摘要
针对目标跟踪中的尺度变化、旋转、遮挡等问题,提出基于高斯混合模型的核相关滤波目标跟踪算法。利用卷积神经网络提取卷积特征并建立目标外观的高斯混合模型,利用核相关滤波算法检测目标位置,使用多尺度、多形状跟踪方法精确定位目标,在线更新高斯混合模型和核相关滤波器。在公开数据集上进行定量和定性分析,并与多种跟踪算法比较,该算法的距离精度和重叠精度相比核相关滤波算法,分别提高了19%、54%。实验结果表明,采用高斯混合模型和多尺度、多形状跟踪方法,较好解决了外观和尺度变化问题,相比其它算法具有更好的鲁棒性和适应性。
Visual tracking remains a challenging problem due to the appearance changes caused by deformation and abrupt motion.To address this issue,a kernelized correlation target tracking approach based on Gaussian mixed model was proposed.With this method,a CNN was introduced to extract convolution features,and target position was estimated using Gaussian mixed model and kernelized correlation filter.The target accurate scale and shape were estimated using multiple scale and shape tracking approach,and the kernelized correlation filters were updated in real time.Experiments were carried out on public data sets,compared with KCF,the distance accuracy and overlapping precision of the algorithm are improved by 19%and 54%,respectively.The results show that the proposed approach exhibits better performance than other algorithms and maintains its good robustness and adaptability even in complex scene.
作者
欧阳城添
汤懿
OUYANG Cheng-tian;TANG Yi(School of Information Engineering,Jiangxi University of Science and Technology,Ganzhou 341000,China)
出处
《计算机工程与设计》
北大核心
2019年第11期3170-3174,3179,共6页
Computer Engineering and Design
基金
国家自然科学基金项目(61561024)
江西省自然科学基金项目(20151BAB207035)
关键词
目标跟踪
卷积特征
相关滤波器
判别模型
高斯混合模型
target tracking
convolution feature
correlation filter
discriminant model
Gaussian mixed model