People are the centre of technologies. Understanding, monitoring and
tracking the behaviour of people will benefit in various areas including
driving assistance, surveillance for safety and caring purposes and applications
for machine-people interaction. Particularly, pedestrians attract
more attention for two reasons: they restrict the behaviours of people to
standing and moving upright; and applications for pedestrian detection
and monitoring have positively impact on the quality of life. Pedestrian
detection and identification, aims at recognising pedestrians fromstill images
and video frames. Together with pedestrian recognition and tracking,
this topic attempts to train computers to recognise a pedestrian.
The problem is challenging. Though frameworks were designed, various
algorithms were proposed in recent years, further efforts are needed to
improve the accuracy and reliability of the performance. In this thesis,
proposing a modifiable framework for pedestrian identification and improving
the performances of current pedestrian detection techniques are
particularly focused. Based on appearance based pedestrian identification,
a modifiable framework is a novel philosophy of developing frameworks
which can be easily improved. For pedestrian identification, a novel
protocol where layers of algorithms are hierarchically applied to solve the
problem. To compare the detected pedestrians, appearance based features
are selected, the "bag-of-features" framework is employed to compare
the histogram descriptions of pedestrians. To improve the performances
of HOG pedestrian detector, the presence of head-shoulder structure
is selected as the evidence of the presence of pedestrian. A novel
appearance based framework is developed to detect the head-shoulder
structure from the detection results of HOG detector. Furthermore, in
order to separate multiple pedestrians detected in one bounding box, a
novel algorithm is proposed to detect the approximated symmetry axes of
pedestrians.