30-01-2013, 04:55 PM
Target tracking algorithm based on optical flow method using corner detection
1Target tracking algorithm.pdf (Size: 368.98 KB / Downloads: 19)
Abstract
The tracking speed and accuracy are two most important parameters for a target
tracking system. In our study, the proposed target tracking algorithm combines the Harris
method and the optical flow method. To improve the tracking speed, the Harris method is
initially used to extract some target corner features, and the optical flow method is then
used to more accurately match corner features for the subsequent video frames. When the
tracked target is rotated or distorted, the barycenter algorithm is employed to compute the
barycenter of those matched features of target. To meet the real-time-tracking requirement,
a small-zone image searching method and a high speed digital signal processing system are
also designed. Our experimental study shows that the method described in this paper has
high accuracy of target tracking, and can be applied to the situations of rotated, distorted,
and/or shielded targets, although it has a limitation that it is only suitable for smaller targets.
Introduction
The target images can be extracted and matched automatically by an automatic target
recognition algorithm in all view fields, i.e. a system with an automatic target recognition
algorithm can implement target capturing, recognition and tracking through a serial of
image processing procedures. An automatic target recognition algorithm may contain the
following parts, i.e. image pretreatment, partition, target detecting, character computing,
correlation matching, recognition and classification, motion analysis, target tracking,
feature-point selection of target, and so on.
Corner feature extraction
In the target tracking process, it is critically important to select a better feature extraction
method [14]. Corner feature is widely used in target tracking algorithms because it contains
a large amount of information about local image characters and target shape features [9].
In an image, target corner is an important local feature which focuses on a great amount
of important image information and is rarely affected by illumination change [1]. Moreover,
it has the rotation invariance property and contains only about 0.05% of image pixels.
Therefore, corner features are the smallest data to deal with, which may significantly
improve the speed of target tracking process. Thus, corner detection has many important
applications in practice, especially, in the real-time target tracking field. At present, the
reported target tracking algorithms using corner detection are mainly based on the SUSAN
and Harris methods [2].
Optical flow method
The algorithm using optical flow matching has higher accuracy in image registration [1].
The optical-flow-based registration is well-known and is widely used in computer vision,
e.g. [10, 11]. The optical flow method employed in this study uses optical flow matching of
the extracted target corner features for processing the subsequent video frames.
The optical flow method was first developed by Gibson in 1950 [4]. An optical
flow is 2-D motion field, i.e. the projection on the image plane of the 3-D velocity field of
a moving scene, which can be used to compute differences of different images using a
feature-point matching as the time changes. Taking target corner as feature points, the
first step of the method proposed in this study is to detect and track target corners with
subsequent images and then determine the target location. Therefore, the optical flow
field can be computed using the displacement of subsequent frame images. Optical flow
method is based on the following assumption, i.e., any change of the image-gray-level
distribution is completely caused by the movement of the target and/or the background
[5].
Hardware design for real-time target tracking system
The real-time target tracking hardware system is mainly constructed by the core chipset
TMS320C6416T [13] as the digital signal processor, programmable logic chipset CPLD,
and field programmable array FPGA. The hardware design is described in Fig. 3.
The main parts of the hardware include video signal processor, high rate analog to digital
(A/D) converter, digital image processor, digital image memory, data analyzer, data
communication interface with computer (or another system), and synchronization and adder
display.
Experimental results
Figure 4 shows the target tracking test results using the proposed method.
Figure 4(a) is the initial image. The target zone is manually selected and the corner
features are extracted by the Harris method after preprocessing. Seven corner features are
denoted by symbol “x” which is also the matched object in the optical flow algorithm.
Figure 4(b)–(e) are individually matching and tracking result of the 5th, 158th, 252nd,
288th frame in the test. The symbol “+” in Fig. 4 denotes the matching point in the optical
flow algorithm. Although the target has little change in Fig. 4(b), a little rotation in Fig. 4
©, and blur phenomenon in Fig. 4(d) compared with that in Fig. 4(a), the proposed method
can correctly yield matched corner features, and successfully track the target. In Fig. 4(e),
since the target image is bigger and rotated, some featured corner points deviate away from
the initial positions. For the cases of Fig. 4(e), the barycenter method should be used to
compute the barycenter of featured corner points in order to determine the target
displacement and location.
Conclusions
In this study a target tracking algorithm based on the optical flow method using corner
detection is proposed. In this algorithm, the Harris method is used to initially extract corner
features, and the optical flow method is then used to more accurately extract corner features
for the subsequent video frames. The barycenter method is employed to determine the
target location. According to the analyses and discussions of the results of numerical
studies, we can conclude that the proposed method has a high accuracy in target tracking,
but it suffers the limitation that it is only suitable for tracking small targets. When the target
is large, or when it is rotated or distorted, it can be easily lost. To overcome this difficulty,
much future research is needed.