Video-based people tracking is extremely challenging, especially when
using one single camera. Among the sources of difficulties are joint
reflection ambiguities, occlusions, cluttered backgrounds, non- rigidity
of tissue and clothing, complex and rapid motions, and poor image
resolution.

In this talk, I will present two complementary approaches to addressing
these issues. The first relies on using spatio-temporal templates to
detect key poses at regular intervals and then to link these detections
into complete trajectories, which results in fully automated tracking.

The second starts from the observation that a common theme in many
recent approaches to capturing human motion from video is to represent
the set of likely poses as a low-dimensional manifold parameterized by a
few latent variables. This mapping, however, is usually learned in a
problem-independent way, which makes it difficult to recover motion
without manually initializing the latent variables and the pose.
Therefore a direct way to derive a mapping between easily observable
image quantities, which serve as latent variables, and pose sequences is
introduced. This also yields full automation but, in addition, only
requires modest amounts of training data.