Information Technology Reference
In-Depth Information
Most recent object tracking algorithms [1], [2], [3] have used tracking-by-
detection approaches to address these challenges. Tracking by detection approach
treats the tracking problem as a detection task applied over time in video se-
quences. The methods first apply detection to the tracked object in the first
frame and then obtain the detection responses to track the object during the
video sequences. These methods maintain a classifier trained online to discrim-
inate the target object from its background. The classifier is used to estimate
object position by searching for the maximum classification score in a local region
around the estimation position from the previous frame.
Structured output tracking with kernels (STRUCK) [1] provides an adaptive
object tracking-by-detection approach with a kernelized structured output Sup-
port Vector Machine (SVM). With an online learning SVM, STRUCK perform
object tracking which integrates learning and tracking, avoiding the need for up-
date strategies in oine learning. While tracking-learning-detection (TLD) [2]
proposes a long-term object tracking in a video by integrates the object detec-
tion task, learning and tracking. Each subtask in TLD is addressed by a single
component and the component operate simultaneously.
However, many tracking-by-detection approaches fail to detect the right target
object because sometimes it mixed with its background and some parts of other
objects which leads to false detection for the features that not belonging to the
object and missed classification in video sequences.
In this paper, we take different approach to track the object in the video
by using the principle of structured output Support Vector Machine (SVM) [1].
Given the initial position of a target object in a frame of a video, we divide
the bounding box into sub-blocks with predefined size. And then we extract the
features from each sub-blocks and learn those features with structured output
classifiers [12]. Given the radius size from the center of a sub-block, we define the
search range which is the area inside the radius, excluding the initial bounding
box area. Then, we divide those search area with the same size of the sub-blocks.
Next, the features from all of the sub-blocks will be learned by using [12] method.
And then we construct a region graph; the nodes are sub-blocks and the edges
link any two adjacent sub-blocks that have 4-connected neighborhoods.
We compare our proposed method with [1] and [2] methods. From the ex-
perimental result, our method obtains better detection and tracking comparing
to those two methods. Our work offers several advantages over existing object
tracking schemes. With the sub-blocks as a unit, we still can detect the target
object if some parts of the object are occlude with other objects.
The rest of this paper is organized as follows: section 1 presents introduction,
section 2 explains our proposed methods to obtain ecient object video tracking
by using structured output SVM. Section 3 will present experimental results
while section 4 presents conclusions.
Search WWH ::
Custom Search