基于时序运动特征的奶牛爬跨行为识别研究  

Research on recognition of crawling behavior of cows based on temporal motion features

作　　者：谢忠红[1] 刘悦怡 宋子阳 徐焕良[1] XIE Zhonghong;LIU Yueyi;SONG Ziyang;XU Huanliang(College of Artificial Intelligence,Nanjing Agricultural University,Nanjing 210095,China)

机构地区：[1]南京农业大学人工智能学院,江苏南京210095

出　　处：《南京农业大学学报》2021年第1期194-200,共7页Journal of Nanjing Agricultural University

基　　金：国家自然科学基金项目(31601545);中央高校基本业务费专项资金(KYZ201670)。

摘　　要：[目的]为了尽早发现发情奶牛并及时进行配种,提高养殖场的经济效益,本文研究了基于机器视觉的奶牛爬跨行为识别。[方法]选取3种行为视频:侧爬跨101段,追随191段,行走343段,合计635段视频,构建VideoROI_set_Extended数据集。针对每一个视频段,在分割出奶牛目标后,使用最小外接矩形(包含运动奶牛对象),计算最小外接矩形框的高度(Height,H),宽度(Width,W)和纵横比(Height/Width,Z)3个特征;以时间(T)为横轴,绘制3条奶牛运动时序曲线,并基于Improve Freeman编码法对3条曲线分别进行编码;最后将VideoROI_set_Extended视频集以8∶2的比例进行随机划分后,使用K最近邻分类器(K-nearest neighbor,KNN)和BP神经网络分类器(back propagation neural network,BP)2种分类器对时序曲线进行训练和识别。[结果]采样数m=10和角度数n=6时,使用KNN分类器进行识别,单一特征Z 10次的平均识别正确率达到97.64%;组合特征W&H&Z的时序曲线识别效果最好,KNN分类器的识别正确率达到99.21%。[结论]本文提出的基于时序运动特征的奶牛爬跨行为识别方法能够有效识别奶牛的侧爬跨行为,为计算机自动识别爬跨行为的奶牛提供依据。[Objectives]In order to improve the economic efficiency of the farm,find the cows in estrus and breed them in time,a research on the recognition of crawling behavior of cows based on machine vision was proposed in this article.[Methods]Three types of behavioral videos were selected:101 video clips of side crawling,191 video clips of following,and 343 video clips of walking.The VideoROI_set_Extended dataset containing 635 video clips was constructed.For each video segment,the cow object was segmented,and the moving cow object was contained by the smallest bounding rectangle.The three characteristics of the minimum bounding rectangle height(height,H),width(width,W),and aspect ratio(height/width,Z)were all calculated;With time(T)as the horizontal axis,3 cow motion timing curves were drawn.Based on the Improved Freeman coding method,3 curves were coded;Finally,after the VideoROI_set_Extended video set was randomly divided at a ratio of 8∶2,K-Nearest Neighbor(KNN)and Back Propagation Neural Network(BP)classifiers were used to train and recognize the cow’s motion curve.[Results]When the number of samples m was 10 and the number of angles n was 6,based on the aspect ratio feature(Z),the KNN classifier was used for recognition,and the average recognition accuracy rate of 10 times was 97.64%.Based on the combined feature(W&H&Z),the KNN classifier was used for recognition,and the average recognition accuracy rate of 10 times was 99.21%.[Conclusions]The method for identifying the crawling behavior of cows based on time-series motion characteristics proposed in this paper can effectively realize the lateral crawling behavior of cows.This research will provide a basis for automatically identifying the crawling behavior of dairy cows.

关 键 词：牛 爬跨行为 发情行为 时序运动曲线 Improve Freeman编码 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]