Advances in Vision-Based Human Body Modeling - 3D Modeling and Animation: Synthesis and Analysis Techniques for the Human Body - page 12

Human Motion Tracking and 3D Pose Recovery
The majority of model-based human motion tracking techniques may be classi-
fied into two main categories. The first one explicitly poses kinematic constraints
to the model parameters, for example, by means of Kalman filtering or physics-
based modeling. The second one is based on learning the dynamics of low-level
features or high-level motion attributes from a set of representative image
sequences, which are then used to constrain the model motion, usually within a
probabilistic tracking framework. Other subdivisions of the existing techniques
may rely on the type of the model or the type of image features (edges, blobs,
texture) used for tracking.
Tracking relies either on monocular or multiple camera image sequences. This
comprises the classification basis in this subsection. Using monocular image
sequences is quite challenging, due to occlusions of body parts and ambiguity in
recovering their structure and motion from a single perspective view (different
configurations have the same projection). On the other hand, single camera
views are more easily obtained and processed than multiple camera views.
In one of the most recent approaches (Sminchisescu & Triggs, 2001), 3D human
motion tracking from monocular sequences is achieved by fitting a 3D human
body model, consisting of tampered superellipsoids, on image features by means
of an iterative cost function optimization scheme. The disadvantage of iterative
model fitting techniques is the possibility of being trapped in local minima in the
multidimensional space of DOF. A multiple-hypothesis approach is proposed
with the ability of escaping local minima in the cost function. This consists of
observing that local minima are most likely to occur along local valleys in the cost
surface. In comparison with other stochastic sampling approaches, improved
tracking efficiency is claimed.
In the same context, the algorithm proposed by Cham & Rehg (1999) focuses on
2D image plane human motion using a 2D model with underlying 3D kinematics.
A combination of CONDENSATION style sampling with local optimization is
proposed. The probability density distribution of the tracker state is represented
as a set of modes with piece-wise Gaussians characterizing the neighborhood
around these modes. The advantage of this technique is that it does not require
the use of discrete features and is suitable for high-dimensional state-spaces.
Probabilistic tracking such as CONDENSATION has been proven resilient to
occlusions and successful in avoiding local minima. Unfortunately, these ad-
vances come at the expense of computational efficiency. To avoid the cost of
learning and running a probabilistic tracker, linear and linearised prediction
techniques, such as Kalman or extended Kalman filtering, have been proposed.
In this case, a strategy to overcome self-occlusions is required. More details on