摘要： 准确高效监测鱼苗应激行为不仅有助于在养殖过程中调控应激源以减少产量损失，同时也可为育种阶段的鱼苗活力评估提供有效手段。针对鱼苗体积小、养殖密度高和高速非线性运动的特点，提出一种改进YOLOv8n-pose和BoTSort的鱼苗应激行为监测方法。改进YOLOv8n-pose作为检测器，将BMS模块与C2f模块相结合，使模型充分学习不同尺度特征；使用SPPCSPC模块替换原模型的特征融合模块，优化鱼苗相互遮挡情形下的检测精度；最后用N-EMASlideLoss替换原模型损失函数，增强模型的稳定性和对小目标的关注度。在跟踪器部分，基于检测器检测出的目标，结合BoTSORT多目标跟踪算法实现了更适合鱼苗应激时非线性运动监测的方法。最后，提取鱼苗的加速度、摆尾角度和聚集度三种特征进行加权融合，根据融合后的特征值判断鱼苗是否处于应激状态。结果显示，改进后的YOLOv8n-pose算法在目标检测和关键点检测的mAP比原模型分别提高了3.6%和4.5%；BoTSORT算法的MOTA为77.628%、MOTP为80.307%、IDF1为79.573%、IDSW为51，优于DeepSORT、ByteTrack、StrongSORT算法。该研究算法基于特征值的应激行为监测准确率为95.24%，为鱼类苗种培育中应激行为监测提供了新的思路和方法。

关键词: 鱼苗, YOLOv8n-pose, BoTSORT, 应激行为监测

Abstract: Accurately and efficiently monitoring the stress behavior of fish fry not only helps to regulate stressors during the breeding process to reduce yield losses, but also provides an effective means for evaluating the vitality of fish fry during the breeding stage. In view of the characteristics of fish fry, such as small size, high stocking density, and high - speed non - linear movement, this study proposes a method for monitoring the stress behavior of fish fry by improving YOLOv8n - pose and combining it with BoTSORT.The improved YOLOv8n - pose is used as a detector. The BMS module is combined with the C2f module to enable the model to fully learn features at different scales. The SPPCSPC module is used to replace the original feature fusion module of the model to optimize the detection accuracy in the case of fish fry occlusion. Finally, N - EMASlideLoss is used to replace the original loss function of the model, enhancing the model's stability and attention to small targets.In the tracker part, based on the targets detected by the detector, a method more suitable for monitoring the non - linear movement of fish fry under stress is achieved by combining the BoTSORT multi - target tracking algorithm.Finally, three features of fish fry, namely acceleration, tail - wagging angle, and aggregation degree, are extracted and weighted for fusion. Based on the fused feature values, it is determined whether the fish fry are under stress. The experimental results show that the mAP of the improved YOLOv8n - pose algorithm in target detection and key - point detection is 3.6% and 4.5% higher than that of the original model respectively. The MOTA of the BoTSORT algorithm is 77.628%, the MOTP is 80.307%, the IDF1 is 79.573%, and the IDSW is 51, which are superior to those of the DeepSORT, ByteTrack, and StrongSORT algorithms. The accuracy of the stress behavior monitoring of this study's algorithm based on feature values is 95.24%, providing new ideas and methods for stress behavior monitoring in fish fry breeding. 

Key words: fish fry, YOLOv8n-pose, BoTSORT;Stress behavior detection