基于支持矢量机和信任域的目标跟踪算法

从模式分类的角度看待跟踪问题,将SVM(支持矢量机)方法、信任域算法和尺度空间理论相结合,提出了一种能精确描述目标尺寸连续变化的新的序列图像目标跟踪算法。首先使用SVM算法对像素分类,将每一帧输入图像转换成目标概率分布图,目标区域在概率分布图中呈现为灰度块。然后通过检测该图在尺度空间中微分滤波器输出的极值,来决定这些灰度块的尺度。最后使用QP-TR信任域算法在尺度空间里和图像平面内快速搜索概率分布图多尺度规范化Laplacian滤波函数的极值,实现了目标定位并同时决定了其尺度,从而完成了跟踪任务。通过和现有算法的比较以及在大量真实序列图像上的实验表明,该算法不但能够更准确地描述目标的大小,而且显著提高了跟踪算法的精度。

Object Tracking by Using SVM and Trust-Region Method through Scale-Space

The tracking problem is tackled from the view point of pattern classification. A new approach of tracking objects in image sequences is proposed, in which the constant changes of the target's size can be precisely described. For each inputt frame, a probability distribution image of the target is created through the classification of each pixel by SVM, where the target's area turns into a blob. The scale of this blob can be determined based on local maximum of differential scale-space filters. Then the QP_-TR trust region algorithm is employed to search for the local maxima of multi-scale normalized Laplacian filter of the probability distribution image for locating the target as well as determining its scale. In the presented tracking examples, the new method can be used to describe the target more accurately and thus achieves much better tracking precision.