基于YOLOv5的猪只盘点方法研究

Research on Pig Inventory System Based on YOLOv5

针对传统生猪养殖场采用人工盘点猪只数量效率低且人力物力耗费较大的问题,本文提出一种基于YOLOv5目标检测算法的猪只盘点系统。采集实验猪厂的猪舍内摄像头图像2000张,除去部分污染严重或猪只重叠难以分辩的图像,经筛选后剩余1700张,网络收集猪舍图像1300张,共计3000张组成数据集,使用vott标注工具对采集图像进行标注。通过YOLOv5目标检测算法对数据集进行训练,并得到训练模型,将训练模型嵌入Jetson Nano开发板,在青岛试验猪场进行猪只盘点实验。实验结果表明,猪只白天识别准确率达95.8%,夜间识别准确率达86%,将训练模型部署到Jetson Nano嵌入式开发板,可稳定运行6路视频流,一台设备能同时对6个猪舍摄像头进行盘点计数,与人工盘点相比,该方法经济高效,应用前景更广阔。

ITraditional pig farms generally use the method of manually counting the number of pigs,which is inefficient and consumes a lot of manpower and time.To solve this problem,this paper proposes a pig inventory system based on yolov5 target detection algorithm.Through the collection of camera images in the pig house of the experimental pig factory and the network,the pig house images are collected to make a data set,in which the camera collects a total of 2000 images in the experimental pig farm.Due to the stains of some cameras in the pig farm,the serious pollution of the images taken,and the problem of overlapping pigs with difficulty to distinguish some images collected by the camera in the pig farm,1700 images remain after screening.A total of 1300 images are collected on the network,and the images collected with the pig farm camera form a data set of 3000 images.The data set is trained by yolov5 target detection algorithm to obtain the training model,and the training model is deployed on the Jetson nano embedded development board.The experimental results show that the accuracy of pig recognition is 95.8%and 86%at night.When the training model is deployed on the Jetson nano embedded development board,it can stably run 6-channel video streams.One device can count 6 pig house cameras at the same time,which is more economical and efficient than the manual counting method.