Automated cucumber harvesting requires a perception system that can detect fruit, assess ripeness and localize the peduncle for cutting under real commercial greenhouse conditions. This study proposes an RGB-D perception framework combining YOLOv8n object detection with depth-based geometric estimation. The framework was evaluated on data collected in commercial greenhouses with dense, continuously varying occlusion, variable illumination and multiple co-visible fruits at different maturity stages. Detections were assigned to five operational classes: ripe, unripe, flagged (non-marketable due to overripeness or deformity), obstructed and peduncle. The detector achieved an overall mAP50 of 0.568, but appearance-based ripeness classification proved unreliable. Confusion concentrated along two boundaries: a continuous, judgment-based occlusion criterion and an absolute length threshold that RGB appearance cannot directly resolve. Depth-based length estimation, by contrast, achieved an MAE of 0.72 cm (RMSE = 1.15 cm, MAPE = 4.04%) directly in the greenhouse. Applying the ripeness length threshold to the estimated lengths misclassified 6.4% of fruits at the RIPE/UNRIPE boundary on the dimensional validation set, indicating that length-based ripeness assignment is feasible. Diameter estimation was markedly less accurate (R² = 0.49) and is not proposed as a validated measurement for grading or yield estimation. The framework also estimated candidate 3D cutting points that remained spatially consistent across frames; cutting accuracy requires future robotic validation. These results suggest that, in complex greenhouse environments, ripeness decisions of cucumber should rely on depth-based geometric measurement, with object detection used primarily for fruit and peduncle localization. Combining visual detection with quantitative measurement therefore appears to be a promising basis for robotic cucumber harvesting.