4-D尺度空间中基于Mean-Shift的目标跟踪

在基于Mean-Shift的目标跟踪算法中，尺度自适应机制是算法研究的一个重要方向。一种典型的方法采用Lindeberg的尺度空间理论以获取目标尺度信息。但现有算法中将尺度由2-D矢量压缩为1-D量，未能精细地刻画目标仿射变换时的尺度变化，从而限制了算法的适用范围。为此，该文将尺度维1-D滤波推广至2-D，构造得到了相应的4-D尺度空间，并利用空间维和尺度维的Mean-Shift交替迭代，实现了同时在空间位置和尺度方向对目标的有效跟踪，提高了算法在目标尺度变化时的自适应性，并扩大了算法的适用范围。

Mean-Shift Object tracking through 4-D Scale Space

The scale self-adaptive mechanism is one of the promising research directions in the object tracking issue based on mean-shift. A typical method is to adopt Lindeberg’s scale-space theory to obtain the scale information of the target. However, the 2-D scale vector is reduced to 1-D in the existing algorithm. So it is not precise enough to portray the changing state of the object scale under affine transformation, which greatly limits the scope of application of the method. Hence, the 1-D filter in the scale dimension is extended to 2-D in this paper. Correspondingly, the 4-D scale space is constructed, then the problem is mapped into two interleaved mean-shift procedure in the spatial and scale dimension through scale space. The modified algorithm achieves the effective object tracking in spatial position and scale direction simultaneously, which enhances the self adaptability when the target scale is changing, and expands the scope of application of the algorithm.