Vision-based maize field zone classification for control of robot automatically dispensing granular fertilizer

Abstract

Importance of the work: Manual application of granular fertilizers is time-consuming
and labor-intensive. Robot design and control are required to replace human labor.
Objectives: to develop and test an automatic granular fertilizer dispenser (AGFD) robot
to replace human labor using a vision-based technique.
Materials and Methods: The problem of zone identification was investigated to locate
the robot’s position based on four defined zones: Front, Middle, Back and No-Row.
A method was proposed to identify these maize field zones using the “You only look once”
classification of red-green-blue (RGB) image frames obtained as the robot navigates in
a maize field based on dead reckoning.
Results: The proposed method for zone identification, implementation of fertilizer system
control and robot navigation in the maize field were successful. The fertilization and
navigation control commands of the AGFD robot were successfully implemented with
a success rate higher than 96% in real-time. The number of frames varied in the range
1,103–2,773 and the number of correct commands for each frame was counted. There was
variation in the field conditions due to light, weeds, growth stages and shadows. When
inconsistency occurred in the commands, the sequence of the designed state machine
could prevent improper state changes and prevent unexpected behavior by the robot.
Main finding: The method of zone classification using front-view RGB images and the
vision-based state machine design could successfully control the robot under various field
conditions. The proposed zone classification could be applied to other types of robots and
other row-crop types under field cultivation.