31-08-2012, 03:00 PM
Real-Time Tracking of Moving Objects with an Active Camera
[attachment=33019]
Introduction
Traditional computer vision methodology regarded the
visual system as a passive observer whose goal was the
recovery of a complete description of the world. This
approach led to systems which were unable to interact
in a fast and stable way with a dynamically changing
environment. Several variations of a new paradigm
appearing under the names active, attentive, purposive,
behavior-based, animate, qualitative vision were introduced
in the last decade in order to overcome the
efficiency and stability caveats of conventional computer
vision systems. A common principle of the new
theories is the behavior-dependent selectivity in the
way that visual data are acquired and processed. To cite
one of the first definitions [1]: ‘‘Active Sensing can be
stated as a problem of controlling strategies applied to
the data acquisition process which will depend on the
current state of the data interpretation and the goal or
the task of the process’’.
Related Work
As pursuit is one of the basic capabilities of an active
vision system, most of the research groups possessing a
camera platform have reported results. We divide the
approaches into two groups. The first group consists of
algorithms that use only motion cues for gaze shifting
and holding, and this is the group to which our system
belongs. Computational basis of this approach group is
the difference between measured optical flow and the
optical flow induced by camera motion.
The Oxford surveillance system [9, 10] uses data from
the motor encoders to compute and subtract the
camera motion-induced flow. It runs in 25 Hz with
processing latency of about 110 ms. Camera behavior is
modeled as either saccadic or pursuit motion. Saccadic
motion is based on the detection of motion in the
coarse scale periphery. Pursuit motion is based only on
the optical flow of the foveal region. This is also the
difference to our system, which can also smoothly
pursue but with repeated motion detection.
Conclusion
We presented a system that is able to detect and pursue
moving objects without knowledge of their form or
motion. The performance of the system with control
rate of 25 Hz, a latency of 80 ms, and average angular
velocities of about 10 degrees per second, is competitive
with respect to the state of the art. The system
needs the minimal number of tuning parameters: a
threshold for normal flow difference, a threshold for the
image gradient, a minimal image area over the mentioned
thresholds, and the maneuvering index.
We have shown that in order to achieve real-time
reactive behavior we must apply the appropriate image
processing and control techniques. The main contribution
of this paper is not only in the achieved high
performance of the system. Our work is different from
other presentations in the study of the individual
components with respect to the given hardware, time
constraints, and desired tracking behavior. We experimentally
studied the responses of the image processing
filters if fixed-point arithmetic is used.
[attachment=33019]
Introduction
Traditional computer vision methodology regarded the
visual system as a passive observer whose goal was the
recovery of a complete description of the world. This
approach led to systems which were unable to interact
in a fast and stable way with a dynamically changing
environment. Several variations of a new paradigm
appearing under the names active, attentive, purposive,
behavior-based, animate, qualitative vision were introduced
in the last decade in order to overcome the
efficiency and stability caveats of conventional computer
vision systems. A common principle of the new
theories is the behavior-dependent selectivity in the
way that visual data are acquired and processed. To cite
one of the first definitions [1]: ‘‘Active Sensing can be
stated as a problem of controlling strategies applied to
the data acquisition process which will depend on the
current state of the data interpretation and the goal or
the task of the process’’.
Related Work
As pursuit is one of the basic capabilities of an active
vision system, most of the research groups possessing a
camera platform have reported results. We divide the
approaches into two groups. The first group consists of
algorithms that use only motion cues for gaze shifting
and holding, and this is the group to which our system
belongs. Computational basis of this approach group is
the difference between measured optical flow and the
optical flow induced by camera motion.
The Oxford surveillance system [9, 10] uses data from
the motor encoders to compute and subtract the
camera motion-induced flow. It runs in 25 Hz with
processing latency of about 110 ms. Camera behavior is
modeled as either saccadic or pursuit motion. Saccadic
motion is based on the detection of motion in the
coarse scale periphery. Pursuit motion is based only on
the optical flow of the foveal region. This is also the
difference to our system, which can also smoothly
pursue but with repeated motion detection.
Conclusion
We presented a system that is able to detect and pursue
moving objects without knowledge of their form or
motion. The performance of the system with control
rate of 25 Hz, a latency of 80 ms, and average angular
velocities of about 10 degrees per second, is competitive
with respect to the state of the art. The system
needs the minimal number of tuning parameters: a
threshold for normal flow difference, a threshold for the
image gradient, a minimal image area over the mentioned
thresholds, and the maneuvering index.
We have shown that in order to achieve real-time
reactive behavior we must apply the appropriate image
processing and control techniques. The main contribution
of this paper is not only in the achieved high
performance of the system. Our work is different from
other presentations in the study of the individual
components with respect to the given hardware, time
constraints, and desired tracking behavior. We experimentally
studied the responses of the image processing
filters if fixed-point arithmetic is used.