Weeding Robot Based on Lightweight Platform and Dual Cameras

Abstract

In the process of planting, weeds will inevitably grow in the farmland, and compete with crops for water, light and space, which obviously affect our normal agriculture. If weeds are not effectively controlled, crops output will be seriously compromised. On the other hand, it also increases the number of pests. Nowadays, the main weeding methods rely on labor and chemical herbicide. However, manual weeding is inefficient, costly, time consuming and cannot remove weeds effectively. The purpose of this paper is to propose a weeding robot. The focus of the research is on how to use dual cameras to accurately detect weeds. The convolutional neural networks (CNNs), deep learning, dual cameras machine vision and mechanical design will be discussed in this paper. The results show that dual cameras robot based on a new lightweight platform can achieve a high accuracy compared to single camera method, while a feasible rail system was proposed for weeding robots. In vegetable detection, this method achieves 98.12% precision, 83.47% recall and 89.91% mAP that is 4.06% higher than a single top view camera. GF-YOLO, a lightweight platform we proposed also outperform other state-of-the-art algorithms in embedded system.