Accurate scale estimation for robust Visual Tracking

Original: http://blog.csdn.NET/gxb0505/article/details/52601613?locationNum=8 profile (accurate scale estimation for robust Visual Tracking)

DSST (discriminative scale Space Tracking) won the first place in the 2014 VOT, the algorithm concise, excellent performance, and I last described the KCF won the third place, both are based on the filter algorithm, this year is the CF rebel projection of the year , it is worth studying these similar excellent algorithms. This algorithm is based on the improvement of Mosse, highlighting the content is added to the scale transformation , the following begins to explain the algorithm content. Correlation Filter

First of all, the relevant filter proposed by Mosse, extract a series of image patches from the target, recorded as F1,f2,... ft as a training sample, the corresponding filter response value is a Gaussian function g1,g2,... GT, and the purpose is to find the minimum mean variance ( Minimum Output Sum of squared Error) optimal filter:
ε=∑j=1t| | ht∗fj−gj| | 2=1mn| | htfj−gj| | 2 (1)
The second equal sign is derived from the Parseval theorem, the equation to the left of the equation, the equation on the right is the frequency domain, it is precisely this equation, so that we transform the problem into the frequency domain solution, the minimum value of ε in the frequency domain solution is as follows:
HT=∑TJ=1GJFJ∑TJ=1FJFJ (2)