Jimin Lee, Jungin Kim*, Occluded Region Restoration and Geometric Positioning of Fruits Using Deep Learning-Based Pose Estimation and Landmark Detection, Journal of the Korean Institute of Intelligent Systems, 2025.

Harvesting robots have received increasing attention in the agricultural sector, but conventional object detection methods often fail to accurately recognize fruits that are partially occluded by leaves or other fruits. To address this issue, we propose a deep learning-based approach that integrates object pose estimation and landmark detection to recover occluded fruit regions and estimate their geometric positioning. A synthetic RGB-D dataset was generated using a Unity-based virtual environment, incorporating diverse occlusion and rotation conditions. Using the YOLO11s-pose model, fruit contour landmarks were automatically extracted from 2D images and used to estimate the rotation angles and spatial orientation of the target fruits. Experimental results demonstrated high accuracy, achieving an mAP50 of 0.995 for object and pose detection, with a mean absolute rotation error of 2.4° and a normalized landmark detection error of 0.04. These findings indicate that the proposed method enables robust fruit recognition under partial occlusion and holds promise for real-world automation in agricultural harvesting systems.