that the three methods have their advantages for different structural features.Their integration can help to improve the overall performance of Kinect for plant phenotyping and,eventually, forma new Kinect-based mobile phenotyping technique. In addition, the proposed leaf geometry delineation (LGD) model proved todraw the contours of leaves and restore the geometries of those partially occluded leaves. Overall, this study developed a novel Kinect-based low-cost but efficient mobile three-dimensional plant structure phenotyping technique, which is of implications for promoting crop monitoring and increasing agricultural production.
Key words：Plant phenotyping; Point cloud; Structural parameter extraction; Kinect; LiDAR
孟祥爽，林 沂. 基于移动Kinect的低成本植物三维结构表型分析[J]. 光谱学与光谱分析, 2020, 40(08): 2352-2357.
MENG Xiang-shuang, LIN Yi. Kinect Sensor Moving for Low-Cost Mobile Phenotyping of 3D Plant Structures. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2020, 40(08): 2352-2357.
[1] White J W, Andrade-Sanchez P, Gore M A, et al. Field Crops Research, 2012, 133: 101.
[2] Fourcaud T, Zhang X, Stokes A, et al. Annals of Botany, 2008, 101(8): 1053.
[3] Lin Y. Computers and Electronics in Agriculture, 2015, 119: 61.
[4] Li L, Zhang Q, Huang D. Sensors, 2014, 14(11): 20078.
[5] Malambo L, Popescu S C, Horne D W, et al. ISPRS Journal of Photogrammetry and Remote Sensing, 2019, 149: 1.
[6] Shalal N, Low T, McCarthy C, et al. Computers and Electronics in Agriculture, 2015, 119: 254.
[7] Bao Y, Tang L, Breitzman M W, et al. Journal of Field Robotics, 2019, 36(2): 397.
[8] Wang Y, Wen W, Wu S, et al. Remote Sensing, 2019, 11(1): 63.
[9] Jay S, Rabatel G, Hadoux X, et al. Computers and Electronics in Agriculture, 2015, 110: 70.
[10] Andújar D, Calle M, Fernández-Quintanilla C, et al. Sensors, 2018, 18(4): 1077.
[11] Vázquez-Arellano M, Paraforos D S, Reiser D, et al. Computers and Electronics in Agriculture, 2018, 154: 276.
[12] Bao Y, Tang L, Srinivasan S, et al. Biosystems Engineering, 2019, 178: 86.
[13] Wu C. Towards Linear-Time Incremental Structure from Motion. Washington, 2013.
[14] Khoshelham K, Elberink S O. Sensors, 2012, 12(2): 1437.
[15] Chéné Y, Rousseau D, Lucidarme P, et al. Computers and Electronics in Agriculture, 2012, 82: 122.
