Trac accidents involving pedestrians are
a very serious matter as it has one of
the highest risks to end with fatalities.
Therefore, a program is needed to help
detect pedestrians in a trac scenario. In
addition the system will be designed to be
able to handle the most common occlusion
problem, which is the lower part of the
body and is therefore designed as an upper
body detector.
The proposed upper body detector in this
project is a CNN based system with three
CNNs in a cascade. This is done in order to
not only have a good performance, but also
a system which is computationally lighter
than a single deep CNN for possible real life
scenario use. The cascade consist rstly of
a very shallow CNN followed by a not as
shallow CNN and lastly a deeper CNN for
nal predictions.
After the conducted experiments, it is
shown that the performance of this upper
body detector does not perform as well as
a full-sized pedestrian detector. This was
expected as the upper body is a subset of
the full-sized pedestrian, but the results
are bad enough to not be usable in a
real life scenario. For that to happen
further research on this topic is needed
for improving on the feature extraction of
upper bodies.