YS3AM: Adaptive 3D Reconstruction and Harvesting Target Detection for Clustered Green Asparagus

Abstract

Green asparagus grows in clusters, which can cause overlaps with weeds and immature stems, making it difficult to identify suitable harvest targets and cutting points. Extracting precise stem details in complex spatial arrangements is a challenge. This paper explored the YS3AM (Yolo-SAM-3D-Adaptive-Modeling) method for detecting green asparagus and performing 3D adaptive-section modeling using a depth camera, which could benefit harvesting path planning for selective harvesting robots. Firstly, the model was developed and deployed to extract bounding boxes for individual asparagus stems within clusters. Secondly, the stems inside these bounding boxes were segmented, and binary masks were generated. Thirdly, high-quality depth images were obtained through pixel block completion. Finally, a novel 3D reconstruction method, based on adaptive section modeling and combining the mask and depth data, is proposed. And an evaluation method is introduced to assess modeling accuracy. Experimental validation showed high-performance detection (1095 field images demonstrated, Precision: 98.75%, Recall: 95.46%, F1: 0.97) and robust 3D modeling (103 asparagus stems, average RMSE: length 0.74, depth: 1.105) under varying illumination conditions. The system achieved 22 ms per stem processing speed, enabling real-time operation. The results demonstrated that the 3D model accurately represents the spatial distribution of clustered green asparagus, enabling precise identification of harvest targets and cutting points. This model provided essential spatial pathways for end-effector path planning, thereby fulfilling the operational requirements for efficient green asparagus harvesting robots.

Code and data availability

The paper describes a custom RGB-D asparagus image dataset (730 original images, augmented to 4380) and the YS3AM pipeline, but no public deposit of the dataset, images, code, or trained models is stated anywhere in the supplied blocks. The only URL given (LabelImg) is a generic third-party annotation tool, not a paper

No evidence-backed public reproduction asset is currently recorded.