基于SE-R(2+1)D网络的自然环境下的奶牛行为识别  

Recognition of cow behavior in natural environment based on SE-R(2+1)D network

作　　者：史学伟 司永胜[1,2] 韩宪忠 王克俭[1,2] SHI Xuewei;SI Yongsheng;HAN Xianzhong;WANG Kejian(College of Information Science and Technology,Hebei Agricultural University,Baoding 071001,China;Key Laboratory of Agricultural Big Data of Hebei Province,Baoding 071001,China)

机构地区：[1]河北农业大学信息科学与技术学院,河北保定071001 [2]河北省农业大数据重点实验室,河北保定071001

出　　处：《河北农业大学学报》2023年第1期97-102,109,共7页Journal of Hebei Agricultural University

基　　金：河北农业大学精准畜牧学科群建设项目(1090064)。

摘　　要：智能行为识别对于奶牛健康的自动诊断和精准养殖具有重要意义。由于接触式传感器会损害动物福利,对奶牛产生应激反应。因此,本文设计了R(2+1)D模型对奶牛进行行为识别。3D网络作为1种时空卷积网络,可以有效识别奶牛的基本时序行为,但该模型针对奶牛的进食行为与反刍行为不易区分,因此对残差网络中的残差映射部分进行改进,在残差网络中添加注意力机制,将SE模块加入到残差映射部分。首先,利用Kinect相机采集奶牛的行为视频;其次,将采集到的奶牛视频分解成连续帧输入到改进后的模型中,连续帧经过二维空间特征和一维时间特征提取,经过残差网络的注意力模块,忽略一些无关的特征;最后,经过模型的Softmax层进行行为分类。实验证明,和原模型比较,准确率提高了2.36%。本文方法实现了精准的奶牛行为识别,可为智慧畜牧业的发展提供技术支持。Intelligent behavior recognition is of great significance for automatic diagnosis and precise breeding of dairy cow health. Stress responses to dairy cows as contact sensors compromise animal welfare. Therefore, this paper designs an R(2+1)D model for behavior recognition of dairy cows. As a spatiotemporal convolutional network, the 3D network can effectively identify the basic temporal behavior of cows. However, the model is not easy to distinguish the feeding behavior and rumination behavior of dairy cows, so the residual mapping part in the residual network is improved, the attention mechanism is added to the residual network, and the SE module is added to the residual mapping part. First, use the Kinect camera to collect the behavioral video of cows;secondly, decompose the collected cow video into consecutive frames and input them into the improved model. The attention module ignored some irrelevant features;finally, it goes through the softmax layer of the model for action classification. Experiments showed that the accuracy rate was increased by 2.36% compared with the original model. The method in this paper realized accurate cow behavior identification and provides technical support for the development of smart animal husbandry.

关 键 词：行为识别 R(2+1)D网络 深度学习 智慧畜牧业 

分 类 号：S823[农业科学—畜牧学]