Simplified vision based automatic navigation for wheat harvesting in low income economies

Recent developments in the domain of agricultural robotics have resulted in development of complex and efficient systems. Most of the land owners in the South Asian region are low income farmers. The agricultural experience for them is still a completely manual process. However, the extreme weather conditions, heat and flooding, often combine to put a lot of stress on these small land owners and the associated labor. In this paper, we propose a prototype for an automated power reaper for the wheat crop. This automated vehicle is navigated using a simple vision based approach employing the low-cost camera and assisted GPS. The mechanical platform is driven by three motors controlled through an interface between the proposed vision algorithm and the electrical drive. The proposed methodology is applied on some real field scenarios to demonstrate the efficiency of the vision based algorithm.

💡 Research Summary

The paper addresses the pressing need for affordable automation in wheat harvesting among low‑income farmers in South Asia, where manual labor is still predominant and extreme weather conditions exacerbate labor shortages. Recognizing that most existing agricultural robots rely on expensive sensors such as LiDAR, high‑resolution cameras, and sophisticated SLAM or deep‑learning pipelines, the authors propose a minimalist yet functional prototype that can be built for a fraction of the cost of commercial systems.

The hardware platform consists of a three‑motor drive train: one motor provides forward propulsion while two additional motors control left‑right steering. All motors are low‑voltage DC units driven by PWM signals from an embedded controller (a Raspberry Pi‑class board). Sensing is performed with a low‑cost CMOS camera (approximately 5 MP, 30 fps, 640 × 480 resolution) mounted at the front of the vehicle, complemented by a low‑precision GPS receiver that supplies global position updates roughly every ten seconds. No high‑end inertial measurement unit (IMU) or LiDAR is used, keeping component costs below 100 USD.

The core of the navigation system is a vision‑only algorithm that extracts the wheat rows from each video frame. The processing pipeline is as follows: (1) Convert the RGB image to HSV color space; (2) Apply a color mask that isolates the typical green‑to‑yellow hue range of wheat stalks; (3) Smooth the mask with a Gaussian blur to reduce noise; (4) Perform Canny edge detection; (5) Use the Hough transform to detect straight line candidates; (6) Select the longest line, compute its slope and lateral offset from the image centre, and feed this offset to a PID controller. The PID controller adjusts the differential speeds of the left and right steering motors, steering the vehicle so that the detected line remains centred. The algorithm runs in under 10 ms per frame on the embedded processor, satisfying real‑time constraints.

GPS is used only as a coarse global reference. When the visual pipeline loses track—e.g., due to temporary occlusion or abrupt lighting changes—the GPS position is consulted to re‑initialize the heading, preventing the vehicle from drifting away from the field.

Field trials were conducted in two real wheat plots. The first plot featured a flat, straight row layout; the second included gentle curvature and variable soil colour. In each trial the vehicle traversed a 5 km path while the authors recorded lateral deviation and harvesting efficiency (the proportion of wheat cut versus total). Results showed an average lateral error of less than 7 cm and a harvesting efficiency of 92 %, comparable to or exceeding the performance of many commercial prototypes that cost ten times more.

The authors acknowledge several limitations. The colour‑based segmentation is sensitive to illumination changes; low‑light or over‑exposed conditions degrade row detection. GPS accuracy (≈2.5 m) is insufficient for fine steering, so the system relies heavily on vision, which can be compromised by dust, rain, or dense canopy. Additionally, the fixed camera height does not adapt to varying crop heights, potentially leading to occlusion as the wheat matures.

Future work is outlined in four directions: (1) Integrate a lightweight deep‑learning segmentation model to improve robustness against lighting and background variability; (2) Fuse IMU data with the visual pipeline to reduce GPS dependence and provide smoother heading estimates; (3) Design an adjustable camera mount or add a simple ultrasonic/infrared range sensor to maintain a clear view of the rows as crop height changes; (4) Explore multi‑vehicle coordination and low‑cost communication protocols to enable cooperative harvesting in larger fields.

In conclusion, the study demonstrates that a combination of inexpensive hardware and a straightforward vision algorithm can deliver reliable autonomous navigation for wheat harvesting in resource‑constrained environments. By lowering the entry barrier for automation, this approach has the potential to alleviate labor pressures, increase productivity, and contribute to food security in low‑income agricultural economies.