Scholarlink

<p>Accurate detection and localization of tea buds remain challenging due to their dense distribution, small size, and significant nonlinear illumination variations in field environments. To address these issues, this study proposes a novel tea bud detection and localization method combining an improved YOLOv8m architecture with point cloud cluster analysis. The proposed method first enhances the original YOLOv8m by incorporating a channel attention mechanism into its backbone to emphasize informative feature channels. Furthermore, a bidirectional feature pyramid network is employed as the neck network to enhance both semantic and detailed features across multi-scale feature maps, thereby improving small target detection performance. Subsequently, a clustering algorithm is applied to segment pixels within each detection bounding box into non-overlapping clusters. The cluster nearest to the camera is identified as the target tea bud, and its minimal enclosing cuboid is computed. The three-dimensional position of the tea bud is determined by the centroid of the cuboid’s bottom surface. Experimental results demonstrated that the improved YOLOv8m achieves significant performance gains, with precision, recall, and mean average precision at 50% intersection over union reaching 67.1%, 65.1%, and 69.5% respectively, representing improvements of 10.7%, 3.1%, and 5.5% over the baseline; the localization errors in the x, y, and z directions are 3.4 ± 6.9&#xa0;mm, 0.8 ± 5.7&#xa0;mm and 0.3 ± 4.5&#xa0;mm, respectively. These findings indicate that the proposed method satisfies the precision requirements for automated tea-picking robotic systems.</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

Tea bud detection and localization based on RGB-D image analysis

  • 期刊论文

  • 发表日期：

    2025年11月10日

查看全文
  • Minglong Wang,
  • Xuxi Huang,
  • Wenyong Zeng,
  • Ruihan Yin,
  • Wenping Fu,
  • Lixue Zhu,
  • Guichao Lin

摘要

Accurate detection and localization of tea buds remain challenging due to their dense distribution, small size, and significant nonlinear illumination variations in field environments. To address these issues, this study proposes a novel tea bud detection and localization method combining an improved YOLOv8m architecture with point cloud cluster analysis. The proposed method first enhances the original YOLOv8m by incorporating a channel attention mechanism into its backbone to emphasize informative feature channels. Furthermore, a bidirectional feature pyramid network is employed as the neck network to enhance both semantic and detailed features across multi-scale feature maps, thereby improving small target detection performance. Subsequently, a clustering algorithm is applied to segment pixels within each detection bounding box into non-overlapping clusters. The cluster nearest to the camera is identified as the target tea bud, and its minimal enclosing cuboid is computed. The three-dimensional position of the tea bud is determined by the centroid of the cuboid’s bottom surface. Experimental results demonstrated that the improved YOLOv8m achieves significant performance gains, with precision, recall, and mean average precision at 50% intersection over union reaching 67.1%, 65.1%, and 69.5% respectively, representing improvements of 10.7%, 3.1%, and 5.5% over the baseline; the localization errors in the x, y, and z directions are 3.4 ± 6.9 mm, 0.8 ± 5.7 mm and 0.3 ± 4.5 mm, respectively. These findings indicate that the proposed method satisfies the precision requirements for automated tea-picking robotic systems.