Deep Learning-based Vision Systems for Agricultural Robots: Overcoming Obscured Apple Orchard Challenges

The thesis focuses on developing vision systems for agricultural robots to perceive crops and obstacles in orchards. It introduces a novel framework to reconstruct 3D obscured branches using stereo depth cameras and deep learning. The second phase estimates fruit orientation using keypoint detection and circle-based segmentation, improving harvesting efficiency. The final phase enhances trellis wire detection with a CNN-based model, improving depth computation for navigation. The proposed methods improve accuracy and robustness, addressing limitations of current depth sensors and enhancing robotic operations in orchards. Experimental results validate their effectiveness for real-world applications in automated harvesting.