Maize Detection and Row Extraction Using Maize-YOLO and IPM-Clustering Method for Autonomous Agricultural Navigation

Area of Science:

• Agricultural Engineering
• Computer Vision
• Robotics

Background:

• Accurate crop row extraction is crucial for autonomous agricultural machinery navigation.
• Challenges in maize fields include variable illumination, occlusion, and complex backgrounds, hindering traditional navigation methods.

Purpose of the Study:

• To develop a robust and efficient plant-oriented framework for maize crop row perception.
• To enhance the reliability and accuracy of vision-based navigation for autonomous agricultural robots.

Main Methods:

• A lightweight object detection model, Maize-YOLO (based on YOLOv11n), was developed with optimized modules (C3k2_Faster_CGLU, Detect_LSH) and pruning strategies.
• A pipeline (IPM-DBSCAN-LSM) was used to estimate row centerlines from detected plant positions, employing inverse perspective mapping, DBSCAN clustering, and least squares fitting.

Main Results:

• Maize-YOLO achieved 97.6% mAP@0.5, reduced GFLOPs by 61.9%, and reached 84.4 FPS on CPU.
• The crop row extraction achieved 98.6% positional accuracy and a mean angular deviation of 0.44°.

Conclusions:

• The proposed framework offers accurate, lightweight, and real-time crop row perception for autonomous agricultural navigation.
• This method enhances precision field operations by overcoming common challenges in maize cultivation environments.