Fire Detection Algorithms Using Multimodal ... - Bilkent University

CHAPTER 1. INTRODUCTION 3
No fire detection experiments were carried out using other sensing modalities
such as ultra-violet (UV) sensors, near infrared (NIR) or middle wave-length
infrared (MWIR) cameras, in this study. As a matter of fact, MWIR cameras
are even more expensive than LWIR cameras. Therefore, deploying MWIR cameras
for fire monitoring turns out to be an unfeasible option for most practical
applications.
There are built-in microphones in most of the off-the-shelf surveillance cameras.
Audio data captured from these microphones can be also analyzed along
with the video data in fire monitoring applications. One can develop fire detection
methods exploiting data coming from several sensing modalities similar to
methods described in [31], [85], [88], [24].
1.1 Contribution of this Thesis
The major contributions of this thesis can be divided into two main categories.
1.1.1 Markov Models Using Wavelet Domain Features for
Short Range Flame and Smoke Detection
A common feature of all the algorithms developed in this thesis is the use of
wavelets and Markov models. In the proposed approach wavelets or sub-band
analysis are used in dynamic texture modeling. This leads to computationally
efficient algorithms for texture feature analysis, because computing wavelet coefficients
is an Order-(N) type operation. In addition, we do not try to determine
edges or corners in a given scene. We simply monitor the decay or increase in
wavelet coefficients’ sub-band energies both temporally and spatially.
Another important feature of the proposed smoke and fire detection methods is
the use of Markov models to characterize temporal motion in the scene. Turbulent
fire behavior is a random phenomenon which can be conveniently modeled in a