【作者】 王毅；

【导师】 王程；

【作者基本信息】 厦门大学 ， 计算机技术， 2021， 硕士

【摘要】 养殖场中活畜的存栏数量在养殖场日常管理、畜牧业财务审计、信贷系统中的资产评估、政府对于市场的调控与政策制定等方面是一个重要的参考指标。随着畜牧业智能化的不断推进,基于计算机视觉技术来实现活畜计数的相关研究已经引起了学术界的广泛研究兴趣。现有的解决方案大多在特定的实验环境和配置下工作良好,然而我国现阶段部分养殖场环境不具备安装固定专用设备的条件。智能手机不仅普及率高并且可以采集丰富的数据,给活畜计数方案的探索带来了新的思路。因此,在这种背景下,本文重点关注以手机为例的低成本相机进行活畜计数的问题。本文研究成果包括以下三个方面:1.构建了数量达10000张的生猪实例分割数据集,在此基础上改进了 Mask R-CNN网络并训练了实例分割算法模型。在湖北襄阳、福建漳州和江西吉安三个养殖场实地采集了不同品种、不同形态、不同生长周期的生猪养殖场景图像,进行了人工实例分割标注和数据增强。最终,实例分割模型OA准确率达到92%。2.提出了一种基于低成本相机在复杂环境下进行活畜计数的处理流程。该流程设计了养殖场环境中拍摄计数视频的操作要求,构建了包含实例分割、全景图拼接以及去重计数的处理框架。最终,该流程实现了典型养殖场环境中活畜的数量统计,相对误差为5%左右。3.提出了畜舍复杂环境下的全景拼接与去重计数方法。针对软件后处理工作方式,设计了基于深度学习网络SuperGlue进行单个畜舍的全景拼接方法;针对视觉-惯导硬件工作方式,设计了基于EXLAM 80X42设备进行单个畜舍的全景图拼接方法,并通过实验验证了两种全景拼接与去重计数方法的可行性。通过在真实的多个不同规模和不同养殖环境的养殖场中进行多项实验和案例研究,实验结果表明本文提出的处理流程能够准确地进行活畜的数量统计,并且具备通用性。更多还原

【Abstract】 The number of live livestock in farms is an important reference index in daily management of livestock farms,financial audit of animal husbandry,asset evaluation in credit system,regulation and control of market by government and policy formulation.With the development of intelligent animal husbandry,counting live livestock based on computer vision technology has aroused extensive research interest in academic circles.Most of the existing solutions work well in specific experimental environments and configurations.However,at the present stage,some of the farms in China do not have the conditions to install special fixed equipment.Smart phones not only have a high popularity rate,but also can collect rich data,which brings new ideas to the exploration of designing a universal live livestock counting scheme.Therefore,in this context,this paper focuses on the problem of counting live livestock with low-cost cameras using mobile phones as an example.The research results of this paper include the following three aspects:1.We build a 10000 live pig instance segmentation dataset,improve the Mask RCNN network and train the instance segmentation algorithm model.We collect live pig breeding scene images of different breeds,different postures and different growth cycles in three pig farms in Xiangyang,Hubei Province,Zhangzhou,Fujian Province and Ji’an,Jiangxi Province.We performe artificial instance segmentation annotation and data enhancement.Finally,the OA accuracy of the example segmentation model reaches 92%.2.We propose a low-cost camera based live livestock counting process in complex environment.We designed the operation requirements of shooting counting video in the breeding farm environment,and construct the processing framework including instance segmentation,panorama mosaic and de-counting.Finally,the process realizes the live livestock counting in the typical farm environment,and the relative error is about 5%.3.We propose a panoramic mosaic and de-repeat counting method in complex environment of livestock house.For the software mode,we design a panoramic mosaic method of a single livestock house based on the deep learning network SuperGlue.For the hardware mode,we design a panoramic mosaic method of a single livestock house based on EXLAM 80X42 equipment.We verify the feasibility of the two panoramic splicing and deconduplication counting methods through experiments.Through a number of experiments and case studies in real farms of different scales and different breeding environments,the experimental results show that the process proposed in this paper can accurately calculate the number of live livestock,and it is universal.更多还原