Real-time Detection and Characterization of Trunks and Upright Branches of Pear Trees for Automatic Dormant Pruning

Abstract

Abstract Purpose Dormant pruning is critical for fruit tree management and maintaining fruit quality. Traditional manual pruning is labor-intensive, driving interest in automated robotic dormant pruning. However, automated robotic dormant pruning meets significant challenges in trunk and branches detection, due to the complexity of the orchard environment and the interlacing of the branches. This paper proposes an automatic method for real-time detection and pruning of pear tree trunks and upright branches using an RGB-D camera. Methods Pear trunk detection was conducted by enhancing the You Only Look Once version 5 Nano (YOLOv5n) model with Squeeze-and-Excitation Networks (SENet) and optimizing the anchor boxes. For branch segmentation, YOLOv8n-seg integrated with Dynamic Snake Convolution (DSConv) and Focal Scale Intersection over Union (Focal_SIoU) was employed. The length and angle of the branches were calculated from the generated mask images, and the pruning position was determined. A PRUNING_ROS package was developed for real-time orchard applications. Results The evaluation demonstrated 96.7% mean average precision (mAP) for trunk detection and 82.6% mAP for branch segmentation in test dataset. Field test results showed that the mean absolute error (MAE) of trunk distance localization compared to manual measurements was 3.71 cm, with the root mean square error (RMSE) of 3.84 cm, and the frames per second (FPS) of 31.6. The MAE was 2.3 cm (RMSE: 2.6 cm) for pruning points in depth direction and 2.34° (RMSE: 2.71°) for upright branches angle, with the FPS of 36.2. The field pruning experiment showed a pruning success rate of 47.6%. Conclusion This method provides technical support for the operation of fruit tree pruning robots, representing a step toward the full automation of fruit tree management.

Plant phenotyping relevance

RGB-D画像からナシ樹の幹・枝を検出・分割し、枝の長さと角度を算出する手法が中心で、単なる対象位置検出を超えた植物器官形態の計測と技術評価を行っている。

Code and data availability

The paper's phenotype data (trunk and upright branch RGB-D images) are available only on request, and the PRUNING_ROS analysis package is promised for GitHub only after acceptance, with no authors' public URL yet. The URLs present in the text (librealsense, LabelImg, Labelme, scikit-image) are generic third-party tools

No evidence-backed public reproduction asset is currently recorded.