Scholarlink

<p>During automatic fruit and vegetable harvesting, interference from plant stems and branches can significantly reduce the success rate and operational efficiency of harvesting robots. Therefore, accurate three-dimensional reconstruction of plant obstacles is essential for effective robotic harvesting. This paper proposes an integrated framework for tomato plant reconstruction based on multi-view point cloud acquisition and fusion combined with deep learning segmentation. Point clouds captured from primary and auxiliary viewpoints are aligned through coordinate transformation to obtain a more complete representation of the plant structure. After preprocessing using voxel filtering, statistical filtering, and radius filtering, a PointNet++ neural network is employed to separate tomato fruits from stems and leaves. Experimental results show that the proposed method achieves a stable Intersection over Union (IoU) of approximately 0.93 on the validation set. An octree-based tomato plant model is then established and deployed in the Robot Operating System (ROS). This provides an effective basis for obstacle avoidance and motion planning of harvesting robots.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-view Point‑Cloud 3D Reconstruction and Deep Segmentation of Tomato Plants for Harvesting Robot Obstacle Avoidance

  • 期刊论文

  • 发表日期：

    2026年7月22日

查看全文
  • Jiantao Zhang,
  • Qixin Chen,
  • Binliang Zhai,
  • Yijian Liu

摘要

During automatic fruit and vegetable harvesting, interference from plant stems and branches can significantly reduce the success rate and operational efficiency of harvesting robots. Therefore, accurate three-dimensional reconstruction of plant obstacles is essential for effective robotic harvesting. This paper proposes an integrated framework for tomato plant reconstruction based on multi-view point cloud acquisition and fusion combined with deep learning segmentation. Point clouds captured from primary and auxiliary viewpoints are aligned through coordinate transformation to obtain a more complete representation of the plant structure. After preprocessing using voxel filtering, statistical filtering, and radius filtering, a PointNet++ neural network is employed to separate tomato fruits from stems and leaves. Experimental results show that the proposed method achieves a stable Intersection over Union (IoU) of approximately 0.93 on the validation set. An octree-based tomato plant model is then established and deployed in the Robot Operating System (ROS). This provides an effective basis for obstacle avoidance and motion planning of harvesting robots.