Scholarlink

<p>Accurate mango yield estimation is crucial for efficient resource management and harvest planning. This study presents a&#xa0;novel framework utilizing a&#xa0;modified version of MangoYOLO5, a&#xa0;custom YOLOv5 model for mango detection proposed in our previous work. It includes an attentive feature fusion mechanism in the backbone, a&#xa0;convolutional block attention module, dense connections in the neck, and an additional <i>10</i> <i>×</i> <i>10</i> detection layer to enhance detection precision and reduce false positives. Furthermore, we integrate the proposed detection model with an enhanced centroid tracking algorithm that employs multi-frame references for occlusion handling and dynamic threshold adjustments to achieve consistent tracking and accurate fruit counting. The system’s effectiveness is validated across varying light conditions and different trees, demonstrating close alignment with ground truth data from manual labeling, achieving a&#xa0;mean average precision (mAP) of <i>93.6%</i>. This work not only advances the state-of-the-art in mango detection and yield estimation but also provides insights into developing robust computer vision solutions for agricultural applications.</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

Enhanced YOLO-Based Mango Detection and Centroid Tracking for Precise Yield Estimation

  • 期刊论文

  • 发表日期：

    2025年6月6日

查看全文
  • Hari Chandana Pichhika,
  • Priyambada Subudhi,
  • Raja Vara Prasad Yerra

摘要

Accurate mango yield estimation is crucial for efficient resource management and harvest planning. This study presents a novel framework utilizing a modified version of MangoYOLO5, a custom YOLOv5 model for mango detection proposed in our previous work. It includes an attentive feature fusion mechanism in the backbone, a convolutional block attention module, dense connections in the neck, and an additional 10 × 10 detection layer to enhance detection precision and reduce false positives. Furthermore, we integrate the proposed detection model with an enhanced centroid tracking algorithm that employs multi-frame references for occlusion handling and dynamic threshold adjustments to achieve consistent tracking and accurate fruit counting. The system’s effectiveness is validated across varying light conditions and different trees, demonstrating close alignment with ground truth data from manual labeling, achieving a mean average precision (mAP) of 93.6%. This work not only advances the state-of-the-art in mango detection and yield estimation but also provides insights into developing robust computer vision solutions for agricultural applications.