Fire Detection Algorithms Using Multimodal ... - Bilkent University
Fire Detection Algorithms Using Multimodal ... - Bilkent University Fire Detection Algorithms Using Multimodal ... - Bilkent University
CHAPTER 7. CONCLUSION AND FUTURE WORK 107- short-range (< 30m) flame detection in long-wave infra-red (LWIR) video,- short-range smoke detection in visible range video,- short-range flame detection using pyro-electric infra-red (PIR) sensor, and- long-range (> 100m) smoke detection in visible range video.Each type of sensor has its own advantages and disadvantages on visualizingdifferent aspects of fire. Flames are visible when viewed by visible and IR sensors.Although flame flicker becomes hard to observe as distance to the location of fireincreases. As a result, methods for short-range flame detection by IR and visiblesensors are proposed. Gradual increase in temperature can be monitored with IRcameras and PIR sensors which can be helpful to take precautions even beforea fire starts. However, smoke, which is an early indicator of smoldering fires, istransparent for LWIR cameras and PIR sensors. Consequently, smoke detection isnot possible using IR sensors. Smoke detection, both short and long-range, is onlypossible with visible range cameras. Since long-range smoke has different spatiotemporalcharacteristics than nearby smoke, separate algorithms are developedfor each case.One common novelty that is introduced in all of the methods developed forfire detection using several types of sensors is that flicker process in fire is characterizedusing Markov models trained with wavelet based feature signals. Flickerwas not thoroughly characterized by previous studies on video based fire detection,which resulted in higher false alarm rates. Flicker modeling in a Markoviansetting suits well with the stochastic nature of fire. Besides, sub-band analysis (orwavelet analysis) extracts non-stationarity inherent in flames and smoke. The useof wavelet transform based feature signals, leads to the development of computationallyefficient algorithms that work in real-time running on a standard PC.Another important contribution of the thesis is the introduction of a novel settheoretic framework for decision fusion and on-line active learning. As describedthroughout the thesis, all of the proposed algorithms are composed of several subalgorithmsyielding their own decisions and confidence values about the observed
CHAPTER 7. CONCLUSION AND FUTURE WORK 108phenomenon. In the last stage of each detection algorithm, individual decisionsfrom sub-algorithms are combined together to reach a final decision. A leastmean-square(LMS) based active decision fusion strategy is proposed and thisframework is successfully applied in the decision fusion step of the long-rangewildfire detection method.The LMS adaptive decision fusion framework takes into account the feedbackfrom an oracle (security-guard) whose decisions are taken as the ground-truth.Individual sub-algorithm weights are updated in such a way that the final decisiontracks the classification results of the oracle at each time step. At each stage ofthe LMS algorithm, the method tracks the changes in the nature of the problemby performing an orthogonal projection onto a hyperplane describing the decisionof the oracle. Therefore, the proposed framework for decision fusion may provideinsight in problems with drifting concepts belonging to many other research areasin machine learning applications including computer vision.The proposed wildfire detection method is integrated into forest fire earlywarning systems which are installed in ten forest watch towers in Turkey. Oursystem successfully did not miss any smoke due to test fires. It also detected threeforest fires in their early stages in Antalya and Aksehir regions in the summerof 2008. The system is developed to assist the guard in fire detection. TheDirectorate of Forestry and Ministry of Environment and Forestry of Turkey willinstall the system to other forestal areas under high risk of fire in upcoming years.In this thesis, we developed signal and image processing algorithms for firedetection dedicated to sensing modalities like visible range and LWIR cameras,and PIR senors. In the proposed setting, just a single type of sensor is assumedto monitor an area of interest for each sensing modality.A natural future direction pertains to extending both the number and typesof sensors covering an indoor or an outdoor area of interest. This extensionrequires the development of algorithms that efficiently extract useful informationfrom multitude of sensors and interpret them to reach a final decision. The sameanalysis algorithms may be used for each sensor. However, overall decision canbe given by linearly combining sensor outputs that define confidence levels for
- Page 75 and 76: CHAPTER 4. SHORT RANGE SMOKE DETECT
- Page 77 and 78: CHAPTER 4. SHORT RANGE SMOKE DETECT
- Page 79 and 80: CHAPTER 4. SHORT RANGE SMOKE DETECT
- Page 81 and 82: CHAPTER 4. SHORT RANGE SMOKE DETECT
- Page 83 and 84: CHAPTER 4. SHORT RANGE SMOKE DETECT
- Page 85 and 86: CHAPTER 4. SHORT RANGE SMOKE DETECT
- Page 87 and 88: Chapter 5Flame Detection Using PIRS
- Page 89 and 90: CHAPTER 5. FLAME DETECTION USING PI
- Page 91 and 92: CHAPTER 5. FLAME DETECTION USING PI
- Page 93 and 94: CHAPTER 5. FLAME DETECTION USING PI
- Page 95 and 96: CHAPTER 5. FLAME DETECTION USING PI
- Page 97 and 98: CHAPTER 5. FLAME DETECTION USING PI
- Page 99 and 100: CHAPTER 5. FLAME DETECTION USING PI
- Page 101 and 102: CHAPTER 6. WILDFIRE DETECTION 82In
- Page 103 and 104: CHAPTER 6. WILDFIRE DETECTION 84are
- Page 105 and 106: CHAPTER 6. WILDFIRE DETECTION 86D 1
- Page 107 and 108: CHAPTER 6. WILDFIRE DETECTION 88tha
- Page 109 and 110: CHAPTER 6. WILDFIRE DETECTION 90The
- Page 111 and 112: CHAPTER 6. WILDFIRE DETECTION 92D 1
- Page 113 and 114: CHAPTER 6. WILDFIRE DETECTION 94as
- Page 115 and 116: CHAPTER 6. WILDFIRE DETECTION 96Sol
- Page 117 and 118: CHAPTER 6. WILDFIRE DETECTION 98as
- Page 119 and 120: CHAPTER 6. WILDFIRE DETECTION 100Un
- Page 121 and 122: CHAPTER 6. WILDFIRE DETECTION 102is
- Page 123 and 124: CHAPTER 6. WILDFIRE DETECTION 104Ta
- Page 125: Chapter 7Conclusion and Future Work
- Page 129 and 130: Bibliography[1] B. Albers and A. Ag
- Page 131 and 132: BIBLIOGRAPHY 112[18] D. Chetverikov
- Page 133 and 134: BIBLIOGRAPHY 114[38] G. Healey, D.
- Page 135 and 136: BIBLIOGRAPHY 116[56] U. Niesen, D.
- Page 137 and 138: BIBLIOGRAPHY 118[77] C. Stauffer an
CHAPTER 7. CONCLUSION AND FUTURE WORK 107- short-range (< 30m) flame detection in long-wave infra-red (LWIR) video,- short-range smoke detection in visible range video,- short-range flame detection using pyro-electric infra-red (PIR) sensor, and- long-range (> 100m) smoke detection in visible range video.Each type of sensor has its own advantages and disadvantages on visualizingdifferent aspects of fire. Flames are visible when viewed by visible and IR sensors.Although flame flicker becomes hard to observe as distance to the location of fireincreases. As a result, methods for short-range flame detection by IR and visiblesensors are proposed. Gradual increase in temperature can be monitored with IRcameras and PIR sensors which can be helpful to take precautions even beforea fire starts. However, smoke, which is an early indicator of smoldering fires, istransparent for LWIR cameras and PIR sensors. Consequently, smoke detection isnot possible using IR sensors. Smoke detection, both short and long-range, is onlypossible with visible range cameras. Since long-range smoke has different spatiotemporalcharacteristics than nearby smoke, separate algorithms are developedfor each case.One common novelty that is introduced in all of the methods developed forfire detection using several types of sensors is that flicker process in fire is characterizedusing Markov models trained with wavelet based feature signals. Flickerwas not thoroughly characterized by previous studies on video based fire detection,which resulted in higher false alarm rates. Flicker modeling in a Markoviansetting suits well with the stochastic nature of fire. Besides, sub-band analysis (orwavelet analysis) extracts non-stationarity inherent in flames and smoke. The useof wavelet transform based feature signals, leads to the development of computationallyefficient algorithms that work in real-time running on a standard PC.Another important contribution of the thesis is the introduction of a novel settheoretic framework for decision fusion and on-line active learning. As describedthroughout the thesis, all of the proposed algorithms are composed of several subalgorithmsyielding their own decisions and confidence values about the observed