A New Approach of GPU Accelerated Visual Tracking - Advanced Concepts for Intelligent Vision Systems - page 353

A New Approach of GPU Accelerated Visual Tracking

Chuantao Zang and Koichi Hashimoto

Graduate School of Information Sciences, Tohoku University
6-6-01 Aramaki Aza Aoba, Aoba-ku, Sendai 980-8579, Japan
{ chuantao,koichi } @ic.is.tohoku.ac.jp

Abstract. In this paper a fast and robust visual tracking approach based on GPU
acceleration is proposed. It is an effective combination of two GPU-accelerated
algorithms. One is a GPU accelerated visual tracking algorithm based on the
Efficient Second-order Minimization (GPU-ESM) algorithm. The other is a GPU
based Scale Invariant Feature Transform (SIFT) algorithm, which is used in those
extreme cases for GPU-ESM tracking algorithm, i.e. large image differences, oc-
clusions etc. System performances have been greatly improved by our combina-
tion approach. We have extended the tracking region from a planar region to a 3D
region. Translation details of both GPU algorithms and their combination strategy
are described. System performances are evaluated with experimental data. Opti-
mization techniques are presented as a reference for GPU application developers.

Keywords: ESM, SIFT, GPU, Visual tracking.

1 Introduction

Visual tracking is the critical task in computer vision applications, such as visual servo,
augmented reality, etc. Visual tracking methods can be mainly divided into two cate-
gories: Feature-based methods and Region-based methods[1]. Feature-based methods
mainly track local features such as points, line segments, edges or corners in the images.
These local feature detections are easy to process but sensitive to illumination change,
occlusion and so on. Region-based methods only use the image intensity information
in a certain region. By minimizing the sum of squared differences (SSD) between a
region in reference image and a warped region in current image, the transformation
parameters can be estimated[2] in these methods . For example, the transformation be-
tween two images of a plane is a homography[3]. A well-known Region-based method
is the Lucas-Kanade algorithm [4][5]. It computes the displacement of points between
consecutive frames when the image brightness constancy constraint is satisfied.
To minimizing the SSD between a template region and a warped region in Lucas-
Kanade algorithm, many nonlinear optimization approaches have been proposed with
different kinds of approximations, such as Standard Newton method [3], Gauss-Newton
approximation. Among these solutions, the Efficient Second-order Minimization (ESM)
algorithm[6] is an elegant idea which obtains the same convergence speed as standard
Newton method while not computing the computationally costly Hessian matrix. Based
on the ESM algorithm, Malis has proposed an efficient “ESM visual tracking algorithm”
and extended it to visual servo[6][7] .