Ground-Based Robotic Phenotyping: Early Season Stand Counting

Researcher(s)

  • Hannah Chen, Computer Engineering, Texas A&M University

Faculty Mentor(s)

  • Yin Bao, Department of Plant and Soil Sciences, Department of Mechanical Engineering, University of Delaware

Abstract

Early season stand count of lima beans provides farmers and researchers with crucial insight into expected crop yield and germination rate. However, manually counting lima bean seedlings can be a time consuming and labor intensive process. High throughput phenotyping (HTTP) is an emerging method that efficiently gathers data regarding plant traits and characteristics, significantly reducing manual labor and time. HTTP utilizes robots equipped with sensors and cameras to efficiently gather and process plant data. This project explores counting early season lima bean seedlings through HTTP. Each week, the Amiga farm-ng robot is driven along crop rows, collecting video data. The video frames are then extracted as images and manually annotated to train a YOLOv11 detection model. After successfully training the model, various tracking algorithms were tested, such as BoTSORT, DeepSORT, and ByteTrack. DeepSORT was found to be the most accurate, however, it struggled to track on plots with a large number of frame jumps. To account for the frame jumps, a custom tracking algorithm was also developed and tested. The custom tracking algorithm crops the YOLO model predictions as templates for matching in the following frames. Feature extraction is utilized to determine when there were frame jumps, and estimate the location of the seedlings. The custom tracking algorithm ultimately had a higher accuracy than DeepSORT. Future applications of this method can include sorghum stand and tiller counting.