Abstract:
Indoor occupancy system has now become dispensable in our lives. The use of this information
is not only limited to the recording of human activities but also serves other applications such as
power control automation and space utilisation examination. A robust system with high accuracy
could also play an essential role in catastrophic situations by assisting the firefighter team in the
evacuation. To achieve this, the convergence between the sensing technology and algorithms should
achieve real-time processing speed with high detection accuracy, and even tracking or prediction
on an individual. High resistance to most environmental factors which could potentially affect the
general consistency is also vital. The system must also show high feasibility and low cost for general
public use, and take privacy into consideration for those under surveillance.
In this thesis, we explore possible solutions to achieve real-time implementation on occupancy
system through Deep Learning and Millimetre wave technology. First, we developed a new unsupervised
learning algorithm SVDBSCAN to achieve a highly accurate ground truth extraction from
raw data, to tackle the uneven density issue existed in general radar sensor. Second, a new function
was proposed to develop an artificial database for Millimetre wave sensor to make up the lack of
public dataset, thereby making the neural network implementation possible. Next, we proposed a
unique data transformation method from the inspiration of an ordinary RGB photograph to serve
as model input. Upon the implementation of an appropriate CNN skeleton, the output result is
more deterministic and time predictable. Finally, system optimization, tracking and prediction were
achieved through a Kalman Filter, which also increases the system robustness and ability to handle
signal attenuation.
Overall, the result of the system proposed in this thesis has demonstrated the ability of real-time
processing with high accuracy, with the ability to handle a multi-object situation and also potential
signal loss. The approach of CNN also provides the ability for features expansion in the future when
public dataset became available and more matured. We hope the contribution of this project could
make the utilisation of this system one step closer.