摘要
为准确实时跟踪羊只目标,进行疾病异常预警,实现奶山羊精细化养殖,本文基于DiMP跟踪模型,利用奶山羊跟踪对象单一且图像样本丰富的特点,结合迁移学习和类特定融合方法,设计了一种类特定的奶山羊目标跟踪模型,能够有效克服DiMP算法在跟踪类特定目标时定位精度不足的缺点。利用构建的奶山羊视频跟踪数据训练集对跟踪算法进行迁移训练,加快模型收敛速度,使评估网络预测出的边界框更贴合奶山羊真实框的位置和尺寸。在线跟踪阶段,针对目标模板仅采用第1帧特征制作整个序列的调制向量,导致该调制向量相对整个跟踪阶段特征不具代表性,与后续帧差异大的缺点,使用训练集制作包含奶山羊各种姿态的类调制向量,以指数消融方式更新奶山羊类调制向量与第1帧调制向量间的比重,增强边界框回归任务中的奶山羊特征与背景的判别性。提出的算法在测试集上的AUC(Area under curve)和精准度(Precision)分别为76.20%和60.19%,比DiMP方法分别提升6.17、14.18个百分点,跟踪速度为30 f/s,满足实时跟踪的要求。实验结果表明,提出的类特定奶山羊目标跟踪方法可用于监测复杂场景下奶山羊的运动,为奶山羊精细化管理提供了技术支持。
In the process of fine breeding of dairy goats,the accurate and real⁃time tracking of goat targets is an important basis for their behavior recognition and disease abnormality warning.Based on the DiMP tracking model,a kind specific target tracking model was designed for dairy goats,which can effectively overcome the disadvantage of insufficient positioning accuracy of DiMP algorithm in tracking specific targets.The migration training of the tracking algorithm was carried out by using the constructed dairy goat video tracking data training set to accelerate the convergence speed of the model and make the boundary box predicted by the evaluation network more fit the position and size of the real frame of the dairy goat.In the online tracking stage,aiming at the disadvantage that the target template only used the first frame features to produce the modulation vector of the whole sequence,which led to unrepresentative characteristics of the modulation vector relative to the whole tracking stage,the training set was used to produce the class modulation vector containing various poses of the dairy goat,and the proportion between the class modulation vector of the dairy goat and the modulation vector of the first frame was updated by exponential ablation to enhance the discrimination between characteristics and background of dairy goats in the boundary box regression task.The AUC and accuracy of the proposed algorithm on the test set were 76.20%and 60.19%,respectively,which were 6.17 and 14.18 percentage points higher than that of the DiMP method.The tracking speed was 30 frames per second(f/s),which met the requirements of real⁃time tracking.The experimental results showed that the proposed target tracking method can be used to monitor the movement of milk goats in complex scenes,and it can provide technical support for fine management of dairy goats.
作者
宁纪锋
张静
杨蜀秦
胡沈荣
蓝贤勇
王勇胜
NING Jifeng;ZHANG Jing;YANG Shuqin;HU Shenrong;LAN Xianyong;WANG Yongsheng(College of Information Engineering,Northwest A&F University,Yangling,Shaanxi 712100,China;Key Laboratory of Agricultural Internet of Things,Ministry of Agriculture and Rural Affairs,Yangling,Shaanxi 712100,China;College of Mechanical and Electronic Engineering,Northwest A&F University,Yangling,Shaanxi 712100,China;College of Animal Science and Technology,Northwest A&F University,Yangling,Shaanxi 712100,China;College of Animal Medical,Northwest A&F University,Yangling,Shaanxi 712100,China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2023年第6期280-286,400,共8页
Transactions of the Chinese Society for Agricultural Machinery
基金
陕西省农业科技创新驱动项目(NYKJ-2021-YL(XN)48)。
关键词
奶山羊
目标跟踪
类特定
迁移学习
DiMP
dairy goat
object tracking
class⁃specific
transfer learning
DiMP
