In this paper a novel background removal technique has
been presented. After dividing an image into a set of 4 × 4
pixel patches, a model for every patch is constructed and
adapted with every new image taken from the camera. A
model consists of a set of several coefficient vectors, which
are obtained by applying a discrete cosine transform to a
corresponding patch. Since only coefficients are used which
correspond to the low frequency components more emphasis
is put on the coarse structure of the patch. Due to this
fact the method was shown to be very robust to illumination
changes.
To reduce the computational burden the amount of vectors
representing the model was not fixed and was adapted
from patch to patch online. In the case of non-stationary
background more model vectors were accumulated. However
if a patch represented a static background just one or
two vectors were sufficient.
Furthermore the presented method also incorporated the
temporal and spacial characteristics of an object motion. To
achieve a low false positive rate an object was detected as
foreground only if its neighbors deviated significantly from
the background in previous images. This procedure let to a
great reduction of noise and false positives.
The presented method was tested in several scenarios
showing its good performance. The system has been successfully
used to track people in indoor environment and
people and cars in outdoor environments. Experiments with
waving plants showed its capability of incorporating nonstationary
background objects into the background model.
Because of its low computational complexity the presented
system was used in real time video surveillance.