Scholarlink

<p>Fruit bagging is an important practice for improving apple quality in structured orchards, whereas its automation remains challenging because immature apples are difficult to detect accurately and efficiently under complex field conditions. Existing methods for immature apple perception still suffer from limited keypoint detection accuracy, insufficient pose estimation precision, and high computational cost. To address these limitations, this study proposes YOLO-LAKP, a lightweight keypoint detection model based on the You Only Look Once (YOLO)v8n-Pose framework. The model improves feature representation and multi-scale fusion for immature apples by integrating contextual anchor attention (CAA), a high-level screening feature pyramid network (HS-FPN), and wavelet transform convolution (WTConv). Experimental results showed that YOLO-LAKP achieved a precision of 91.10%, a recall of 86.00%, and an mAP@0.5 of 88.50%, while requiring only 1.54&#xa0;million parameters, 5.8 GFLOPs, and 3.2&#xa0;MB of storage. Based on the detected keypoints, a depth-assisted method was further developed for 3D pose estimation of immature apples. Under different lighting conditions with minor occlusion, the proposed method achieved an average angular error of 7.43°. These results indicate that YOLO-LAKP is a practical lightweight method for immature apple keypoint detection and pose estimation, with potential for robotic fruit bagging applications.</p> Graphical abstract <p></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

Lightweight keypoint detection and pose estimation for immature apples based on YOLOv8n-pose

  • 期刊论文

  • 发表日期：

    2026年6月27日

查看全文
  • Kequan Feng,
  • Mustafa Mhamed

摘要

Fruit bagging is an important practice for improving apple quality in structured orchards, whereas its automation remains challenging because immature apples are difficult to detect accurately and efficiently under complex field conditions. Existing methods for immature apple perception still suffer from limited keypoint detection accuracy, insufficient pose estimation precision, and high computational cost. To address these limitations, this study proposes YOLO-LAKP, a lightweight keypoint detection model based on the You Only Look Once (YOLO)v8n-Pose framework. The model improves feature representation and multi-scale fusion for immature apples by integrating contextual anchor attention (CAA), a high-level screening feature pyramid network (HS-FPN), and wavelet transform convolution (WTConv). Experimental results showed that YOLO-LAKP achieved a precision of 91.10%, a recall of 86.00%, and an mAP@0.5 of 88.50%, while requiring only 1.54 million parameters, 5.8 GFLOPs, and 3.2 MB of storage. Based on the detected keypoints, a depth-assisted method was further developed for 3D pose estimation of immature apples. Under different lighting conditions with minor occlusion, the proposed method achieved an average angular error of 7.43°. These results indicate that YOLO-LAKP is a practical lightweight method for immature apple keypoint detection and pose estimation, with potential for robotic fruit bagging applications.

Graphical abstract