基於連續影像技術的煙霧偵測

Abstract

煙霧是大部分火災發生的先兆，因此選擇一個適切有效的煙霧偵測方法，是即早發報火警的基本措施。基於連續影像的火警系統已普遍建立於影像監控系統中。然而，想要於存在各干擾(如行人、汽車)的開放式環境下，快速地偵測煙霧，同時不產生誤報，一直是在影像式煙火偵測領域的一大挑戰。因此，改善影像式煙霧偵測的兩大目標為：1. 縮短反應時間，2. 減低誤報的機率 本研究將目標投注在藉由分析可線性分割煙霧與其它物體的特徵，進而迅速地偵測煙霧並排除誤報的情況。本論文提出了三個可線性分割的特徵，並介紹一個簡單實用的設計框架，使用時間、空間的分析來設計基於連續性影像處理技術的偵測系統。本研究以時間、空間上的分析，萃取三種重要的煙霧特徵：邊緣模糊化、緩慢地能量變化、以及緩慢地色澤結構變化。本研究提出的時空分析技術增進了對煙霧緩慢變化的特徵的解析程度。更甚者，本研究所提出之基於統計區域性特徵的全域驗證機制，亦進一步地降低系統誤報的機率。 測試本研究所提出之演算效能的PC平台為Intel® Core™2 Duo CPU (2.2 GHz) and 2 GB RAM，每張影像的平均處理時間為32.27 ms，相當於每秒30.98張影像，實驗結果顯示，本研究提出之系統於各種真實環境中，可在誤報很少的情況下迅速有效地偵到測煙霧。

Smoke is an early sign of most fires; therefore, selecting an appropriate smoke-detection method is essential. Video-based fire detection is currently a fairly common application with the growth in the number of installed surveillance video systems. However, fast detecting smoke without creating a false alarm remains a challenging problem for open or large spaces with the disturbances of common moving objects, such as pedestrians and vehicles. Hence, two important targets for improving video-based smoke detector are the short reaction time and the low false alarm rate. This thesis aims to fast detect smoke while eliminate false alarm cases by analyzing features which is linearly separable. This thesis proposed three linear separable features and also introduces a framework for design a simple Video Technology-based system with temporal and spatial analysis. In this study, the process of extracting smoke features from candidate regions was accomplished by analyzing the spatial and temporal characteristics of video sequences for three important features: edge blurring, gradual energy changes, and gradual chromatic configuration changes. The proposed spatial-temporal analysis technique improves the feature extraction of gradual energy changes. Moreover, this thesis proposed a global verification stage based on the statistic of calculated local features to further lower the false alarm rate. The effectiveness of the proposed algorithm was evaluated on a PC with an Intel® Core™2 Duo CPU (2.2 GHz) and 2 GB RAM. The average processing time was 32.27 ms per frame, i.e., the proposed algorithm can process 30.98 frames per second. Experimental results showed that the proposed system can detect smoke effectively with a low false-alarm rate and a short reaction time in many real-world scenarios.