Validating Point Cloud Registration for Precision Agriculture Using AprilTags

3D data allows for better tracking of the health of plants, allowing for optimal care for each plant. This reduces water consumption, decreases reliance on pesticides, and increases crop yield. One major barrier to collecting 3D plant data is registration of point clouds collected from different perspectives. Unlike traditional man-made targets, such as buildings and consumer products, organic targets have high variability in shape and structure. Traditional methods of point cloud registration struggle with these highly variable features. To overcome this, a method of data collection involving fixed artificial markers is presented. A D435i depth camera is attached to the end of a my Cobot 320 M5 robot arm, and a frame of AprilTags is built around the scaffolding the plants grow on. As the plants are scanned, the camera position and orientation is determined by viewing the AprilTags. This information is used to register the point clouds. The localization of this setup is measured to be less than 0.7 cm by placing the depth camera in two positions that are fixed relative to each other, then subtracting the results. The registration error is measured by comparing a known object across scans from different positions. The average root mean square error of the registration is 2.16 mm. This system is then validated by successfully scanning and registering a pink hyacinth and a green pepper sprout.