| 摘 要: 在智能水产养殖中,如何精准、无创地评估鱼类应激状态仍是一个关键挑战。传统应激评估方法多依赖侵入性采样或主观行为判读,难以实现连续、客观的动态监测。为解决这一问题,本研究提出一种基于三维行为特征方法,用于动态评估罗非鱼在梯度空气暴露应激下的响应强度。该方法融合双目视觉系统与YOLOv11多关键点检测模型,通过视频采集与三维轨迹重建,系统提取8项运动行为参数；经分析筛选出5项对应急强度呈极显著单调变化的敏感指标(p < 0.001)。进一步采用主成分分析(PCA)降维,构建以“趋深贴底–活动抑制”为核心内涵的行为综合应激指数(Behavioral Stress Index, BSI),其第一主成分方差贡献率达86.5%。基于BSI均值中点确立四级应激判别标准,并训练线性判别分析(LDA)分类模型,在留一法交叉验证下准确率达91.7%(Kappa = 0.89)。结果表明,所提方法能够实现鱼类应激状态的连续量化、自动分级与无损监测,为智能水产养殖中的动物福利评估提供了可靠的技术路径。 |
| 关键词: 双目视觉 YOLOv11 关键点检测 三维行为分析 应激评估 |
|
中图分类号:
文献标识码:
|
| 基金项目: 国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| Acquisition method of 3D motion parameters of fish based on YOLOv8 and binocular visionWU Xin1, Wang Lei2, Xie Yufeng1, Fan Shengli3, CAI Weiming3 |
|
wuxin, wanglei, xieyufeng, fanshengli, caiweiming
|
Zhejiang Sci-Tech University
|
| Abstract: In intelligent aquaculture, accurately and non-invasively assessing fish stress status remains a critical challenge. Traditional stress evaluation methods primarily rely on invasive sampling or subjective behavioral interpretation, making continuous and objective dynamic monitoring difficult to achieve. To address this issue, this study proposes a three-dimensional behavioral feature-based approach for dynamically evaluating the stress response intensity of tilapia under graded air-exposure stress. The method integrates a binocular vision system with a YOLOv11 multi-keypoint detection model to capture video data and reconstruct 3D trajectories, systematically extracting eight motion-related behavioral parameters. Statistical analysis identified five of these parameters as highly sensitive indicators that exhibit extremely significant monotonic changes with increasing stress intensity (p < 0.?001). Subsequently, principal component analysis (PCA) was employed for dimensionality reduction, leading to the development of a Behavioral Stress Index (BSI) centered on the core behavioral pattern of “depth-seeking and bottom-hugging with activity suppression.” The first principal component accounted for 86.5% of the total variance. Using the median BSI value as the threshold, a four-level stress classification criterion was established, and a Linear Discriminant Analysis (LDA) classification model was trained. Under leave-one-out cross-validation, the model achieved an accuracy of 91.7% (Kappa = 0.89). These results demonstrate that the proposed method enables continuous quantification, automated grading, and non-invasive monitoring of fish stress status, offering a reliable technical pathway for animal welfare assessment in intelligent aquaculture systems. |
| Keywords: Binocular vision YOLOv11 Keypoint detection 3D behavioral analysis Stress assessment |
