摘要
在可视化跟踪过程中目标窗经常会由于遮挡、光照、姿势等变化而发生跟踪漂移,影响目标跟踪的准确性和稳定性。为解决该问题,提出一种基于图层的离散域均值漂移算法,在离散域提取基于核的直方图作为目标模型,并对离散分区中的目标函数进行平滑以避免寻优搜索陷入局部极小值,从而提高目标跟踪性能。实验结果表明,与多示例学习算法相比,该算法的跟踪精度提高了16%,具有更好的实时性和鲁棒性。
In the process of visual tracking, target window is always drifted for the illumination light change, deformation and poses change which affect the accuracy and robust tracking performance. To solve this problem, this paper proposes a novel Mean Shift （MS） algorithm based on picture layer which represents target model by the kernel histogram in the discrete fields. In order to improve the tracking performance, the objective function is smoothed to avoid falling into the local minimum in the search procedure. Experimental results show that the tracking precision of proposed algorithm increases by 16% compared with Multiple Instance Learning （MIL） algorithm, and it has better real-time and robustness.
出处
《计算机工程》
CAS
CSCD
北大核心
2015年第9期281-285,291,共6页
Computer Engineering
基金
福建省教育厅基金资助项目(JA12263
JB11127)
福州市科技合作基金资助项目(2013-G-86)
关键词
目标跟踪
目标表示
离散域模型
均值漂移
迭代寻优
object tracking
object representation
discrete domain model
Mean Shift（ MS ）
iterative optimization
