系留球载光学影像的草原牲畜数量自动核算方法

Automatic Grassland Livestock Quantity Accounting Method for Optical Images Acquired by Tethered Balloon

开展草原放牧型牲畜数量的精准分类核算对评估草地的实际载畜能力、评估草原资源风险等应用具有重要意义。系留气球可提供定点驻空的观测平台,具有载重量大、驻空时间长的特点,对于重点地区的草地载畜量监测具有重要应用价值。系留气球载荷通常是以较大倾角斜视观测,不同观测角度下图像的分辨率存在差异,且出现目标的模糊与粘连现象,对牲畜数量核算带来较大挑战。针对上述问题,提出了一种基于异常检测与模糊均值聚类的球载可见光影像牲畜数量核算方法,通过异常检测算法将牲畜目标与草地背景区分,并采用形态学方法去除噪声点,利用模糊均值聚类算法计算目标的中心位置并进行数量统计。通过基于2种球载光学相机载荷获取的牲畜影像中对1376头羊与273头牛的数量核算试验结果表明,所提方法对羊与牛的数量核算精度分别为90.5%与94.1%,总计核算精度93.0%,比目前主流的深度学习目标检测算法得到的核算精度(77.9%)有明显提升,对不同角度获取的影像的牲畜核算有较好的适应性,对草原牲畜数量核算与草地载畜量估计应用具有重要价值。

Accurate classification and accounting of the quantity of grassland grazing livestock is of great significance for applications such as grassland actual carrying capacity assessment and grassland resources risk assessment.Tethered balloons can provide a fixed-point observation platform which has the characteristics of large carrying capacity and long observation time,and have important application value for grassland carrying capacity monitoring in key areas.Tethered balloon loads usually observe at large inclination angle,the resolution of the image under different observation angles is different,and the objects in the image are usually blurred and overlapped,which brings great challenges to the calculation of the quantity of livestock.To solve the above problems,an accounting method for the quantity of livestock in balloon-borne visible light images based on anomaly detection and fuzzy mean clustering is proposed.An anomaly detection algorithm is used to distinguish the livestock objects from the grassland background,and the morphological method is used to remove noise points.Finally,the fuzzy mean clustering algorithm is used to calculate the center position of the objects and obtain quantitative statistics.The results of the quantity accounting experiment of 1376 sheep and 273 cattle in the livestock images acquired from the two types of balloon-mounted optical cameras show that the accuracy of the method for the quantity accounting of sheep and cattle is 90.5%and 94.1%respectively,and the total accounting accuracy is 93.0%,which is significantly improved as compared with the current mainstream deep learning object detection algorithm(77.9%).At the same time,the proposed method has better adaptability for the images acquired from different angles.Therefore,the method has important value for grassland livestock quantity accounting and grassland livestock capacity estimation application.