VISION-BASED POINT CLOUD PROCESSING FRAMEWORK FOR HIGH THROUGHPUT PHENOTYPING

High throughput phenotyping is an emerging field that aims to bring rapid, non-invasive sensing technology in agriculture to accelerate the estimation of plant traits significantly. This paper presents a computer vision-based automated 3D point cloud processing framework for accurate estimation of essential phenotypic traits. The framework relies on three steps involving sub-plot detection, extraction of the crop from each sub-plot, and estimating the required trait. Four essential phenotypic traits are estimated as a use case of the proposed framework, namely, plant height, leaf area index (LAI), leaf inclination, and plant count. The crop of interest is mung bean. The obtained estimates for plant height, leaf area index (LAI), and leaf inclination are statistically validated by comparing the results with ground truth data in terms of coefficient of determination, root mean squared error (RMSE), and correlation coefficient. These metrics are found to be, on average, 0.87, 0.05, and 0.93 respectively. The regression analysis has also been performed to gain analytical insights into the data. For plant count, deep learning based segmentation method have been explored and the best accuracy achieved is 86%.