Method for Improving Speed of Instance Segmentation by Limiting Image Area Using High Precision Maps in Agricultural Field

Abstract

In order to make precision agriculture more accurate, real-time operations for individual crops is a challenge. It is essential to measure the crop area accurately when approaching individual crops. However, existing instance segmentation methods are computationally expensive, and it is difficult to recognize crops on devices in real-time. In this paper, we develop a crop recognition method for a gripper weeding machine. By creating a high-precision crop map in advance and limiting the image area based on the map, the location and the azimuth of the vehicle from GPS units, we succeeded in the real-time and high-accurate individual crop recognition on Jetson AGX Xavier.