200
N. Blanik
11.5.2.4 Further Tracking Algorithms
Besides the tracking algorithms mentioned above, there exists a wide range of other
approaches, which can be applied to eliminate movement from video sequences while
tracking objects along their trajectory through the image plane. Though they are not
(yet) implemented for concrete PPGI-applications, their individual approaches are
promising to improve the technique of tracking objects.
Firstly, this applies to the above-mentioned SIFT and SURF algorithms. As can be
understood from their names, they are especially suited to compensate scaling effects
of tracked objects, or to decrease required calculation time, respectively [31, 32].
Alternate tracking and filter techniques are derivable from other image and video
processing applications, each approach opening up new and interesting vistas of
application. In this method, the particle filter algorithm is a model and probabilistic
approach [33, 34]. Originally, as the name suggest, it was used to track the distribution
of particles. This model assigns every single particle with a position inside the frame
and tags a particle weight. Again, such particles can be extracted from images by
suitable feature detectors like color or gray level distribution, couture curves or curve
of gradients. For the purpose of tracking objects, the probability distribution of the
particles is predicted by an underlying movement model. The area around the new
estimated position, i.e., the center of the areas of the highest particle density, is
compared to the features extracted from the next frame. If necessary, the estimation
is corrected. Due to the nature of particles, this algorithm facilitates easy tracking of
multiple objects such as particles and particle clouds. However, the quality of motion
detection depends markedly on the number and the kind of used particles/features,
as well as the used motion model. The latter is also crucial for tracking based on
Kalman Filtering, in which the procedure is described as a time-discrete dynamic
problem solvable by a process and a measurement equation [35, 36]:
x t+1 = F · x t + r t
(11.15)
z t = H · x t + v t
(11.16)
The process matrix F determines change of states of the state vector x t to x t+1 r t
considers process noise. The current state z t can be described by the measurement
matrix H and the measurement noise v t . In this model, the current state is tracked by
continuously comparing predictions with the measured states. The classic Kalman
Filter approach is restricted to linear models only. For more complex models (like in
most motion processes), the Extended Kalman Filter can be utilized instead. Among
its benefits, it allows short-time masking of the tracked object. The intrinsic state
vector is updated for each point of time, in contrast to particle filtering, where the
consideration of a multitude of probable positions is not possible.
Finally, without claiming to be exhaustive, the mean shift tracking method is
briefly touched upon here [37]. This method is based on an in-depth study of extreme
values of distribution function that are retrieved from the image plane. In the simplest
N. Blanik
11.5.2.4 Further Tracking Algorithms
Besides the tracking algorithms mentioned above, there exists a wide range of other
approaches, which can be applied to eliminate movement from video sequences while
tracking objects along their trajectory through the image plane. Though they are not
(yet) implemented for concrete PPGI-applications, their individual approaches are
promising to improve the technique of tracking objects.
Firstly, this applies to the above-mentioned SIFT and SURF algorithms. As can be
understood from their names, they are especially suited to compensate scaling effects
of tracked objects, or to decrease required calculation time, respectively [31, 32].
Alternate tracking and filter techniques are derivable from other image and video
processing applications, each approach opening up new and interesting vistas of
application. In this method, the particle filter algorithm is a model and probabilistic
approach [33, 34]. Originally, as the name suggest, it was used to track the distribution
of particles. This model assigns every single particle with a position inside the frame
and tags a particle weight. Again, such particles can be extracted from images by
suitable feature detectors like color or gray level distribution, couture curves or curve
of gradients. For the purpose of tracking objects, the probability distribution of the
particles is predicted by an underlying movement model. The area around the new
estimated position, i.e., the center of the areas of the highest particle density, is
compared to the features extracted from the next frame. If necessary, the estimation
is corrected. Due to the nature of particles, this algorithm facilitates easy tracking of
multiple objects such as particles and particle clouds. However, the quality of motion
detection depends markedly on the number and the kind of used particles/features,
as well as the used motion model. The latter is also crucial for tracking based on
Kalman Filtering, in which the procedure is described as a time-discrete dynamic
problem solvable by a process and a measurement equation [35, 36]:
x t+1 = F · x t + r t
(11.15)
z t = H · x t + v t
(11.16)
The process matrix F determines change of states of the state vector x t to x t+1 r t
considers process noise. The current state z t can be described by the measurement
matrix H and the measurement noise v t . In this model, the current state is tracked by
continuously comparing predictions with the measured states. The classic Kalman
Filter approach is restricted to linear models only. For more complex models (like in
most motion processes), the Extended Kalman Filter can be utilized instead. Among
its benefits, it allows short-time masking of the tracked object. The intrinsic state
vector is updated for each point of time, in contrast to particle filtering, where the
consideration of a multitude of probable positions is not possible.
Finally, without claiming to be exhaustive, the mean shift tracking method is
briefly touched upon here [37]. This method is based on an in-depth study of extreme
values of distribution function that are retrieved from the image plane. In the simplest