Proceedings Article | 7 May 2007
KEYWORDS: Detection and tracking algorithms, Motion models, Affine motion model, Image segmentation, Image processing algorithms and systems, Algorithm development, Computing systems, Video, Expectation maximization algorithms, Feature extraction
Object tracking is an important component of many computer vision systems. It is widely used in video surveillance,
robotics, 3D image reconstruction, medical imaging, and human computer interface. In this paper, we
focus on unsupervised object tracking, i.e., without prior knowledge about the object to be tracked. To address
this problem, we take a feature-based approach, i.e., using feature points (or landmark points) to represent
objects. Feature-based object tracking consists of feature extraction and feature correspondence. Feature correspondence
is particularly challenging since a feature point in one image may have many similar points in another
image, resulting in ambiguity in feature correspondence. To resolve the ambiguity, algorithms, which use exhaustive
search and correlation over a large neighborhood, have been proposed. However, these algorithms incur
high computational complexity, which is not suitable for real-time tracking. In contrast, Tomasi and Kanade's
tracking algorithm only searches corresponding points in a small candidate set, which significantly reduces computational
complexity; but the algorithm may lose track of feature points in a long image sequence. To mitigate
the limitations of the aforementioned algorithms, this paper proposes an efficient and robust feature-based tracking
algorithm. The key idea of our algorithm is as below. For a given target feature point in one frame, we first
find a corresponding point in the next frame, which minimizes the sum-of-squared-difference (SSD) between the
two points; then we test whether the corresponding point gives large value under the so-called Harris criterion.
If not, we further identify a candidate set of feature points in a small neighborhood of the target point; then find
a corresponding point from the candidate set, which minimizes the SSD between the two points. The algorithm
may output no corresponding point due to disappearance of the target point. Our algorithm is capable of tracking
feature points and detecting occlusions/uncovered regions. Experimental results demonstrate the superior
performance of the proposed algorithm over the existing methods.
