Meanshift tracking based on spatial histogram

Recently read an article "Spatiograms versus histograms for region-based tracking", in this article's core ideas and algorithmic reasoning to organize.

Spatial histogram

The traditional histogram is visualized as the zero-order spatial histogram, and the second space histogram includes the spatial mean value and covariance of each bin of the histogram, which can obtain the richer feature description of the target, thus improving the robustness of the tracking.

Probability density function

The spatial information of spatial histogram obeys the Gaussian distribution by default, and it is judged by the distribution characteristics of Dogause (GMM) for the similarity of two histogram.

Gaussian distribution: The probability density function mean of the normal distribution is a μ variance of σ2 (or standard deviation σ) is an example of a Gaussian function:

The smaller the σ, the more concentrated the distribution, the larger the σ, the more dispersed the distribution. If a random variable x obeys This distribution, we write x ~ N (μ,σ2). If μ= 0 and σ= 1, this distribution is called the standard normal distribution, this distribution can be simplified to

Dovigos Distribution formula:

D represents the dimension of X, which represents the d*d covariance matrix, defined as

The spatial histogram is similar to GMMs, but the multi-Gaussian weights and similar values obtained from the region I are GMMs, and the spatial histogram only obtains the values from a Gaussian distribution region. GMMs are non-parametric in their region, their range is semi-parametric, histograms are parameterless in their region and range, and spatial histograms are parameterless in their range, but their intervals are semi-parametric.

Meanshift

The Meanshift algorithm is a target tracking algorithm based on pattern matching, first manually selects the Tracking window, calculates the target template under kernel function weighting according to the color histogram distribution, and obtains the histogram distribution of the selected area in the subsequent tracking frame using the same method. The statistical iterative calculation makes each point drift in the direction of the maximum distribution similarity of two.

These are the core ideas of this article algorithm, and then I'll post the implementation code.

Meanshift tracking based on spatial histogram