Scholarlink

<p>To address the challenge of non-invasive and dynamic monitoring of fish body size in deep-sea aquaculture, this paper proposes a three-dimensional fish body measurement method based on stereo vision. A dataset of seven categories of fish keypoints was constructed, and high-precision keypoint detection in complex and dynamic underwater environments was achieved by integrating the You Only Look Once version 8n with pose estimation (YOLOv8n-pose) model. By combining stereo camera calibration with the Semi-Global Block Matching (SGBM) algorithm, a 3D coordinate fitting model was developed to estimate fish body length and height. Experimental results show that the dataset used for constructing the 3D non-intrusive fish detection model passed the normality test for body length (<i>p</i> = 0.027), and the 95% limits of agreement in the Bland–Altman analysis ranged from −2.1 to 3.5&#xa0;cm. For height, the <i>p</i>-value was 0.104, and the 95% difference was within ± 1.2&#xa0;cm, indicating that the method meets the precision requirements of aquaculture monitoring. The proposed method achieved a mAP@50 of 98.8%, representing a 13.2% improvement over the traditional High-Resolution Network (HRNet) model. The average relative errors in body length and height measurements were 12.76% and 11.64%, respectively, demonstrating the effectiveness and applicability of the method in intelligent aquaculture.</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

An integrated method for non-intrusive underwater fish measurement based on keypoint detection and stereo vision

  • 期刊论文

  • 发表日期：

    2025年10月24日

查看全文
  • Hua Zhang,
  • Yue Guo,
  • Yonghe Xie,
  • Zhi Zheng

摘要

To address the challenge of non-invasive and dynamic monitoring of fish body size in deep-sea aquaculture, this paper proposes a three-dimensional fish body measurement method based on stereo vision. A dataset of seven categories of fish keypoints was constructed, and high-precision keypoint detection in complex and dynamic underwater environments was achieved by integrating the You Only Look Once version 8n with pose estimation (YOLOv8n-pose) model. By combining stereo camera calibration with the Semi-Global Block Matching (SGBM) algorithm, a 3D coordinate fitting model was developed to estimate fish body length and height. Experimental results show that the dataset used for constructing the 3D non-intrusive fish detection model passed the normality test for body length (p = 0.027), and the 95% limits of agreement in the Bland–Altman analysis ranged from −2.1 to 3.5 cm. For height, the p-value was 0.104, and the 95% difference was within ± 1.2 cm, indicating that the method meets the precision requirements of aquaculture monitoring. The proposed method achieved a mAP@50 of 98.8%, representing a 13.2% improvement over the traditional High-Resolution Network (HRNet) model. The average relative errors in body length and height measurements were 12.76% and 11.64%, respectively, demonstrating the effectiveness and applicability of the method in intelligent aquaculture.