With the assumption of small motion displacements in sequences, templates can be matched successfully by SSD (sum-squared-difference) optimization technique. In this paper, two crucial modifications are proposed to improve the original SSD tracking. First,
SSD Tracking Using Dynamic Template and Log-polar Transformation
Qiang Zhu1 Kwang-Ting Cheng1 Hongjiang Zhang2
1
Electrical and Computer Engineering, University of California, Santa Barbara, CA 93106
2
Microsoft Research Asia, Beijing, P.R.China, 100080
Abstract
With the assumption of small motion displacements in sequences, templates can be matched successfully by SSD (sum-squared-difference) optimization technique. In this paper, two crucial modifications are proposed to improve the original SSD tracking. First, templates are dynamically updated for each frame. Thus, instead of designing complicated parametric models to explain various image distortions in tracking sequences, more efficient and compact parametric models, say the pure translation model, can be adopted. Second, rotation and scale can be incorporated into a translation motion model after a Log-polar transformation. Extensive experimental results performed on live video sequences show the effectiveness and advantage of our approach.
Keywords: SSD Optimization, Dynamic Template, Log-polar Transformation, Face Tracking
1. Introduction
Visual tracking has emerged as an active research field in several applications including HCI [2] [11] and video surveillance [5]. The techniques [7] [8] related to image correlation imply a good solution to this challenging problem. With the assumption of small displacements between two consecutive frames, object can be tracked successfully by optimizing some matching criterions with respect to image deformations. Hager [3] proposed a general and efficient SSD (sum-squared-difference) framework to track objects undergoing geometric distortion, changing illumination and partial occlusion. Their work further showed that the computations needed to perform SSD optimization can be factored to greatly improve algorithm efficiency. After introducing twist motion and exponential maps, Bregler [1] reformulated SSD framework to the problem of articulated human body tracking in 3D space. Most of the previous attempts mainly focused on developing more accurate and complex 2D/3D parametric models to account for the image changes caused by geometry deformations, perspective projection, non-rigid object distortions and illumination changes etc. Unfortunately, complicated parametric models may bring several unexpected results. First, considerable increases in the computational effort required to establish correspondence are conflicted with the real-time applications. Secondly, for some applications, say face tracking, pose relative to viewing camera leads to dramatic appearance changes usually failed to be
modeled accurately. On the other hand, we noticed that a legal or mild human motion implies continuous image changes in sequences. Hence, instead of using complicated parametric models to interpret the image warps between the current frame and a static template obtained at the initial time, we propose a simple and low-order parametric (pure translation) model to describe the appearance changes between the current frame and a dynamic template, which is updated according to the content in the previous frame. Furthermore, by a Log-polar transformation, a pure translation model can account for the image variations caused by rotation and scale.
The remainder of this paper is organized as follows. In section 2, we show the principle of SSD optimization. Then, the comparisons between two types of SSD trackers (static template plus affine motion model vs. dynamic template plus pure translation model) will be performed in detail. Section 3 introduces Log-polar transformation to incorporate the rotation and scale into the tracker with the pure translation model. Section 4 shows the work of a real-time tracking platform. Finally, we present a short discussion of conclusion.
2. SSD Region Tracking
2.1 Algorithm definition
The 2D changes of image appearance under orthographic projection can be viewed as affine motion described by a six-parameter vector U(a1, a2, a3, a4, dx, dy). For a point (x t, y t) at time t, the new point (xt+1, y t+1) at time t+1 after an affine motion can be defined as:
xt+1 xt dx a1,a2 xt y(1)
t =A× + dx
+1 yt dy = a3,a4 × yt + dy Then the SSD objective function can be evaluated with respect to a set of points in the target region R:
O(u)=
X∑(I(f(P,U),t+1) I(P,t))
2
(2)
∈R
where I (P, t) is the image intensity value of a point P. f (P,U) indicates the affine transformation for P from time t to t+1. The linearization of Equation (2) can be written by expanding I(f(P,U),t+1)in a Taylor series with respect to motion U and time t:
O(u)=∑
(Ia1a1+Ia2a2+Ia3a3+Ia4a4+Idxdx+Idydy+Itτ)2 (3)
X∈R
Optimization above function yields the solution:
U= (MTM) 1MTIt (4)
Face Tracking Using Motion-Guided Dynamic Template Matching Liang Wang, Tieniu...(SSD) distance measure between the template and a search region is ...
Face Tracking Using Motion-Guided Dynamic Template Matching Liang Wang, Tieniu...(SSD) distance measure between the template and a search region is ...
热门文档