基于深度学习YOLO v4的舍养育肥猪行为识别  

Behavior recognition of fattening pigs based on deep learning YOLO v4 model

作　　者：季照潼 李东明[1] 王娟[1] 张莉[1] 胡顺斌[1] JI Zhaotong;LI Dongming;WANG Juan;ZHANG Li;HU Shunbin(College of Mechanical and Electrical Engineering,Hebei Agricultural University,Baoding 071001,China)

机构地区：[1]河北农业大学机电工程学院,河北保定071001

出　　处：《黑龙江畜牧兽医》2021年第14期39-42,151,共5页Heilongjiang Animal Science And veterinary Medicine

基　　金：河北省重点研发计划项目(19227205D)。

摘　　要：为了增加生猪福利,实时检测猪只的健康状况,预防异常现象发生,试验进行了舍养育肥猪行为检测,即使用萤石云摄像头对5头散养的4月龄左右的健康舍养育肥猪进行视频数据采集,针对不同光照强度及猪只行为从视频中截取了2 401张图片,并将图片按照躺卧、站立、进食、坐立和侵略性行为类型进行标注,通过深度学习YOLO v4模型对猪只行为进行了训练、验证、测试和评估。结果表明:训练模型中猪只的躺卧、站立、进食、坐立和侵略性行为的识别精度分别为98.80%、95.05%、89.40%、79.41%、97.30%,平均精度为91.99%;利用该模型进行测试的精度分别为90.70%、90.16%、88.38%、80.75%、96.69%,平均精度则为89.34%;利用该模型可在侵略性行为出现前1~2 s内进行预判。说明基于深度学习YOLO v4的舍养育肥猪行为识别模型有效克服了环境中不同光照强度和噪声的影响,达到了较好的识别效果,并且利用模型可以有效地检测猪只的侵略性行为。In order to increase the welfare of pigs,detect the health status of pigs in real time and prevent the occurrence of abnormal phenomena,the behavior of house-bred fat pigs was detected in this test.Fluorite cloud camera was used to collect the video data from 5 four-month-old fattening pigs in the natural environment.A total of 2401 images were selected from the video based on different lighting and pig posture,and labelled as behavior types such as lying,standing,eating,sitting and aggressive behavior.Then deep learning YOLO v4 model was used to train,test,verificaction and assess the behaviors of pigs.The results showed that the average precision values of lying,standing,eating,sitting and aggressive behavior of pigs in the training model was 98.80%,95.05%,89.40%,79.41%and 97.30%,respectively,and the average precision was 91.99%.On the test set,the average precision tested by this model were 90.70%,90.16%,88.38%,80.75%and 96.69%,respectively,while the average precision are 89.34%,and predict the aggressive behavior within 1-2 seconds before it appears using this model.The results indicated that YOLO v4 model could effectively overcome the influence of different light intensity and noise in the environment,and achieve good recognition effect.The model can be used to effectively detect aggressive behavior in pigs.

关 键 词：深度学习 YOLO v4模型 目标检测 猪只行为 侵略性行为 

分 类 号：S828.4[农业科学—畜牧学] TP391.4[农业科学—畜牧兽医]