Evaluation of a Robotic System for Automatic Weed Identification Using Computer Vision

Abstract

This work presents an experimental analysis of the performance of the weed detection system that equips an agricultural robot in different scenarios and operating conditions. The system is based on computer vision, cloud processing, and the Internet of Things (IoT). Through an online web application, a digital twin coordinates tasks such as seeding, weed control, and irrigation in a real vegetable growing bed where the physical robot operates. In order to evaluate the performance of the weed detection system, the robot was tested experimentally in two different scenarios, one more conservative and the other with a higher degree of difficulty because artifacts that make weed detection more difficult were added to the scenario. In each scenario, experiments were carried out in three different lighting conditions. From the results presented, it is possible to assess in which situations the robot performs best in detecting weeds.