Alarm Decision Unit

n xH n

 



(3.34)

This feature is used to accumulate the fire surface area change rate over a period of time. n is representative of the amount of data, and  feature is inversely proportional to the fire probability.

3.9 Alarm Decision Unit

Image-based detection system often receive the wrong image information to the camera due to external influences, such as the shaking or change in brightness caused by strong winds or the bus passing by. Although the image is returned to normal after a short time, this type of wrong information can cause the detection system’s momentary judgment error. However, this type of alarm is not accurate and instantaneous.

In order to make the entire detection system more reliable and widespread, these sudden false alarms must be filtered out. Because f will not suddenly be produced or suddenly disappear, we can use the alarm buffer to eliminate this type of situation, as demonstrated in the figure. The buffer collects all the system produced false alarms from the image sequence and ratio of alarm data and the non-alarm data in this time period.

Fig. 3-19 Alarm decision unit

29

If the alarm ratio is over 50% of the time interval, then a real alarm is issued. This method passes the signal through the buffer to produce the real fire alarm, and can handle the noise of the camera caused by sudden external influence and reducing false alarm reports. Clearly, this method requires adding a little reaction time, but the system reliability and stability is beyond doubt. Below picture is a real fire alarm produced by the buffer

Fig. 3-20 Final fire alarm

30

4 Chapter 4 Experimental Results

In this chapter, several results of fire detection will be presented. Our algorithm was implemented on the platform of PC with Intel Core i5 2.53GHz and 2GB RAM.

Borland C++ Builder is our development tool and operated on Windows 7. All of our testing inputs are uncompressed AVI video files and DVD video data acquired by USB video capture. The resolution of video frame is 320*240.

In section 4.1, we will show the experimental results of the proposed algorithm on different scenes. Besides, accuracy rate and comparison between features are demonstrated in section 4.2. In section 4.3, we have a brief discussion of efficiency of our proposed algorithm.

4.1 Experimental Results of Fire Detection

In the following, we use “yellow boundary” to represent global candidate fire region and the region become “red” region that means this region exist fire object.

The columns of left side contain original video sequence and the columns of right side contain detection results of the proposed algorithm.

Figure 4-1 illustrates the outdoor environment situation. There is no wind in Fig.

4-1(a).But the wind is blowing hard in Fig. 4-2(b).Our features wouldn’t be affected by the external environment and fire region can be detected correctly.

31

(a)

(b)

Fig. 4-1 Outdoor environment

Figure 4-2 illustrates the indoor environment situation. Fire region can be detected correctly.

(a)

32

(b)

(c)

(d)

Fig. 4-2 Indoor environment

33

Figure 4-3 For situations when fire is reflected on metal, it can still capture the correct fire region.

(a)

(b)

Fig. 4-3Fire is reflected on metal Figure 4-4 Fire in an extremely dark space.

Fig. 4-4 extremely dark space

34

Figure 4-5 Is a high intensity fire case, as shown in figure, can still find the correct fire region.

Fig. 4-5 High intensity fire

Figure 4-6 are non-fire test videos, most of which are traffic video or pedestrian videos. Because generally these videos do not have fire characteristics, thus no strong candidate blocks are generated, and will not produce global candidate fire region. Therefore, we show the foreground movement candidate block to represent that this video has a foreground moving object. 4-6 (d) is a special case. The entire screen is reddish hued, and color signal judgment becomes useless. However, we can still rely on other fire features to accurately judge if a fire exists or not.

(a)

35

(b)

(c)

(d)

Fig. 4-6 Non-fire situations

Figure 4-7 event is flashing red car lights. This event is in line with most of the fire features, including fire color, fire source, temporal difference, and variance analysis.

Fire surface area feature was finally used for filtering out.

36

Fig. 4-7 Filter out flashing red light

A large variety of conditions are tested including indoor, outdoor and sunlight variation each containing fire events, pedestrians, bicycles, motorcycle, tourist coaches, trailers, waving leaves, etc in Table 4-1. In our experiments, there are 13257 positive samples (fire events) and 62614 negative samples (ordinary moving objects).

Table 4-1 Properties of the testing videos Movie List Descriptions

Movie_01 Outdoor fire with pedestrians

Movie_02 Outdoor fire with pedestrians at farmland Movie_03 Fire in the living room

Movie_04 Fire in yellow tone living room Movie_05 Sofa on fire

Movie_06 The boiler burn up

Movie_07 High intensity fire under bridge Movie_08 Fire in the dark room

Movie_09 Fire in the wide storehouse Movie_10 Christmas tree burn up

Movie_11 Car with blinking light in tunnel

37

Movie_12 Cars in red tunnel at night

Movie_13 A trailer tow away a truck with pedestrians Movie_14 Cars with dark shadow

Movie_15 Cars in tunnel in day time Movie_16 Cars in tunnel at night

4.2 Accuracy Discussion

Data in Table 4-1 show the evaluations of each feature and the global testing result without ADU (Alarm Decision Unit). The reaction time is obtained by the ratio between frames to detect and frames per second. The detection rate and the false alarm rate are calculated as follow:

detected

From Table 4-1 we can see that the false positive rate of the detection rate is also high.

Fire source represent the higher flame intensity and fixed location feature. This has good affects in moving light or objects that does not have enough intensity. Temporal difference means that fire in a fixed block have continuous movement. This feature can eliminate fixed position, but non-moving lights, i.e. street light. Next, using global feature for verification, the variance can be filter out objects with fire like colors, but with smooth texture, i.e. red clothes, yellow car. After match the previous four feature values and the block becomes a strong candidate block, and has a high likelihood of being a fire body. Lastly, the fire surface area analysis will be used to make the last judgment.

38

We can see from the table that each feature will complement each other in filtering out false positives. Each algorithm forms a complementary relationship with each other, with variance as a strong feature, can lower the most false positive rates.

Table 4-2 Experimental results without ADU based on single frame Detection rate can further decrease from 2.8% to 0.8%.

Table 4-3 Experimental results with ADU based on single frame Detection

39

4.3 Comparison

The analysis of experiments implementing the proposed process derived in previous sections is presented in this section. Although most of the papers in the literature don’t provide experimental data, we implement some approaches and calculate the experimental results. The comparison results are presented in Table 4-3.

We were using the same test video in the proposed method in this thesis [1], [3], but the result of the comparison is based on the video number as a unit rather than frame unit. Because the performance of the image processing algorithms is more dependent on video scenes, and different test film are different length, specific algorithms will result in significant increase in false positive rates if the film is longer.

In order to avoid this situation, we changed to using the video number as a calculation unit for detection rate and false alarm rate.

Our test includes 10 fire videos, 6 non-fire videos, using three algorithms for testing respectively, and the test results are as in the following table:

Table 4-4 Comparison between the proposed method and work in [1][3]

Detected differential to measure the degree of fire disorder. This has the same effect as using the temporal difference, both measuring fire disorder feature. However, [1] not raise

40

more algorithms for verification, thus we can see from the table that method [1] has high detection rate, but also high false alarm rate. And this paper does not consider the high intensity fire, so the only fire that is undetectable is the high intensity fire.

Method [3] in this thesis refers to and improves the YCbCr fire color model.

This paper has very precise fire color judgment; however, this thesis only used color information, and did not use any other fire object feature for further confirmation.

Thus the reaction time is the shortest, but the false positive rate is the highest for the three methods, and similarly does not have high intensity fire detection affects.

The method proposed in this thesis has the highest detection rate and the lowest rate of false positives. This is using the fire’s different features to come up with different algorithms that filter out a variety of false positives. This is the result of mutual authentication of all the algorithms, and even lacking just one is indispensable.

However, the proposed algorithm uses more time accumulated data, thus the reaction time will be longer, which is a pity. But we hope that this set of algorithm calculation can be practically applied to daily life, so system stability is our emphasis.

41 block is a fire body. Decision Unit will enable the system to allow false alarm caused by instantaneous changes.

Most of other’s algorithms are only seeking higher detection rate. It does not provide enough information on how accurate the system might be. When considering accuracy of fire detection, people care more about how to decrease the false alarm rate and detect fire events quickly, rather than just increase the detection rate. Here the false alarm rate of the proposed system is significantly lower than others. This system can also locate the fire regions correctly even when both fire and non-fire objects exist in the same frame due to block processing, while other systems only detect whether there is fire existing in the whole video or the single frame.

The proposed algorithm can be operated well in variant environment. However, to further improve the performance of our system, some enhancements or trials can be made in the future. First, the reaction time slightly increased. This is because the proposed algorithm uses more information from the timeline, and the accumulation time of the data will cause the reaction time to become longer. Second, a chandelier equipped with spiral fan will cause the chandelier to have minor shaking, thus matching the fire feature algorithms and cause a false alarm. Therefore, if these

42

problems can be solved, our algorithms will be reliable.

This paper demonstrates a robust and efficient system for fire detection, and it involves the global and local features analysis for each candidate block. Experimental results show the opportunity of the real-time operation of surveillance systems and advanced applications. The false alarm rate of the proposed system is lower than that of the state of art. The proposed algorithm has low computation load and has been implemented on embedded system.

43

References

[1] T. H. Chen, P. H. Wu and Y. C. Chiou, “Application of panoramic annular lens for motion analysis tasks: surveillance and smoke detection,” 2004 International Conference on Image Processing (ICIP).

[2] W. B. Horng, J. W. Peng and C. Y. Chen, “A new image-based real-time flame detection method using color analysis,” Networking, Sensing and Control, 2005.

Proceedings. 2005 IEEE.

[3] J. Ebert and J. Shipley, “Computer vision based method for fire detection color videos,” Pattern Recognition Letters 1 January 2006, Pages 49–58.

[4] B. U. Toreyin, Y. Dedeoglu, U. Gudukbay and A. E. Cetin, “Computer vision based method for real-time fire and flame detection,” Pattern Recognition Letters 27 (2006) 49–58 .

[5] O. Gunay, K. Tasdemir, B. U. Toreyin and A. E. Cetin, “Fire detection in video using LMS based active learning,” Fire Technology Volume 46, Number 3(2010), 551-557.

[6] P. V. K. Borges and E. Izquierdo, “A probabilistic approach for vision-based fire detection in videos,” IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 20, NO. 5, MAY 2010.

[7] T. Celik and H. Demirel, “Fire detection in video sequences using a generic color model,” Fire Safety Journal 44 (2009) 147– 158.

[8] W. B. Hrong, J. W. Peng and C. Y. Chen, “A new image-based real-time flame detection method using color analysis,” Networking, Sensing and Control, 2005.

Proceedings. 2005 IEEE

[9] J. Zhao, Z. Zhang, S. Han, C. Qu, Z. Yuan and D. Zhang, “SVM based forest fire detection using static and dynamic features,” Computer Science and Information

44

Systems Volume 8, Issue 3 (June 2011)

[10] Z. Zhang, J. Zhao, D. Zhang, C. Qu, Y. Ke and B. Cai, “Contour based forest fire detection using FFT and wavelet,”2008 International Conference on Computer Science and Software Engineering

[11] Stauffer. C, Grimson. W.E.L, “Adaptive background mixture models for real-time tracking,” IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 2, June 1999.

[12] http://www.openvisor.org

[13] R. C. Gonzales and R. C. Woods, Digital image processing, Prentice Hall,2002