Optimizing Indoor Farm Monitoring Efficiency Using UAV: Yield Estimation in a GNSS-Denied Cherry Tomato Greenhouse

TL;DR

The paper addresses efficient, non-destructive yield estimation in GNSS-denied greenhouse environments by deploying a compact UAV equipped with RGB-D, 3D LiDAR, and IMU. It presents a LiDAR-inertial odometry pipeline for robust navigation and a 3D multi-object tracking system to count and size cherry tomatoes, using YOLOv10 detections fused with depth to form 3D detections. On a harvesting row, the system achieves 94.4% counting accuracy and 87.5% weight estimation accuracy across a 13.2 m lane flown in 10.5 s, demonstrating the potential for rapid, scalable greenhouse monitoring. The study also investigates tracking unripened, densely occluded fruits, identifying occlusion and depth-visibility as key challenges and outlining directions for active perception and improved validation data in real-world settings.

Abstract

As the agricultural workforce declines and labor costs rise, robotic yield estimation has become increasingly important. While unmanned ground vehicles (UGVs) are commonly used for indoor farm monitoring, their deployment in greenhouses is often constrained by infrastructure limitations, sensor placement challenges, and operational inefficiencies. To address these issues, we develop a lightweight unmanned aerial vehicle (UAV) equipped with an RGB-D camera, a 3D LiDAR, and an IMU sensor. The UAV employs a LiDAR-inertial odometry algorithm for precise navigation in GNSS-denied environments and utilizes a 3D multi-object tracking algorithm to estimate the count and weight of cherry tomatoes. We evaluate the system using two dataset: one from a harvesting row and another from a growing row. In the harvesting-row dataset, the proposed system achieves 94.4\% counting accuracy and 87.5\% weight estimation accuracy within a 13.2-meter flight completed in 10.5 seconds. For the growing-row dataset, which consists of occluded unripened fruits, we qualitatively analyze tracking performance and highlight future research directions for improving perception in greenhouse with strong occlusions. Our findings demonstrate the potential of UAVs for efficient robotic yield estimation in commercial greenhouses.

Figures (6)

• Figure 1: UAV-based monitoring of GNSS-denied greenhouse environment. The proposed system, designed for yield estimation, achieves the accuracy of 94.4% for fruit counting and 87.5% for weight estimation, traveling 13.2 meters in just 10.5 seconds. Actual yield estimation computations are performed offline.
• Figure 2: Images to help understanding the indoor farm environment. (a) Validation lane for yield estimation. Blue line indicates a boundary between a harvesting row and a growing row. Pink allows shows the UAV's flight path. All ripened cherry tomatoes in the harvesting row are harvested after the flight. (b) Lengths of the smallest and the biggest cherry tomato among all harvested cherry tomatoes, which are 28mm and 45 mm respectively. (c) Top-view, perspective view, and side-view from the farm's 3D point cloud map. Note that the UAV's flight path is aligned with 3D map's x-axis.
• Figure 3: Configurations of the proposed UAV. With a flight speed of 2 m/s and a duration of 9 minutes, the UAV can cover a distance of over 1 km.
• Figure 4: Overview of the proposed 3D multi-object tracking framework. The system takes depth images, 2D bounding boxes from color images, and 6D camera poses as input. By fusing multi-sensor data, the 3D contexts of cherry tomatoes are estimated. Skyblue: Creation of 3D detections. Cherry tomatoes are approximated as cubes. Light peach: Association rule for tracking. Tracks are created or updated based on the similarity between existing tracks and new 3D detections. After association, the 3D contexts of cherry tomatoes are updated.
• Figure 5: Qualitative results on tracking ripened and unripened cherry tomatoes. All visible tracks are reprojected onto the images with their corresponding IDs. A total of four types of dataset were collected. The blue line represents the boundary between the growing row and the harvesting row, as shown in Fig. \ref{['fig:farm_analysis']}. The light green line indicates the minimum height threshold used to reject false positive tracks (e.g., IDs 105, 106, 107, 112, and 115 in the bottom-left case) during experiments in the harvesting row.