Fire Detection Algorithms Using Multimodal ... - Bilkent University

Fire Detection Algorithms Using Multimodal ... - Bilkent University Fire Detection Algorithms Using Multimodal ... - Bilkent University

theses.eurasip.org
from theses.eurasip.org More from this publisher
11.07.2015 Views

Chapter 1IntroductionDynamic textures are common in many image sequences of natural scenes. Examplesof dynamic textures in video include fire, smoke, clouds, volatile organiccompound (VOC) plumes in infra-red (IR) videos, trees in the wind, sea andocean waves, as well as traffic scenes, motion of crowds, all of which exhibit somesort of spatio-temporal stationarity. They are also named as temporal or 3-D texturesin the literature. Researchers extensively studied 2-D textures and relatedproblems in the fields of image processing and computer vision [32], [30]. Onthe other hand, there is comparably less research conducted on dynamic texturedetection in video [18], [63], [12].There are several approaches in the computer vision literature aiming atrecognition and synthesis of dynamic textures in video independent of theirtypes [71], [15], [16], [50], [29], [51], [33], [35], [89], [48], [49], [57], [79], [55],[96], [62], [75], [28], [68], [90]. Some of these approaches model the dynamictextures as linear dynamical systems [71], [15], [16], [50], some others use spatiotemporalauto-regressive models [48], [79]. Other researchers in the field analyzeand model the optical flow vectors for the recognition of generic dynamic texturesin video [29], [89]. In this dissertation, we do not attempt to characterizeall dynamic textures but we present smoke and fire detection methods by takingadvantage of specific properties of smoke and fire.1

CHAPTER 1. INTRODUCTION 2The motivation behind attacking a specific kind of recognition problem isinfluenced by the notion of ‘weak’ Artificial Intelligence (AI) framework which wasfirst introduced by Hubert L. Dreyfus in his critique of the so called ‘generalized’AI [25], [26]. Dreyfus presents solid philosophical and scientific arguments on whythe search for ‘generalized’ AI is futile [61]. Current content based general imageand video content understanding methods are not robust enough to be deployedfor fire detection [21], [53], [43], [54]. Instead, each specific problem should beaddressed as an individual engineering problem which has its own characteristics.In this study, both temporal and spatial characteristics related to flames andsmoke are utilized as clues for developing solutions to the detection problem.Another motivation for video and pyroelectric infra-red (PIR) sensor basedfire detection is that conventional point smoke and fire detectors typically detectthe presence of certain particles generated by smoke and fire by ionization orphotometry. An important weakness of point detectors is that they cannot providequick responses in large spaces. Furthermore, conventional point detectorscannot be utilized to detect smoldering fires in open areas.In this thesis, novel image processing methods are proposed for the detectionof flames and smoke in open and large spaces with ranges up to 30m. Flickerprocess inherent in fire is used as a clue for detection of flames and smoke in (IR)video. A similar technique modeling flame flicker is developed for the detectionof flames using PIR sensors. Wildfire smoke appearing far away from the camerahas different spatio-temporal characteristics than nearby smoke. The algorithmsfor detecting smoke due to wildfire are also proposed.Each detection algorithm consists of several sub-algorithms each of which triesto estimate a specific feature of the problem at hand. For example, long distancesmoke detection algorithm consists of four sub-algorithms: (i) slow moving videoobject detection, (ii) smoke-colored region detection, (iii) rising video object detection,(iv) shadow detection and elimination. A framework for active fusion ofdecisions from these sub-algorithms is developed based on the least-mean-square(LMS) adaptation algorithm.

Chapter 1IntroductionDynamic textures are common in many image sequences of natural scenes. Examplesof dynamic textures in video include fire, smoke, clouds, volatile organiccompound (VOC) plumes in infra-red (IR) videos, trees in the wind, sea andocean waves, as well as traffic scenes, motion of crowds, all of which exhibit somesort of spatio-temporal stationarity. They are also named as temporal or 3-D texturesin the literature. Researchers extensively studied 2-D textures and relatedproblems in the fields of image processing and computer vision [32], [30]. Onthe other hand, there is comparably less research conducted on dynamic texturedetection in video [18], [63], [12].There are several approaches in the computer vision literature aiming atrecognition and synthesis of dynamic textures in video independent of theirtypes [71], [15], [16], [50], [29], [51], [33], [35], [89], [48], [49], [57], [79], [55],[96], [62], [75], [28], [68], [90]. Some of these approaches model the dynamictextures as linear dynamical systems [71], [15], [16], [50], some others use spatiotemporalauto-regressive models [48], [79]. Other researchers in the field analyzeand model the optical flow vectors for the recognition of generic dynamic texturesin video [29], [89]. In this dissertation, we do not attempt to characterizeall dynamic textures but we present smoke and fire detection methods by takingadvantage of specific properties of smoke and fire.1

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!