摘要
在火災救援工作中，除了滅火和救援工作外，戰術指揮及火場情況判斷也非常重要。在現今技術發展下，許多實驗都有做過很多關於火焰的偵測，不管是單純火焰影像偵測、室內火災偵測或森林火災偵測等，都希望能透過AI技術幫助火災救援能更早發現，或是提供更多資訊給消防指揮中心進行決策。 室內火災發生時，並不單純是建築物是否在燃燒及火勢大小判斷，而是透過室內火災發生時會有初期、成長期、旺盛期、衰退期四個階段，且四個階段都有不同搶救的重點，為此，本研究希望透過影像辨識室內火災階段的判別，因此使用了YOLO v7演算法且其對於火災四階段之mAP達88.1%，但在使用新驗證影片供給YOLO v7測試時，會發現火災階段不能穩定判斷的情形發生，因此想透過長短期記憶神經網路(LSTM)其時間序特性使火災階段判斷更為穩定，從而結合兩者優點衍生出了YoLstm-FSN(YOLO-LSTM Fire Stage Net)模型。 為了證明YoLstm-FSN模型相較於YOLO模型及LSTM模型是否能更穩定判斷火焰階段，本研究三個模型進行相同室內火災燃燒影片火焰階段判斷，並使用混淆矩陣計算其模型準確率、精確度、召回率、F1-Score，已能評估整體模型優劣之F1-Score為例，YOLO模型之F1-Score為0.51，LSTM模型之F1-Score為0.31，YoLstm-FSN模型之F1-Score為0.86。因此，YoLstm-FSN模型對於火災階段辨識的準確性是優於其他兩者，能更準確提供當下火焰發生階段的訊息，從而使消防指揮中心戰術計畫有所幫助，以減少救災傷亡。 關鍵字： 消防救援、火災四階段、YOLO物件辨識模型、LSTM模型、YoLstm-FSN模型
並列摘要
In the fire scene rescue not only to put out the fire and rescue, tactical command and fire scene situation judgment is also very important, in today's technology development, many experiments have done a lot of flame detection, whether it is a simple flame image detection, indoor fire detection or forest fire detection, etc., all hope to help the fire rescue through AI technology can be found earlier, or provide more information to the fire command center for decision making. In this study, we used YOLO v7 algorithm to identify the stages of indoor fires through images, and its mAP for the four stages of fires reached 88.1%. However, when using the newly verified video for YOLO v7 testing, it was found that the fire stage could not be judged stably. Therefore, the YoLstm-FSN(YOLO-LSTM Fire Stage Net) model was derived by combining the advantages of the two by using the temporal sequential characteristics of the long and short-term memory neural network (LSTM) to make the fire stage judgment more stable. In this study, in order to prove whether the YoLstm-FSN model is more stable than YOLO and LSTM, the three models were used to determine the flame stage of the same indoor fire burning video, and their model accuracy, precision, recall, and F1-Score were calculated using confusion matrix. The F1-Score of the YOLO model is 0.51, the F1-Score of the LSTM model is 0.31, and the F1-Score of the YoLstm-FSN model is 0.86. Therefore, the stability of the YoLstm-FSN model for fire stage identification is better than the other two models, and it can provide more accurate information about the flame stage, which can help the tactical planning of the fire command center. Keywords: fire rescue, fire stage 4, YOLO object recognition model, LSTM model, YoLstm-FSN model