Scholarlink

<p>The abundance of farmed Holothurian provides valuable data to support Holothurian feeding and management in the farming process. However, the traditional counting method is manual, subjective, inefficient, and labor-intensive. This study proposes an in situ automatic counting framework for farmed Holothurian based on multi-object tracking method. This framework includes three parts: detecting, tracking, and counting. In the detection part, we developed YOLOv8s-BB to identify the presence of Holothurians. YOLOv8s-BB enhances YOLOv8s by integrating the BiFormer Convolutional Attention Module (BCAM) and BiFormer attention modules. In the tracking part, we enhanced the ByteTrack algorithm by replacing the aspect ratio of bounding box with its width in the Kalman filter (KF) state vector to track the detected Holothurians over time. The counting part is to tally the number of detected Holothurians. Experimental results indicate that the improved detector YOLOv8s-BB achieves a mean average precision (mAP) of 88.9%, a recall rate of 77.8%, and an F1 score of 84.2%, representing improvements of 4.5%, 6.3%, and 4.2%, respectively, compared to the original YOLOv8s model. The enhanced ByteTrack model achieves a High Order Tracking Accuracy (HOTA) of 67.63%, a Multi-Object Tracking Accuracy (MOTA) of 71.12%, an IDF1 score of 82.53%, and a frame rate of 44.82 frames per second (FPS). The counting model achieves an accuracy of 93.37%, with mean absolute error (MAE) and root mean square error (RMSE) values of 3.25 and 4.09, respectively. This study provides a real-time (44.82 FPS) abundance estimation method with 93.37% accuracy, which may assist in aquaculture management decisions.</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

Dynamic in situ counting of farmed Holothurians based on improved YOLOv8s-ByteTrack

  • 期刊论文

  • 发表日期：

    2025年6月24日

查看全文
  • Ke-Xin Liu,
  • Fang Wang,
  • Sheng-Cheng Hong,
  • Jun-Yi Wang,
  • Xin-Yu Zheng,
  • Yuan-Shan Lin

摘要

The abundance of farmed Holothurian provides valuable data to support Holothurian feeding and management in the farming process. However, the traditional counting method is manual, subjective, inefficient, and labor-intensive. This study proposes an in situ automatic counting framework for farmed Holothurian based on multi-object tracking method. This framework includes three parts: detecting, tracking, and counting. In the detection part, we developed YOLOv8s-BB to identify the presence of Holothurians. YOLOv8s-BB enhances YOLOv8s by integrating the BiFormer Convolutional Attention Module (BCAM) and BiFormer attention modules. In the tracking part, we enhanced the ByteTrack algorithm by replacing the aspect ratio of bounding box with its width in the Kalman filter (KF) state vector to track the detected Holothurians over time. The counting part is to tally the number of detected Holothurians. Experimental results indicate that the improved detector YOLOv8s-BB achieves a mean average precision (mAP) of 88.9%, a recall rate of 77.8%, and an F1 score of 84.2%, representing improvements of 4.5%, 6.3%, and 4.2%, respectively, compared to the original YOLOv8s model. The enhanced ByteTrack model achieves a High Order Tracking Accuracy (HOTA) of 67.63%, a Multi-Object Tracking Accuracy (MOTA) of 71.12%, an IDF1 score of 82.53%, and a frame rate of 44.82 frames per second (FPS). The counting model achieves an accuracy of 93.37%, with mean absolute error (MAE) and root mean square error (RMSE) values of 3.25 and 4.09, respectively. This study provides a real-time (44.82 FPS) abundance estimation method with 93.37% accuracy, which may assist in aquaculture management decisions.