基于姿态传感器的肉牛行为特征识别研究

发布时间：2018-05-03 17:14

本文选题：肉牛行为 + 姿态传感器　；参考：《沈阳农业大学》2017年硕士论文

【摘要】：随着大规模、集约化的肉牛养殖业的发展,动物行为识别的机器化实现表现得尤其重要。现在对肉牛行为特征的观察主要依赖于饲养员的实际经验,即借助自己对牛的了解程度,对肉牛的健康及行为特征作出判断,这不仅加重饲养员的负担,而且工作效率低下,对肉牛个体行为状态的改变不够快速,不能很好的适应大规模农场饲养的需求。最近几年,信息化技术不断发展,对肉牛精准饲养提供了技术支持,各国相关工作者借助图像处理技术对肉牛的行为状态进行观察与监测,但这种技术对环境的要求相对较高。本文以姿态传感器为工具,以育肥期肉牛为主体,将肉牛的行为数据通过主成分分析进行降维,采用卡尔曼滤波器算法、BP神经网络算法对肉牛行为特征数据建立分类模型。主要内容包括:(1)本文有针对性的对肉牛的姿态传感器数据进行采集并分析,通过主成分分析算法将肉牛9个行为特征数据集进行降维,得到的降维数据既可以代表肉牛的所有行为特征,在建模计算中又可以减少计算量,增加测量精度。(2)采用卡尔曼滤波器算法对肉牛的行为特征数据集展开迭代运算,探讨其与分类模型之间的关系及影响程度,这也是采用卡尔曼滤波器在肉牛行为分类方面的初次尝试。结合肉牛行为活动发生的时间,对肉牛的日常行为特征对应的姿态数据模型展开研究与分析,经过卡尔曼滤波器得到的肉牛行为特征分类模型识别率为62.5%,识别率偏低,说明线性系统在肉牛行为识别领域不能作为理论支持,对肉牛行为分类的研究需要在非线性系统中进行进一步的研究。(3)由于采用卡尔曼滤波器对肉牛行为分类模型的识别率过低,不足以作为肉牛行为分类的理论依据,本文还采用BP神经网络算法对肉牛行为分类进行研究,结果表示:BP神经网络算法在肉牛行为特征分类方面变现比卡尔曼滤波器要强很多,分类精度在卡尔曼滤波器算法的基础上提高了近30%,这一结果表明,肉牛行为特征分类模型的准确度在采用非线性系统计算时取得较大的提升,该方法在肉牛行为识别领域可以作为理论支持。(4)本文还对传感器在肉牛身上部署的位置和记录时间间隔进行了相应的研究,研究表明:传感器安放在肉牛的颞骨位置时,传感器节点所测得的数据最能符合肉牛的实际行为活动;在不影响肉牛行为特征分类精度的情况下,传感器记录时间间隔为2s时可以减少计算量,加快数据处理的速度。
[Abstract]:With the development of large-scale and intensive beef farming, the realization of machine-based recognition of animal behavior is particularly important. Now the observation of beef cattle behavior characteristics mainly depends on the actual experience of the keeper, that is, with the help of one's own understanding of the cattle, to judge the beef cattle's health and behavior characteristics, which not only increases the burden of the keeper, but also the work efficiency is low. The change of individual behavior of beef cattle is not fast enough to meet the needs of large-scale farms. In recent years, information technology has been developing, which provides technical support for precision feeding of beef cattle. Relevant workers in various countries use image processing technology to observe and monitor the behavior of beef cattle, but this technology requires relatively high environmental requirements. In this paper, the behavior data of beef cattle are reduced by principal component analysis (PCA) with attitude sensor as tool and beef cattle in fattening period as main body. Kalman filter algorithm and BP neural network algorithm are used to establish classification model of beef cattle behavior characteristic data. The main contents include: (1) this paper collects and analyzes the pose sensor data of beef cattle, and reduces the dimension of nine behavioral data sets of beef cattle by principal component analysis (PCA) algorithm. The reduced dimension data can not only represent all the behavioral characteristics of beef cattle, but also reduce the computational cost and increase the measurement accuracy in the modeling calculation. (2) the Kalman filter algorithm is used to iterate the behavioral feature data set of beef cattle. The relationship between the model and the classification model and the degree of influence are discussed. This is also the first attempt to use Kalman filter in beef cattle behavior classification. Combined with the time of beef cattle behavior, the attitude data model corresponding to the daily behavior characteristics of beef cattle was studied and analyzed. The recognition rate of the classification model of beef behavior characteristics obtained by Kalman filter was 62.5 and the recognition rate was on the low side. It shows that the linear system can not be used as theoretical support in beef cattle behavior recognition. The study of beef behavior classification needs further study in nonlinear system. (3) because the recognition rate of beef behavior classification model using Kalman filter is too low, it can not be used as the theoretical basis of beef behavior classification. This paper also uses BP neural network algorithm to study beef cattle behavior classification, the results show that the performance classification of beef cattle behavior is much stronger than Kalman filter. The classification accuracy is improved by nearly 30 percent on the basis of Kalman filter algorithm. The results show that the accuracy of beef behavior feature classification model is greatly improved by using nonlinear system calculation. This method can be used as a theoretical support in the field of Behavioral recognition of Beef cattle. (4) in this paper, the location and recording time interval of the sensor deployed on beef cattle are also studied. The results show that the sensor is placed in the temporal bone position of beef cattle. The measured data of sensor nodes can best accord with the actual behavior of beef cattle. Without affecting the classification accuracy of behavioral characteristics of beef cattle, the sensor recording time interval of 2s can reduce the calculation amount and speed up the data processing.
【学位授予单位】：沈阳农业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S823;TP212

【相似文献】