基于CenterNet的草原牛羊计数研究  

CenterNet⁃based grassland cattle and sheep counting study

作　　者：孙强 郝敏[1] Sun Qiang;Hao Min(College of Mechanical and Electrical Engineering,Inner Mongolia Agricultural University,Hohhot 010000,China)

机构地区：[1]内蒙古农业大学机电工程学院,呼和浩特010000

出　　处：《现代计算机》2023年第11期1-8,共8页Modern Computer

基　　金：内蒙古自然科学基金(2017MS0606)。

摘　　要：对草原放牧牛羊数量快速准确的统计是进行科学放牧管理和草原草畜平衡分析的重要前提。随着市场对畜牧产品需求的增加,内蒙古地区作为国内畜牧产品主要产地,其草原放牧牛羊的数量逐年上升,而目前以人工进行放牧牛羊数量统计工作存在难度大、效率低、统计准确度不稳定等问题。为实现牛羊数量的快速统计,使用无人机进行航拍采集的草原放牧牛羊图像并建立数据集,使用多尺度特征提取网络Res2Net提升anchor-free的CenterNet模型对航拍情景下放牧牛羊图像粘连小目标的检测效果,加入ECANet注意力机制进一步提升模型性能,并调节相关参数进行算法优化。研究结果表明,该模型在无人机航拍获取的数据集中的类平均精度达到88.22%,体现密集目标检测效果的牛羊对数平均漏检率分别下降了10%和11%。这一结果说明引入无接触式的计算机视觉技术在草原牛羊计数方面具有一定的价值。Rapid and accurate statistics on the number of cattle and sheep grazing on grassland is an important prerequisite for scientific grazing management and grassland grass-livestock balance analysis.As the market demand for livestock products in-creases,the number of grassland grazing cattle and sheep in Inner Mongolia,the main production area of livestock products in China,increases year by year,while the current manual work of counting the number of grazing cattle and sheep has problems such as difficulty,low efficiency and unstable statistical accuracy.In order to realize the fast counting of cattle and sheep,we use the aerial photography of the cattle and sheep images collected by UAV and build the dataset,use the multi-scale feature extraction net-work Res2Net to improve the detection effect of the anchor-free CenterNet model on the small targets adhered to the grazing cattle and sheep images under the aerial photography scenario,add the ECANet attention mechanism to further improve the model The performance of the model is further improved by adding ECANet attention mechanism,and the relevant parameters are adjusted for algorithm optimization.The research results show that the class average accuracy of the model in the dataset acquired by UAV aerial photography reaches 88.22%,and the logarithmic average missed detection rate of cattle and sheep,which reflects the effect of dense target detection,decreases by 10%and 11%,respectively.This result indicates the value of introducing contactless com-puter vision technology in grassland cattle and sheep counting.

关 键 词：CenterNet 目标检测 计数 Res2Net 注意力机制 

分 类 号：S818.9[农业科学—畜牧学] TP391.41[农业科学—畜牧兽医]