摘要
为了增强目标跟踪的有效性,提出了一种以粒子滤波作为跟踪框架,基于多特征级联的目标跟踪算法。以log-Gabor滤波器作为粒子判别级,滤除一定数量的无效粒子以提高粒子滤波的性能;再级联融合了log-Gabor特征、局部二值模式(LBP)特征和方向梯度直方图(HOG)特征的粒子加权级,实现目标跟踪。应用log-Gabor滤波器良好的频率响应对粒子做出总体评估以决定其有效性,同时以log-Gabor滤波器输出张成的频域特征。配合LBP和HOG局部特征,处理目标总体信息和细节信息,利用混合高斯模型突出后验概率分布中的峰值状态。实验结果表明,该文算法能快速去除无效粒子,实现目标的鲁棒跟踪。
In order to improve the effectiveness of visual object tracking,an object tracking algorithm based on multiple features cascade is proposed with particle filters as a tracking frame. Invalid particles are removed by log-Gabor filters as a discrimination stage; particle tracking is realized by cascade particle weighting stage combing log-Gabor features,local binary pattern（ LBP） features and histograms of oriented gradients（ HOG） features. The particles are estimated totally by the frequency response of log-Gabor filters,so that their effectiveness is decided,and the frequency domain characters are outputted by the log-Gabor filters. The total information and detail information are handled by considering LBP and HOG local features. The peak valve of posterior probability distribution is pro-truded using Gaussian mixture model（ GMM）. Experimental results indicate that the proposed method can remove invalid particles efficiently and realize robust tracking effectively.
出处
《南京理工大学学报》
EI
CAS
CSCD
北大核心
2015年第3期286-292,共7页
Journal of Nanjing University of Science and Technology
基金
国家自然科学基金(61373055)
关键词
多特征级联
目标跟踪算法
粒子滤波
LOG-GABOR滤波器
局部二值模式
方向梯度直方图
频率响应
混合高斯模型
后验概率分布
multiple features cascade
object tracking algorithm
particle filter
log-Gabor filter
local binary pattern
oriented gradients histograms
frequency response
Gaussian mixture model
posterior probability distribution
