FAST Keypoint detection on excess green images for navigation line extraction in maize fields

Abstract

Maize field interrow navigation requires high precision for mobile robot heading control because of the narrow soil area between maize rows. This paper presents a new method for extracting a navigation line in the interrow space of maize fields of an automatic granule fertilizer dispenser (AGFD) robot. The proposed algorithm is a crop row detection approach named FAST keypoint detection, which is applied to the excess green image. The FAST keypoints detected are distributed throughout the crop rows in the image. The crop row lines are fitted by the least squares method. Afterward, the navigation line is extracted. The initial region of interest (ROI) is automatically identified using YOLO-based soil segmentation and the adaptive boundary lines are generated. The RMSE and R² value of the navigation lines obtained from the proposed method is compared with that of the excess red- and excess green-based methods. The proposed method achieved a minimum RMSE of 10.06 pixels and a maximum R² value of 0.75. The results reveal that the FAST keypoint detection method on the excess green image outperforms the benchmarks methods for various stages of maize growth with respect to both higher accuracy and less computation time. The proposed method achieved a maximum RMSE improvement of 10 pixels (43.74%), reducing the error from 22.86 pixels, while requiring a computational time of only 53 ms, corresponding to 17% of that of the benchmark method. The results of the proposed method under various field conditions, including differences in illumination, soil color, and vegetation density, are also presented. The proposed method is capable of real-time implementation on the AGFD robot.