Real-time people tracking in a camera network
Abstract
Visual tracking is a fundamental key to the recognition and analysis of human behaviour.
In this thesis we present an approach to track several subjects using multiple
cameras in real time. The tracking framework employs a numerical Bayesian estimator,
also known as a particle lter, which has been developed for parallel implementation on
a Graphics Processing Unit (GPU). In order to integrate multiple cameras into a single
tracking unit we represent the human body by a parametric ellipsoid in a 3D world.
The elliptical boundary can be projected rapidly, several hundred times per subject per
frame, onto any image for comparison with the image data within a likelihood model.
Adding variables to encode visibility and persistence into the state vector, we tackle the
problems of distraction and short-period occlusion. However, subjects may also disappear
for longer periods due to blind spots between cameras elds of view. To recognise
a desired subject after such a long-period, we add coloured texture to the ellipsoid surface,
which is learnt and retained during the tracking process. This texture signature
improves the recall rate from 60% to 70-80% when compared to state only data association.
Compared to a standard Central Processing Unit (CPU) implementation, there
is a signi cant speed-up ratio.