基于人眼视觉非均匀特性的实时Mean Shift跟踪方法

由于在成像制导过程中需要实时处理大量的信息,为了在尽可能保留有效信息情况下降低计算量,采用了一种人眼视觉非均匀采样模型——对数极坐标模型,来压缩信息量以提高计算速度；另外,由于对数极坐标变换对目标形状具有旋转和缩放不变性,在跟踪非刚性变形目标时该模型能表现出很好的稳健性；考虑到在成像跟踪末段,质心、角点之类的跟踪方法会产生匹配点漂移,为了抑制匹配点漂移,采用基于目标强度特征的Mean Shift跟踪方法,并对采用Gauss核函数的Mean Shift方法进行了优化来加快计算速度；实验结果表明,该方法能够有效抑制匹配点漂移,是一个稳健的目标跟踪方法.

Real Time Mean Shift Tracking Method Based on Human vision Non-uniform Sampling

Because large quantity information needs to be dealt during imaging processing, the proposed method use human vision non-uniform sample model, called log-polar coordinate model, to compress quantity and increase calculating speed. Invariant feature of image rotation, scaling and gray scale can be gotten by log-polar coordinates transform, so it is a robust approach for tracking Non-rigid deformation of target. Tracking methods using centroid or corner feature maybe bring matching point drift; and in order to supress matching point drift Mean Shift tracking method based on target intensity character was adopt. Mean Shift method used Gauss kernel function was optimized to accelerate calculating speed. The results show that the effectiveness and stability of the proposed approach for supressing matching point drift.