Fire Detection Algorithms Using Multimodal ... - Bilkent University

CHAPTER 2. FLAME DETECTION IN VISIBLE RANGE VIDEO 9
There are several video-based fire and flame detection algorithms in the literature
[64], [17], [48], [78], [38], [81], [80], [94]. These methods make use of
various visual signatures including color, motion and geometry of fire regions.
Healey et al. [38] use only color clues for flame detection. Phillips et al. [64]
use pixel colors and their temporal variations. Chen et al. [17] utilize a change
detection scheme to detect flicker in fire regions. In [78], Fast Fourier Transforms
(FFT) of temporal object boundary pixels are computed to detect peaks
in Fourier domain, because it is claimed that turbulent flames flicker with a characteristic
flicker frequency of around 10 Hz independent of the burning material
and the burner in a mechanical engineering paper [1], [14]. We observe that flame
flicker process is a wide-band activity below 12.5 Hz in frequency domain for a
pixel at the boundary of a flame region in a color-video clip recorded at 25 fps
(cf. Fig. 2.1). Liu and Ahuja [48] also represent the shapes of fire regions in
Fourier domain. However, an important weakness of Fourier domain methods is
that flame flicker is not purely sinusoidal but it’s random in nature. Therefore,
there may not be any peaks in FFT plots of fire regions. In addition, Fourier
Transform does not have any time information. Therefore, Short-Time Fourier
Transform (STFT) can be used requiring a temporal analysis window. In this
case, temporal window size becomes an important parameter for detection. If
the window size is too long, one may not observe peakiness in the FFT data. If
it is too short, one may completely miss cycles and therefore no peaks can be
observed in the Fourier domain.
Our method not only detects fire and flame colored moving regions in video but
also analyzes the motion of such regions in wavelet domain for flicker estimation.
The appearance of an object whose contours, chrominance or luminosity values
oscillate at a frequency higher than 0.5 Hz in video is an important sign of the
possible presence of flames in the monitored area [78].
High-frequency analysis of moving pixels is carried out in wavelet domain in
our work. There is an analogy between the proposed wavelet domain motion analysis
and the temporal templates of [21] and the motion recurrence images of [43],
which are ad hoc tools used by computer scientists to analyze dancing people
and periodically moving objects and body parts. However, temporal templates