摘要
针对传统的视频跟踪算法对视频跟踪的精度不足以及主成分分析(PCA)的非线性拟合能力较弱的问题,将卷积神经网络与交互似然(IL)算法相结合,在深度学习的基础上对粒子滤波算法进行了优化改进。将核主成分分析(KPCA)网络应用于视频跟踪来获取目标的深层次特征表达,并采用一种新的交互似然图像跟踪器,非迭代地计算,对不同区域进行跟踪取样来减少数据之间的关联需求。在图像集上将所提算法与多种改进算法进行评估对比,结果表明所提算法具有非常好的鲁棒性及精确性。
The traditional video target tracking methods usually prossess low accuracy.This paper proposed an improved scheme based on convolution neural network and the interactive likelihood algorithm,and optimized the particle filter algorithm on the basis of deep learning.To address the issue of deficient nonlinear fitting ability of the principal component analysis(PCA),a kernel principal component analysis(KPCA)tracking algorithm was provided to obtain the deeper characteristic expression of the target.Then,a novel interactive likelihood(ILH)method was performed for image-based trackers,which can non-iteratively compute the sampling of areas belonging to different targets and thus reducing the requirement for data associations.The performance of the presented algorithm was evaluated in comparison with several related algorithms on image datasets.The experimental results demonstrate the great robustness and accuracy of the proposed algorithm.
作者
张明月
王静
ZHANG Ming-yue;WANG Jing(School of Computer Science and Technology,Nanjing Tech University,Nanjing 211816,China)
出处
《计算机科学》
CSCD
北大核心
2019年第2期279-285,共7页
Computer Science
关键词
目标跟踪
深度学习
卷积神经网络
核主成分分析
交互似然
Target tracking
Deep learning
Convolutional neural network(CNN)
Kernel principal component analysis(KPCA)
Interactive likelihood(IL)
