大型动物行为模型与高级行为智能视频感知新方法研究

It is difficult to perceive the high-level behavior of large animals (dairy cattle, et al.) for precision livestock farming, predicting diseases and experiments on animal behavior. In this project, the co-relationship of basic movement, complex movement and high-level behavior and dairy cattle will be studied as research basis and object respectively, and animal ethology, the theory of hierarchical structure and video analysis will be integrated to research hierarchical structure based behavior model of large animals and the methods to perceive animal behavior. With theoretical research, mathematical modeling, computer simulation and experiments on observing behavior, we will analyze the mechanism of animal behavior to build hierarchical structure based behavior model. By studying the methods to detect and track moving object with interaction between environments in complex background, new models will be built to recognize basic motion of large animals. Based on the hierarchical model of behavior, a method will be proposed to perceive the high-level behavior. Accordingly, we will present new theories and methods with the fusion of different levels of behavior information to perceive and understand high-level behavior of large animals. This project provides new theoretical and methodological basis for recognition and understanding high-level behavior of large animals, theoretic foundation for precision livestock farming and forecasting health of large animals and new ideas and ways for experiments on behavior of large animals with great researching significances and prosperous applications.

大型动物如奶牛的精准养殖、疫病预测及行为实验面临如何高效准确感知与识别高级行为这一关键科学问题。本项目将大型动物简单运动、复杂运动和高级行为的互作用关系作为研究切入点，以大型动物奶牛为研究载体，运用动物行为学、分层结构理论与视频分析方法相结合，研究并建立分层结构的大型动物行为新模型和感知方法。通过理论分析、数学建模与计算机仿真和行为观察实验，深入分析大型动物行为机理，构建分层结构的大型动物行为模型；研究复杂背景及动物与环境交互作用下的目标检测与跟踪新方法，建立大型动物基本运动识别模型，提出基于分层行为模型的高级行为感知方法，从而建立不同层次行为信息融合的大型动物高级行为感知与理解新理论和新方法，为大型动物的高级行为识别与理解提供新的理论和方法基础，为大型动物的精准养殖和健康预测提供理论依据，同时为大型动物行为实验研究提供新思路、探索新途径，具有重要的研究意义和广阔的应用前景。

对动物个体信息和行为进行智能感知与分析是实现精准畜牧业的核心。本项目以大型动物奶牛为研究对象，采用视频分析技术和人工智能技术，研究复杂养殖环境下目标检测、个体识别、体况信息获取、简单与复杂行为检测，及高级行为检测与识别方法，在以下方面取得了理论和方法上的突破：（1）提出动态背景减去法和改进的GMM目标检测方法。动态背景减去法检测准确率为88.34%，改进GMM的前景和背景误检率分别比经典GMM降低了19.50%和13.37%，可有效检测复杂环境下的奶牛；（2）提出基于深度学习的牛只个体识别方法，牛只个体识别准确率为90.55%，对光线变化、前景遮挡均有较高的鲁棒性，能够对奶牛个体进行较高精度识别；（3）提出基于图像深度信息的奶牛身体区域精细分割和体尺测量方法，奶牛各身体区域平均识别率为95.15%，体尺参数平均相对误差为3.07%，为奶牛体况获取与评价提供了技术和理论基础；（4）提出基于视频分析的多目标奶牛反刍行为监测方法，基于相关核滤波算法的奶牛反刍行为监测方法平均误检率为7.25%，且可在夜间跟踪多目标，适合全天候监测；（5）提出基于决策树的跛行检测和基于头颈部轮廓拟合直线斜率特征的跛行检测方法。根据四肢运动信息的决策森林分类的识别总体精度为90.18%，基于头颈部斜率数据集的KNN分类算法对正常、轻度和中重度跛行平均检测准确率达93.89%，解决了奶牛跛行程度的非接触识别问题；（6）提出基于视频分析的犊牛基本行为识别和与场景交互行为的识别方法，犊牛躺、站、走和跑跳的正确识别率分别为100%、96.17%、95.85%和97.26%，交互行为中进入休息区、离开休息区、采食、饮水行为的识别率是：94.38%，92.86%， 79.69%，81.73%；（7）开发了基于Storm的奶牛发情实时监测系统，奶牛发情预测准确率为85.71%；研发了奶牛体温植入式传感器及实时监测系统，体温检测误差在±0.05℃内。项目成果为大型动物高级行为识别及理解提供了方法和技术。发表论文17篇；其中，SCI论文4篇，EI论文13篇。授权专利3项，授权软著15项。