Thesis note "Tracking Using Dynamic programming for appearance-based sign Language recognition"

I. Overview

This is when I do gesture recognition, in the resolution of the gesture screen when looking at a paper, the key is to use dynamic programming as a tracking algorithm, the effect is comparable to cameshift and KF, but in occlusion, Gaps or offline tracking is a good time to do it.

Second, the algorithm steps

　　Step1: For the time of the T frame such as x_t each pixel (x, y), first calculated a score Q (t,x,y), called the local score, this later will say, score function is your own choice, Then we need to figure out a Q (t,x,y), that is, Global score,q (t,x,y) is the sum of the scores of best tracking at the time t (x, Y) (personal understanding is the choice of the largest score path), so here's the idea of DP, That is, for Q (t,x,y), its optimal value, basically definitely by Q (t-1,x ', y ') come over, X ', y ' is the neighborhood of X, Y.

　　　　　　

Here the comparison of the two formula of the paper can be seen, and then the paper in the calculation of the time added to the L2 of the regular optimization, after all, optimization method home plus not talk about it ... The reason is also very simple, (x, Y) and (× ', y ') distance cannot be too large, because the object is continuous motion.

　　Step2: Here is the use of the above obtained Q (t,x,y) and B (t,x,y) to do traceback, and finally reconstruct out the best path.

Tracking method with dynamic programming there is a problem, that is, he generally limited his tracking size, otherwise, the calculation will be very large, so there are two of trick

1, one point (x, y) only satisfies Q(t,x,y) > max(q(t,x,y))−t0 is considered as a t+1 time.

predecessor, here T0 is a parameter set by itself

2, using the above method and jump penalty function to do local adjustment, so that you can do a adaptive window

Third, score function

　　In short, you choose a suitable function for distinguishing between the different image, here is a simple motion information Scor function, which calculates the sum of the value of pixel in a size. Next said eigenfaces and skin Color score Function, in fact, you are only in the actual scene when a self-balance, the key point is that you can choose two score Function, At last, we'll do it with a W-weight as a super-parameter.

Iv. Summary

Later also said the use of hmm do recognition and tracking fusion, to tell the truth, did not read ... Have a chance to see it again.

　

　　

Thesis note "Tracking Using Dynamic programming for appearance-based sign Language recognition"