Meanshift algorithm Introduction

Meanshift is widely used in clustering, image smoothing, segmentation, and tracking. The concept of meanshift was first proposed by fukunage in 1975. Its initial meaning is just like its name: mean vector of the Offset. However, with the development of theory, the meaning of meanshift has changed a lot. Now, the meanshift algorithm we are talking about generally refers to an iterative step, that is, to first calculate the offset mean of the current point, and then use this as the new starting point to continue moving, until a certain ending condition is met.

For a long time, the meanshift algorithm was not paid enough attention until another important paper was published in 1995. Yizong Cheng, the author of this paper, defines a family of kernel functions, so that the contribution of the offset to the mean offset vector varies with the distance between the sample and the offset point. Secondly, he also set a weight coefficient to make the importance of different sample points different, which greatly expands the application scope of meanshift. In addition, researchers approximate non-rigid-body tracking to a meanshift optimization problem so that tracking can be performed in real time. Currently, using meanshift for tracking is quite mature.

The meanshift algorithm is actually a kernel density estimation algorithm. It moves each vertex to the local maximum point of the density function, that is, the point with a density gradient of 0, also known as the pattern point. In the non-parameter estimation section (see the http://blog.csdn.net/carson2005/article/details/7243425), we mention that multidimensional Kernel Density estimates can be expressed:

It is estimated that it is 0. The meanshift vector always points to the direction where the density is increased to the maximum, which can be ensured by the molecular items in the above formula, while the denominator item reflects the step of moving each iteration's core function, in a region that does not contain the feature of interest, the step size is longer, while in a region of interest, the step size is shorter. That is to say, the meanshift algorithm is a gradient rise algorithm with a variable step, or an Adaptive Gradient rise algorithm.