基於時空分析之煙霧偵測系統

Abstract

大多數火災發生時都會伴隨產生煙霧，若能在災害擴大前發出警報是一項重要的工作。因此，近來影像式之煙霧偵測技術在智慧型安全監控領域是一項研究的重點，而如何能夠在最短的時間內偵測出煙霧以及有效的辨別非煙霧的物體、減少誤報機率是此研究必須克服的問題。 本研究利用連續影像上的時空分析，分別針對時域、空域進行高通與低通的運算，並產生八種不同的組合，藉此分析出可代表煙霧在時間上緩慢移動、空間上變化平緩之區域型特徵，並能夠與其它非煙霧物體做有效的線性區隔。考量在有限的實際火災事件影片狀況下，為了獲得更佳的一般性，我們採用基於支持向量機(Support Vector Machines)的分類器將八項特徵做結合。接下來透過面積比例、面積擴散、成分一致性三項全域性驗證，降低更多誤報。實驗結果表現出本篇論文所提出的系統在煙霧偵測上有很低的誤報率以及快速的反應時間，並且可以在無建立背景模型的情況下有效的偵測煙霧，不易受攝影環境變動而有所影響以達到在實際防災應用上的穩健性與可靠性。

Smoke always accompanies with fires as a early sign. Hence, it's practical to send alerts to prevent fire disasters. Visual-based smoke detection techniques in surveillance systems have been studied for years. However, there are some challenging problems to detect smoke within a short reaction time, recognize non-smoke objects correctly and decrease the false alarm rate.

This study presents temporal spatial and spatio-temporal analysis on image sequences. High-pass filter and low-pass filter are exploited on both domain to generate eight features by cross combinations. This approach can extract information that smoke moves swiftly with time and changes slowly on spatial domain. In order to obtain the proper generalization ability with respect to sparse training samples, a support vector machine is to combine the eight features as a classifier. The global verifications including area ratio, area spread and spatial variance is used to decrease false alarm.

Experimental results show the false alarm rate is lower and the reaction time is shorter than other approaches. The proposed approach in this study can detect smoke without background modeling and provide better performance in the field of smoke detection.