1.1 Motivation
In last few years, there were average 5622.8 fire accidents per year according to the statistic report from the National Fire Administration. The number of dead and injured people was nearly 700 and the property loss was about 2 billion NT dollars each year. If the fire accident could be found much earlier, it is more likely to reduce the loss of life.
The process of fire development mostly divided into four periods: Ignition, Fire Growth Period, Fully Developed Period and Decay Period as shown in Fig.1-1.
Fig. 1-1 Processing diagram of fire development
In general occurrences of fire, a great quantity of thick smoke instead of fire is produced in the initial stage. After flashover, fire spreads quickly and burns all spaces continuously. If people don’t escape from the scene of a fire before flashover, they
probably wouldn’t save their life. Therefore, the duration of flashover is the prime time for people to flee from fire.
Conventional point-based smoke and fire detectors typically detect the presence of certain particles generated by smoke and fire by ionization or photometry. An important weakness of point detectors is that in large space, it may take a long time for smoke particles to reach a detector and they can’t be operated in open spaces such as hangers, tunnels, storage, and offshore platform.
Owing to the limitation of the traditional concept, point-based detectors can’t detect fires or smokes in early stage. In recent years, many researches are devoted to video smoke detection that doesn’t rely on proximity of smoke to the detector. This enables it to incorporate standard video surveillance cameras with sophisticated image recognition and processing software to identify the distinctive characteristics of smoke patterns. In most cases, smoke usually appears before ignition. Therefore, the beginning of fire can be observed soon before it causes any real damage.
1.2 Related Work
Fig. 1-2 Four categories of video smoke detection
There are four categories of video smoke detection in the literature as shown in Fig.
1-2. The first category is Motion-Based approaches. Kopilovic et al. [1] observed that the irregularities in motion due to non-rigidity of smoke. They apply a multiscale optical flow computation and the entropy of the motion distribution in Bayesian classifier to detect the special motion of smoke. In order to save computational time, Yuan [2] proposed a fast orientation model that produces more effective way to extract the motion characteristics. Although significant advances have been made in the development of this work, their adoption in general surveillance systems is not widely reported.
The second one is Appearance-Based approaches. Toreyin et al. [3] indicated that smoke of an uncontrolled fire expands in time which results in regions with convex boundaries. Chen [4] found that airflows will make the shape of smoke to be variously changed at any time. Therefore, a disorder measure, the ratio of circumference to area for the extracted smoke region, is introduced to analyze shape complexity. Growth rate is obtained by increment of smoke pixels due to the diffusion process existed in generation of smoke. Two thresholds are determined by the statistical data of experiments to verify the real smoke; furthermore, the changing unevenness of density distribution is proposed in [5]. The difference image provides a natural way to represent the attribute which has more internal information in smoke frames than non-smoke frames. While much research has been devoted to these techniques, few studies have investigated the situation that smoke and non-smoke objects exist in the same time and the presence of moving objects from the outside of video scenes.
The third one is Color-Based approaches. Smoke usually displays grayish colors during the burning process [4]. Two thresholds of I (intensity) component of HSI color space depend on statistical data and this implies that three components R, G and
B of the smoke pixel are equal or so. Since smoke color can’t be represented accurately by a single unimodal, the 3D joint probability density function can be decomposed in three marginal unidimensional distributions over each color axis to accommodate different ranges of color [6]. Independent of the fuel type, smoke naturally decreases the chrominance channels U and V values of the candidate region [3]. In spite of the early alarm capability, few experimental results have been conducted in the range of grayish or dull non-smoke objects.
The fourth one is Energy-Based approaches. It is well-known that wavelet coefficients contain the high frequency information of the original image [7]. Since smoke obstructs the texture and edges in the background of an image [3], a decrease in wavelet energy is an important clue for smoke detection. Piccinini et al. [8] further improved the concept by on-line modeling the ratio between the current input frame energy and the background energy. This method performs well in many real cases but needs long reaction time and more exact validation of the input data extracted from surveillance systems that operate 24 hours a day.
1.3 Thesis Organization
The remainder of this thesis is organized as follows. Chapter 2 describes system overview. Chapter 3 shows smoke detection algorithm including background modeling, candidate selection, feature extraction, classification and verification. Chapter 4 shows experimental results and comparisons. Finally, the conclusions of this system and future work will be presented in chapter 5.