摘要
视觉跟踪最大的挑战在于能否建立一种能有效适应目标外观变化的观测模型,这就需要跟踪算法能对不断变化的目标外观模式进行在线学习。提出一种基于综合子空间的观测算法,在贝叶斯估计的前提下,用PCA子空间和正交子空间来描述目标外观。该算法结合了PCA子空间和正交子空间的优点,既能学习到目标的低维描述子空间,又能迅速学习到最新的目标外观变化模式。通过构建跟踪观测模型,并在粒子滤波框架下进行实验。结果表明,该方法能够有效地跟踪目标,性能优于PCA算法,而且其在光照变化,目标转动等外观变化大情况下仍能稳定地跟踪目标。
To build an observing model adapting to the rapidly changing object appearance is the biggest challenge of visual tracking.So the tracking algorithm requires learning the object appearance on line which may vary quickly.An algorithm based on mixed subspace is proposed which colligates PCA subspace and orthogonal subspace together and builds a tracking observation model.The experiment is executed under the framework of particle filtering.The result shows that this kind of algorithm can track the object more effectively than only using PCA subspace,especially under situations of drastically changing appearance such as illumination and object rotation.
出处
《控制工程》
CSCD
北大核心
2009年第S1期101-103,188,共4页
Control Engineering of China
