Improved Real-Time Semantic Segmentation Network Model for Crop Vision Navigation Line Detection.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Robotics

Background:

• Efficient field crop management relies on precise row detection for automated machinery.
• Unmanned aerial vehicles (UAVs) require robust visual navigation for real-time precision agriculture operations.
• Existing methods for crop row segmentation often lack the efficiency and accuracy needed for UAV applications.

Purpose of the Study:

• To develop an advanced semantic segmentation model for accurate crop row detection in farmland imagery.
• To enable real-time visual navigation and precise operations for agricultural UAVs.
• To enhance the efficiency and applicability of smart agricultural management systems.

Main Methods:

• An improved ENet semantic segmentation network was designed, featuring compressed convolutions and a residual network structure for efficient feature extraction.
• A shunting process within the network facilitates the backward flow of low-dimensional boundary information, improving segmentation accuracy.
• An enhanced random sampling consensus (RANSAC) algorithm with a novel scoring index was employed for robust navigation line extraction and fitting.

Main Results:

• The proposed model achieved accurate and efficient segmentation of crop rows in farmland images.
• The navigation line extraction algorithm demonstrated strong robustness and high applicability.
• The method significantly improved the accuracy of boundary localization and row-to-row segmentation.

Conclusions:

• The developed approach provides a reliable technical foundation for autonomous navigation and precise operations of agricultural UAVs.
• The improved ENet model and RANSAC algorithm offer a viable solution for real-time smart agricultural management.
• This research contributes to advancing precision agriculture through enhanced UAV capabilities.