Vision-based Fruit Recognition and Three-dimensional Localization for Tomato Harvesting Robots: A Review
Zongxi Xu *
North China University of Water Resources and Electric Power, Zhengzhou 450045, China.
*Author to whom correspondence should be addressed.
Abstract
Tomato harvesting is a labour-intensive, costly step in agricultural production, and machine-vision-based harvesting robots have become an important approach to automated harvesting. This review systematically synthesises advances in fruit recognition and localisation for tomato-harvesting robots recent advances along the technical chain of recognition, three-dimensional (3D) localisation, and system integration. For recognition, traditional colour-threshold methods are strongly affected by illumination; deep-learning methods represented by the YOLO family have reported markedly higher accuracy and speed than traditional methods on their respective datasets and have further evolved toward instance segmentation and picking-keypoint detection, although high accuracy, lightweight design, and occlusion robustness are difficult to achieve simultaneously. For localisation, binocular stereo vision is cost-effective but sensitive to lighting and texture; RGB-D depth cameras provide reliable depth and are easy to integrate but recover 3D coordinates less reliably under occlusion; multi-view and point-cloud fusion are robust but computationally complex. For system integration, hand-eye calibration links visual coordinates to arm coordinates, and reported end-to-end greenhouse trials have reached harvesting success of roughly 80-87% alongside ~85% reachability in dense scenes, yet they remain limited by long cycle times, difficult damage control, and limited commercial adoption. On this basis, this review identifies three key research gaps: robust perception under occlusion, dataset scale and cross-region generalisation, and end-to-end integration of recognition, localisation, and harvesting; future directions are outlined to guide the development of tomato harvesting robots. Commercial adoption, however, also depends on economics: the reviewed studies rarely report cost per harvested fruit, payback period or comparison with manual picking, so the approaches therefore cannot yet be ranked by cost.
Keywords: Tomato harvesting robot, fruit recognition, three-dimensional localisation, deep learning, visual servoing