以全域及區域特徵基礎之煙霧偵測

Abstract

本篇論文提出了使用區域特徵分析及全域特徵驗證的煙霧偵測方法，調查中指出近年來基於影像式的煙霧偵測技術在智慧型監控系統中受到廣泛的重視與研究。然而，在一個廣大的開放空間處理煙霧偵測事件如何不被其他常見的干擾物例如行人和車輛所影響，建立一個無誤報的煙霧偵測系統仍是一項具有挑戰性的問題。因此能夠在不同的環境下仍然能找出可區別煙霧以及非煙霧的特徵是一件重要的任務。本篇論文分析影片中每個候選區塊的區域特徵：邊緣模糊化、能量的逐步變化與色彩結構的逐步變化，其中每個區域特徵都對煙霧有足夠的偵測能力以及低誤報。此外，本論文所提出的三種類型的區域特徵都各具互補性，因此，藉由Boosting學習演算法加上串聯式架構的方式結合區域特徵以降低更多的誤報。為了更進一步的克服誤報的情況，提出了全域特徵的統計方式來驗證候選區域的邊緣地帶及整個感興趣移動區域內的資訊。實驗結果指出本篇論文所提出的系統對於不同的環境地點在煙霧偵測上有良好的偵測率以及很低的誤報率以及快速的反應時間。整個系統在執行上具有高效率可對影像做即時處理，並已將此煙霧偵測技術移植到嵌入式系統中。

This study presents a novel smoke detection approach using local feature analysis and global feature verification. Studies have investigated visual-based smoke detection techniques in surveillance systems for years. However, given an image in open or large spaces with typical smoke and disturbances of commonly moving objects such as pedestrians or vehicles, detecting smoke without false alarm is still a challenging problem. It is important to find features to distinguish smoke from various environments. This study analyzes characteristics of candidate blocks in video sequences to exploit local features: edge blurring, gradual energy change and gradual color configuration change. Each local feature is strong enough to detect smoke with few false alarms. Moreover, proposed features are complementary to each other. Hence, local features are combined to lower the false alarm rates by boosting cascade architecture. To further overcome some false situation, global feature verification is proposed to gather statistics of information on contour and in the whole area of each candidate region. Experimental results show that the proposed system can well detect smoke with low false alarm rate within a short reaction time in various environments. The whole system can run in real time and has been implemented on embedded system.