基于多模态联合稀疏表示的视频目标跟踪.doc


基于多模态联合稀疏表示的视频目标跟踪#
段喜萍1,2,刘家锋1,唐降龙1**
(1. 哈尔滨工业大学计算机科学与技术学院,哈尔滨 150001;
5
10
15
20
25
30
35
40
2. 哈尔滨师范大学计算机科学与信息工程学院,哈尔滨,150025)
摘要:复杂场景下的目标跟踪过程中,单模态特征往往不能很好地区分目标与背景,从而影
响跟踪精度。针对这一问题,本文提出使用多模态进行目标表示。在多模态表示上,考虑到
分别对各模态进行稀疏表示存在的潜在不稳定性以及多模态间存在的关联,提出对多模态特
征进行联合稀疏表示,以提高稀疏表示的可靠性。具体跟踪过程是:在粒子滤波框架下,对
表示粒子的各模态特征进行联合稀疏表示,使得各模态特征趋于产生相同的稀疏模式;然后
通过组合粒子的各模态重建误差计算各粒子的观察概率,进而最终将具有最大观察概率的粒
子确定为跟踪目标。实验结果表明,与基于单模态的稀疏表示方法相比,本文方法具有更高
的精度;与基于多模态分别进行稀疏表示的跟踪算法相比,本文方法的精度也更高。
关键词:计算机视觉;目标跟踪;粒子滤波;稀疏表示
中图分类号:TP391
Multi-cue Visual Tracking Based on Joint Sparse
Representation
DUAN Xiping1,2, LIU Jiafeng1, TANG Xianglong1
(1. School puter Science and Technology, Harbin Institute of Technology, Harbin, 150001;
2. Computer Science and Information Engineering college, Harbin Normal University, Harbin,
150025)
Abstract: plex tracking environment, the single feature usually can’t distinguish the
target from background well and the corresponding tracking accuracy is low, so that multiple
features are considered to be used to represent the tracked object. In regard to the representation of
multiple features, respective sparse representation of each feaure may cause the potential
instability and multiple features are correlated to each other, so the joint sparse representation is
proposed to improve reliability of sparse position. The tracking process is as follows. In
particle filter framework, multiple features of each particle are extracted and the joint sparse
representation is conducted, then its observation probality is calculated binating
reconstruction errors of all features, and at last the particle with the maximum observation is
determined as the target. Experimental results show that the proposed method is much better than
single feature method. And paring with respective sparse representation of multiple
features, the proposed method also performs much better.
Key words: Computer vision; visual tracking; particle filter; sparse representation
0