Answers are grounded in this paper and cite their sources.
Rapid and stable extraction of maize-row navigation lines is essential for autonomous inter-row operations by agricultural robots. Existing deep-learning methods typically detect individual plants or segment maize rows, then derive navigation feature points from geometric cues including bounding boxes, region centroids, edges, or skeletons. These stages can accumulate errors, while additional post-processing increases computational cost. This study developed point annotations representing navigation-line geometry for end-to-end prediction of navigation keypoints from field images. A navigation-keypoint dataset covering complex maize-row scenarios was constructed from seedling-stage field images collected under diverse natural conditions. Each row was an independent instance. Keypoints generated at fixed vertical positions from manually annotated navigation-line endpoints were directly predicted as structured sequences using YOLOv11n-pose and fitted by least squares. Field-image experiments yielded Pose AP50 and Pose AP50–95 of 99.3% and 95.5%, respectively. Navigation-line normalized root mean square error, normalized lateral error, and angular error were 1.221%, 1.090%, and 1.206°, respectively, with all test navigation lines matched. Processing reached 91.90 frames per second without separate navigation feature-point calculation, balancing detection completeness and real-time performance. In 200-frame video sequences, maximum adjacent-frame heading-angle changes were 8.774° and 2.920° under normal-row and sparse-row conditions, respectively. An observed transient spike under complex interference subsided in the next frame. These results demonstrate that spatially constrained point annotations enable direct learning of navigation-line geometry and reduce dependence on intermediate region processing, providing an effective path-perception approach for real-time field navigation by agricultural robots.
No takes yet. Share an insight, caveat, or question.
Shang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: