With the rapid advancements in genotyping technologies, plant phenotyping has become a bottleneck in exploiting the massive genomic data for crop improvement. The common practice of plant phenotyping relies on human efforts, which is labor-intensive, time-consuming, and prone to human errors. This dissertation documents our innovative research in automated plant phenotyping using 3D machine vision and robotics.

Sorghum and maize are important economic crops for food, feed, fuel, and fiber production. Manipulation of plant architecture plays a vital role in yield improvement via plant breeding. A high-throughput, field-based robotic phenotyping system was developed to characterize plant architecture for tall-growing sorghum plants having dense population and canopies. Side-viewing stereo cameras were used for 3D reconstruction of plants. A novel data processing pipeline was developed to measure plant height, width, convex hull volume, surface area, and stem diameter. These image-derived features were highly correlated with the in-field manual measurements, and with high repeatability.

Additionally, Time-of-Flight 3D imaging was used to collect side-view point clouds of maize plants under field conditions. Algorithms for extracting plant height, leaf angle, plant orientation, and stem diameter at plant level were developed. A customized skeletonization algorithm was developed to effectively reduce a large point cloud to a skeleton graph; and a 3D Hough line detection algorithm was implemented to find individual stems. The image-derived traits showed satisfactory accuracies, except for stem diameter due to the limitations of the sensor’s depth sensing precision.

Various instrumentation devices for plant physiology study require accurate placement of their sensor probes toward the leaf surface. A robotic leaf probing system was developed for a controlled environment using a Time-of-Flight sensor, a laser profilometer, and a six-DOF robotic manipulator. The Time-of-Flight sensor and the laser profilometer were utilized for environment mapping and high-precision scanning of plant canopies, respectively. The environment point cloud was used for collision-free motion planning and individual plant segmentation, while the high-resolution canopy point cloud was analyzed for leaf segmentation and probing point extraction. The system achieved an average motion planning time of 0.4 s with an average probe positioning error of 1.5 mm and probe orientation error of 0.84 degrees.