基于图像特征融合的农事活动行为的识别  

Recognition of the agricultural activities based on image feature fusion

作　　者：李子茂[1,2] 余慧 夏梦 郑禄[1,2] 徐杰 LI Zimao;YU Hui;XIA Meng;ZHENG Lu;XU Jie(College of Computer Science,South-Central University for Nationalities,Wuhan,Hubei 430074,China;Hubei Provincial Engineering Research Center for Intelligent Management of Manufacturing Enterprises,Wuhan,Hubei 430074,China)

机构地区：[1]中南民族大学计算机科学学院,湖北武汉430074 [2]湖北省制造企业智能管理工程技术研究中心,湖北武汉430074

出　　处：《湖南农业大学学报（自然科学版）》2021年第5期603-608,共6页Journal of Hunan Agricultural University(Natural Sciences)

基　　金：湖北省科学技术厅创新专项(2019ABA101)。

摘　　要：针对农事活动图像中人体姿态所隐含的行为信息以及人与农具所隐含的关联信息,提出了一种基于图像特征融合的农事活动行为的识别方法:利用人体姿态估计技术OpenPose提取农事行为关节点位置信息,利用目标检测YOLOv3提取农事行为中农具的位置和分类信息,用以构建农事行为的距离空间特征矩阵和角度空间特征矩阵,并将这些特征进行图像特征融合,建立基于图像显式特征和隐式特征融合的农事活动行为识别方法EI–SVM,实现农事活动行为的识别。试验结果表明,EI–SVM方法对农事活动行为识别的准确率可达94.87%,在公用数据集上准确率达到92.39%。Aiming at the behavior information implied by human posture in agricultural activity images and the association information implied by human and agricultural tools,a recognition method of agricultural activity behavior based on image feature fusion is proposed.Human posture estimation technology OpenPose is used to extract the joint position information of agricultural behavior,and target detection YOLOv3 is used to extract the position and classification information of agricultural tools in agricultural behavior.These information is used to construct the distance space feature matrix and angle space feature matrix of agricultural behavior,to fuse the above image features.Based on explicit and implicit features,the recognition method EI-SVM was establish to realize the recognition of agricultural activity behavior.The experimental results show that the accuracy of EI-SVM method for agricultural activity behavior recognition is 94.87%,and the accuracy on public data set is 92.39%.

关 键 词：农事活动 人体姿态 行为识别 目标检测 图像特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]