Visual-SLAM for the detection of hidden tomatoes in greenhouses by Hierarchical Localization and GLOMAPfor robotized harvesting

Analysis

Why This Paper Matters

Greenhouse crop monitoring and robotized harvesting are increasingly important for sustainable agriculture, but high-performance sensors like LiDAR and stereo cameras remain prohibitively expensive for many growers. This paper addresses that cost barrier by proposing a monocular Visual-SLAM system tailored for detecting hidden tomatoes—fruit that is occluded and inaccessible to classical vision technologies. By leveraging a single low-cost camera, the work opens the door to affordable, scalable robotic perception in controlled-environment agriculture.

The focus on occluded fruit is particularly significant. In dense tomato clusters, many ripe tomatoes are hidden behind leaves or other fruit, making them difficult to detect with conventional 2D vision. The proposed system not only identifies the cluster but also reconstructs a 3D model that characterizes the occluded tomato, which is a prerequisite for a robot to plan a successful harvest without damaging the plant.

Moreover, the integration of Hierarchical Localization with GLOMAP represents a practical marriage of robust place recognition and accurate Structure-From-Motion mapping. This combination is well-suited to the repetitive, self-similar textures of greenhouse environments, where naive feature matching often fails.

Technical Contributions

• Monocular Visual-SLAM pipeline: A cost-effective alternative to LiDAR or stereo setups, specifically designed for agricultural mapping of tomato crops.
• Hierarchical Localization integration: A coarse-to-fine strategy that first performs global retrieval to generate location hypotheses, then refines with local features within candidate regions.
• GLOMAP mapper: A Structure-From-Motion-based 3D reconstruction module that generates a mapped model of the crop.
• ROS 2 Humble node: A robot-mounted capture system that stores images for offline processing, enabling practical deployment on harvesting robots.
• Validation against ground truth: The reconstructed 3D model is checked against manual measurements of fruit size, centroid position, and orientation.

Results

The abstract reports that the system correctly identified the tomato cluster and successfully characterized the occluded tomato that classical vision technologies cannot access. The 3D reconstruction was validated against manual ground-truth measurements of fruit size, centroid position, and orientation, confirming the geometric accuracy of the low-cost monocular pipeline. However, no quantitative metrics (e.g., error in millimeters, success rate, or comparison to baselines) are provided in the abstract. The tests were carried out on a real tomato bunch in the Agroconnect experimental greenhouse, demonstrating real-world feasibility.

Significance

This work demonstrates that a monocular camera combined with modern SLAM and localization techniques can achieve geometric accuracy sufficient for agricultural tasks like harvesting. By avoiding expensive sensors, it lowers the barrier for adoption in commercial greenhouses. The initial mapping also serves as a foundation for future algorithms to analyze growth patterns and optimize agricultural management. As robotics and AI continue to penetrate agriculture, such low-cost, high-accuracy perception systems will be critical for sustainable food production and labor reduction.