AUTHOR=Wang Jianxia , Sun Wenbing 

TITLE=Cluster segmentation and stereo vision-based apple localization algorithm for robotic harvesting

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 16 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2025.1598414

DOI=10.3389/fpls.2025.1598414

ISSN=1664-462X

ABSTRACT=IntroductionAutomated apple harvesting is hindered by clustered fruits, varying illumination, and inconsistent depth perception in complex orchard environments. While deep learning models such as Faster R-CNN and YOLO provide accurate 2D detection, they require large annotated datasets and high computational resources, and often lack the precise 3D localisation required for robotic picking.MethodsThis study proposes an enhanced K-Means clustering segmentation algorithm integrated with a stereo-vision system for accurate 3D apple localisation. Multi-feature fusion combining colour, morphology, and texture descriptors was applied to improve segmentation robustness. A block-matching stereo model was used to compute disparity and derive 3D coordinates. The method was evaluated against Faster R-CNN, YOLOv7, Mask R-CNN, SSD, DBSCAN, MISA, and HCA using metrics including Recognition Accuracy (RA), mean Average Precision (mAP), Mean Coordinate Deviation (MCD), Correct Recognition Rate (CRR), Frames Per Second (FPS), and depth-localisation error.ResultsThe proposed method achieved >91% detection accuracy and <1% localisation error across challenging orchard conditions. Compared with Faster R-CNN, it maintained higher RA and lower MCD under high fruit overlap and variable lighting. Depth estimation achieved errors between 0.4%–0.97% at 800–1100 mm distances, confirming high spatial accuracy. The proposed model exceeded YOLOv7, SSD, FCN, and Mask R-CNN in F1-score, mAP, and FPS during complex lighting, occlusion, wind disturbance, and dense fruit distributions.Discussion and ConclusionThe clustering-based stereo-vision framework provides stable 3D localisation and robust segmentation without large training datasets or high-performance hardware. Its low computational demand and strong performance under diverse orchard conditions make it suitable for real-time robotic harvesting. Future work will focus on large-scale orchard deployment, parallel optimisation, and adaptation to additional fruit types.