Adaptive Multi-ROI Agricultural Robot Navigation Line Extraction Based on Image Semantic Segmentation

Abstract

Automated robots are an important part of realizing sustainable food production in smart agriculture. Agricultural robots require a powerful and precise navigation system to be able to perform tasks in the field. Aiming at the problems of complex image background, as well as weed and light interference factors of the visual navigation system in field and greenhouse environments, a Faster-U-net model that retains the advantages of the U-net model feature jump connection is proposed. Based on the U-net model, pruning and optimization were carried out to predict crop ridges. Firstly, a corn dataset was trained to obtain the weight of the corn dataset. Then, the training weight of the obtained corn dataset was used as the pretraining weight for the cucumber, wheat, and tomato datasets, respectively. The three datasets were trained separately. Finally, the navigation line between ridges and the yaw angle of the robot were generated by B-spline curve fitting. The experimental results showed that the parameters of the improved path segmentation model were reduced by 65.86%, and the mPA was 97.39%. The recognition accuracy MIoU of the Faster-U-net model for maize, tomatoes, cucumbers, and wheat was 93.86%, 94.01%, 93.14%, and 89.10%, respectively. The processing speed of the single-core CPU was 22.32 fps/s. The proposed method had strong robustness in predicting rows of different crops. The average angle difference of the navigation line under a ridge environment such as that for corn, tomatoes, cucumbers, or wheat was 0.624°, 0.556°, 0.526°, and 0.999°, respectively. This research can provide technical support and reference for the research and development of intelligent agricultural robot navigation equipment in the field.

Keywords

• semantic segmentation
• transfer learning
• navigation
• image processing
• path recognition
• multiple regions of interest