A New Approach of GPU Accelerated Visual Tracking - Advanced Concepts for Intelligent Vision Systems - page 354

However, when considering a visual tracking system, the main requirements of the
tracking algorithms are about efficiency, accuracy and stability. From our experience,
with the increase of tracking region size (for example a 360

360 pixels region), the
ESM computation still costs too much time and induces a relative low processing speed.
This low processing speed will cause a larger image difference in the two successive
images of a fast moving object. As ESM tracking algorithm can only work well with
small image differences, these large differences will cause tracking failure.
To deal with these problems, we propose a novel approach of using GPU as copro-
cessor to enhance the system performance. Our contributions are mainly as follows.
We present a GPU based ESM tracking algorithm (GPU-ESM tracking) to address
the need for faster tracking algorithms. The speedup allows for a higher speed cam-
era so that there will be smaller difference between two successive frames, which will
make the ESM tracking result more reliable and robust. Besides GPU-ESM, we adopt
GPU based object recognition algorithms to solve those extreme cases for GPU-ESM
tracking, such as large image differences and occlusions, etc. We implement Lowe's
Scale Invariant Feature Transform (SIFT) algorithm[8] on GPU (GPU-SIFT) and ex-
tend GPU-SIFT algorithm with “RANdom SAmple Consensus” (RANSAC) method
to increase its accuracy. With an approximately 20 times GPU speedup, our extended
GPU-SIFT tracking greatly enhances the system reliability.
We propose an effective combination strategy of both algorithms mentioned above.
When GPU-ESM tracking failure happens, GPU-ESM will automatically load the re-
sult from GPU-SIFT so that it can continue tracking. Therefore, the whole system can
work smoothly with high reliability at a high processing speed. The previous paper [9]
mentions the ESM tracking and visual servo and in this paper 3D region tracking is
developed with this combination strategy.
The rest of this paper is organized as follows. Section II reviews the relative works
on ESM tracking and SIFT algorithms. Section III introduces the translation details of
two GPU algorithms so as to fully utilize the parallel capacity of GPU. This part also
covers the combination model of both algorithms in detail. Section IV describes the
experimental results to validate our proposed approach. Section V describes the key
optimization techniques in our GPU applications. Section VI concludes this paper.

×

2 Related Works

Our proposed approach is a combination of GPU-ESM tracking and GPU-SIFT algo-
rithms. For simplicity, we review the ESM tracking algorithm and SIFT algorithm.

2.1 ESM Tracking Algorithm

ESM tracking algorithm was proposed by Malis in 2004[6]. By performing second
order approximation of the minimization problem with only first order derivative, ESM
algorithm can get a high convergence rate and avoid local minima close to the right
global minima. Different kinds of its applications have been realized, such as visual
tracking of planar object and deformable object[10], visual servo[7] etc.
Suppose the tracking object is planar and projected in a reference image I ∗ with a
“Template” region of m pixels. Tracking this region consists in finding the homography