In this talk, I will present the tracking system we developed at the CVLAB,
EPFL. Our system is able to reliably track multiple people in a multi-camera
setting. The obtained trajectories can be used for understanding individuals
and
group behavior. There are numerous applications for such a system; we are
currently involved in a project whose goal is to understand the behavior of
basketball teams and players from video cameras.


Our system represents the ground floor of the scene, as a grid of cells. The
goal of the system is to estimate, at each time step, which grid cells are
occupied and by whom. The system is composed of three core components:
detection, identification and tracking. Detection is based on a generative
model
which can effectively handle occlusions in each time frame independently. This
produces what we call a Probability Occupancy Map (POM) which is then used by
the next components. The identification component recognizes the identity of
the
person according to his color histogram, facial descriptor and, in the case of
sport matches, his jersey number. In the final tracking component, the
multi-person tracking problem is formulated as a multi-commodity network flow
problem. The tracker links the detections of people in individual frames across
time, while taking into account the appearance and identity constraints.


The full system demonstrates excellent results on long and challenging video
sequences, including a pedestrian benchmark dataset and several sports
datasets.
We show that our system works reliably in spite of significant occlusions and
delivers metrically accurate trajectories for each tracked individual.