Trellis wire reconstruction by line anchor-based detection with vertical stereo vision

Abstract

Detecting and reconstructing thin trellis wires in agricultural environments, particularly under occluded conditions, presents a significant challenge for current depth sensors, which struggle to capture the depth of such thin structures. This study introduces Wire-CLRNet, a line anchor-based convolutional neural network architecture designed to detect trellis wires under occluded outdoor conditions. Wire-CLRNet is integrated into a novel framework that translates the detected planar information into accurate spatial information critical for robotic operations in orchards using vertically configured stereo vision. The proposed system improves depth estimation and provides a comprehensive solution for both planar and spatial wire reconstruction. The framework is tested on real-world and simulated environment built within Nvidia Isaac Sim. Experimental results demonstrate that Wire-CLRNet achieved 0.9345 m F 1 score for planar reconstruction and 0.0303 m mean distance error for spatial reconstruction. The study demonstrates that the system can achieve better accuracy under occlusion conditions, offering a practical solution for agricultural robots tasked with harvesting and pruning. • A novel deep learning-based object-focused stereo vision system was developed. • Line anchor-based method was proposed for wire 2D detection and reconstruction. • Proposed method yields better accuracy and robustness in obscured settings. • Proposed scheme recovers 3D information of obscured trellis wires in orchard.