基于改进YOLOv8n—ByteTrack模型的海参原位计数方法

打开文本图片集

中图分类号：TP391.4；S432 文献标识码：A 文章编号：2095-5553（2026）03-0089-08

Abstract：Aiming at theproblemsof loweficiencyand high cost inmanualcounting for holothurian farming，a dynamic in-situcounting methodbasedonanimprovedYOLOv8nandByteTrack isproposed.Thegoalis toachievehigh precisionautomaticcounting in complexunderwater environments.Firstly，GhostConv and GhostC3 are used to replace the Convand C2flayers in YOLOv8n toreduce thefloating-point computational loadand parameter countof the model.At the same time，a Focal Modulation（FM） module is usedto replace the YOLOv8n’s Fast Spatial Pyramid Pooling （SPPF）module，and the Squeeze-and-Excitation（SE）atention mechanism isintroduced to enhance therecognition accuracyof holothurians.Secondly，the Kalman filtering model in ByteTrack tracking is improved to enhance tracking performance.Experimental results show that the improved model achieves an average detection accuracy，precision rate，and recall rate of 88.90% 93.20% ，and 81.20% ，respectively，for detection tasks. In multi-object tracking task， the higher-order tracking accuracy，multi-object tracking accuracy，IDFl score，and frame rate are improved to 69.95% 72.68% 84.43% ，and 49.84 fps，respectively. In the test of real-world farming scenarios，the model's counting accuracy reaches 95.58% ，with a mean absolute error of 2.OO and a root mean square error of only 2.55. This research method enables fasterand more accuratecounting of holothurian quantities in videos，providing a scientific basis for decision-making in the intelligent management of holothurian farming.

Keywords：holothurian counting；in-situ counting；unmanned ship；object detection；multi-object tracking

0 引言

养殖海参原位计数为投饵量、海参产量估计和销售价格等提供重要的决策支持。（剩余10964字）