Abstract
This study proposes a navigation system for the headland autonomous turning of a safflower picking robot. The proposed system includes binocular cameras, differential satellites, and inertial sensors. The method of extracting the headland boundary line combining the hue, saturation, and value-fixed threshold segmentation method and random sample consensus algorithm and planning the headland-turning trajectory of a robot based on the multiorder Bezier curve are used as control methods. In addition, a headland-turning tracking model of a safflower picking robot is designed, and a path-tracking control algorithm is developed. A field test verifies the performance of the designed headland-turning navigation system. The test results show that the accuracy of the judgment result regarding the existence of a headland is higher than 96%. In headland boundary detection, the angle deviation is less than 1.5˚, and the depth value error is less than 50 mm. The headland-turning path tracking test result shows that at a turning speed of 0.5 km/h, the average lateral deviation is 37 mm, and the turning time is 24.2 seconds. The results verify that this navigation system can accurately extract the headland boundary line and can successfully realise the headland-turning path tracking of a safflower picking robot.
Introduction
In precision agriculture, autonomous inter-row navigation is relatively well-established, but "headland turning"—the process of a robot exiting one row and accurately entering the next—remains a significant hurdle. Safflower (Carthamus tinctorius L.) fields present a particularly unstructured environment where lighting is inconsistent, the ground is uneven, and the boundary between the crop and the turning area is often blurred by vegetation.
Manual picking is labor-intensive and inefficient. While robots offer a solution, they must perceive where a row ends and plan a smooth, continuous turn that aligns perfectly with the next row without damaging the crop. This paper addresses the dual challenge of perception (identifying the headland boundary) and control (executing a smooth turn).
Methodology
The authors propose a multi-sensor fusion approach using a ZED2 binocular camera, a high-precision Differential Global Navigation Satellite System (DGNSS), and an Inertial Measurement Unit (IMU).
The workflow is divided into three main stages:
- Boundary Perception: Instead of standard RGB color space, the system uses HSV (Hue, Saturation, Value) thresholding to segment the headland, making it resilient to changing outdoor light. A RANSAC (Random Sample Consensus) algorithm is then applied to fit a straight line to the detected boundary points, filtering out noise from stray weeds.
- Path Planning: To ensure smooth movement, the system generates a turning trajectory using third-order Bezier curves. These curves provide continuous curvature, which is easier for a differential-drive robot to follow without jerky movements.
- Path Tracking: The robot uses a Fuzzy PID controller. The "Fuzzy" component allows the system to adaptively tune PID parameters in real-time based on the robot's lateral and angular deviation from the planned path.
Results and Discussion
Field tests conducted at a safflower plantation in Xinjiang, China, yielded impressive quantitative metrics. The system correctly identified the existence of a headland with over 96% accuracy. In boundary detection, the angular deviation was less than 1.5°, with a depth error of less than 50 mm. At an operational speed of 0.5 km/h, the robot achieved an average lateral deviation of only 37 mm and an angular deviation of 1.3°.
A critical insight from the study is the superiority of HSV over RGB color space for outdoor robotics; by isolating the "Value" (brightness), the system maintains accuracy despite shadows or direct sunlight. However, the results also highlight a speed-accuracy trade-off. Doubling the speed from 0.5 km/h to 1.0 km/h significantly increased lateral deviation, suggesting that "slow and steady" remains a technical requirement for autonomous precision in delicate crop environments.
Conclusion
The developed navigation system successfully enables autonomous headland turning for safflower picking robots by accurately extracting boundary lines and tracking planned paths. Future work will focus on adaptive speed control to handle varying soil conditions and power optimization to extend the robot's operational life in the field.
Summary updated