The development and utilization of low-cost near infrared imagers for fire detection and diagnostics

This research explores the use of near infrared spectrum for video fire
detection and combustion diagnostics. The near infared spectrum is
somehow very under-explored in these fields. A low-cost silicon-based
backside illuminated CMOS image sensor was modified to become a
monochrome sensor exposing its full spectral sensitivity. The sensor was
tested to show its improved spectral sensitivity. The multi-spectrum
fire detection combines stereo cameras with NIR only and NIR-visible
spectrum respectively for robust fire detection. Flame image properties
in both conditions are extensively studied, whereby the NIR-only channel
corresponds to ROI extraction and texture feature extraction; the NIRvisible channel give rise to a unique colour model for false positive
classification rejection. Machine learning algorithms were employed for
fire recognition. Practical considerations of designing a fire detection
system have been discussed in terms of sensor selection, feature extraction
as well as choice of classification algorithms. Dual-band stereo video fire
detection not only showed great potential for robust fire detection, but
also for vision-based automated firefighting. In combustion diagnostics,
low-cost NIR sensors were used for the imaging of combustion products
of fuel-lean premixed flames, which demonstrated its effectiveness in
potential flame instability related diagnostics. Moreover, the thermalcapability of NIR camera sensor were applied in conjunction with visible
and schlieren imaging, to study the mechanism of fire propagation on
solid fuel combustion. Finally, a summary was made with additional
suggestions and speculations on the subject as a whole.