【摘要】 水产养殖中，鱼类跟踪是实现鱼类行为监测、水质异常报警、鱼类生长状况评估的基础，但现有方法存在计算耗时长、模型占用空间大、在边缘端设备部署困难等问题。针对上述问题，本文以石斑鱼为研究对象，提出一种基于轻量化YOLO v8与BoT-SORT的石斑鱼跟踪方法，该方法包括目标检测和目标跟踪两个阶段。在目标检测中，采用YOLO v8m作为基线网络，引入卷积模块FasterConv以减少参数量；加入EMA(Excitation and modulation attention)机制以保持模型精度；使用多尺度特征融合模块Fusion并调整Neck网络结构以提高模型的特征融合能力。在目标跟踪部分，BoT-SORT算法简化了鱼体的运动状态变量，加入相机运动补偿(Camera motion compensation, CMC)以应对鱼体外观剧烈变化，最后利用ResNeST50网络提取较高置信度检测框内鱼体的外观特征，实现了鱼体跟踪。在自建的石斑鱼数据集上进行了训练和验证，目标检测模型mAP@0.5为95.80%;其模型内存占用量为23.7 MB,相较原始YOLO v8m模型降低54.42%;将本文的轻量化目标检测模型应用到BoT-SORT算法，MOTA为78.774%,FPS达到28.20 f/s,在对比实验中综合性能大幅超过SORT、DeepMoT等算法。本方法可以实现石斑鱼的检测与跟踪，为石斑鱼的养殖提供技术支撑。更多还原

【Abstract】 In aquaculture, fish tracking is fundamental for monitoring fish behavior, detecting water quality anomalies, and assessing the growth conditions of fish. However, existing methods suffer from issues such as computational time consumption, large model size, and difficulties in deploying on edge devices. To address these challenges, taking grouper fish as the research subject, a tracking method was proposed based on a lightweight YOLO v8 and BoT-SORT. This method consisted of two stages: target detection and target tracking. In the target detection phase, YOLO v8m was used as the baseline network. Firstly, a convolutional module, FasterConv, was introduced to reduce the number of parameters. Then the excitation and modulation attention(EMA) mechanism was incorporated to maintain model accuracy. Finally, a multi-scale feature fusion module, Fusion, was employed and the structure of the Neck network was adjusted to enhance the model’s capability for feature integration. For the target tracking part, the BoT-SORT algorithm simplified the motion state variables of the fish body and included camera motion compensation(CMC) to deal with drastic changes in fish appearance. Subsequently, the ResNeST50 network was utilized to extract the appearance features of the fish within high-confidence detection bounding boxes, achieving fish tracking. The method was trained and validated on a self-built grouper dataset, achieving a mAP@0.5 of 95.80% for the object detection model, with a model size of 23.7 MB—a 54.42% reduction compared to that of the original YOLO v8m model. When applying the lightweight object detection model to the BoT-SORT algorithm, the MOTA reached 78.774%, with an FPS of 28.20 f/s, significantly outperforming SORT, DeepMoT, and other algorithms in comparative experiments. The results demonstrated that this method can achieve detection and tracking of grouper fish, providing technical support for the cultivation of groupers.更多还原