early detection system for vegetation fire in the ...

4 downloads 0 Views 52KB Size Report
Réka Kovács. Department of Control Engineering and Information Technology,. Budapest University of Technology and Economics,. Magyar tudósok körútja 2.
EARLY DETECTION SYSTEM FOR VEGETATION FIRE IN THE AGGTELEK NATIONAL PARK Réka Kovács Department of Control Engineering and Information Technology, Budapest University of Technology and Economics, Magyar tudósok körútja 2. 1117, Budapest, Hungary e-mail: [email protected]

Bálint Kiss Department of Control Engineering and Information Technology, Budapest University of Technology and Economics, Magyar tudósok körútja 2. 1117, Budapest, Hungary e-mail: [email protected]

Ákos Nagy Department of Control Engineering and Information Technology, Budapest University of Technology and Economics, Magyar tudósok körútja 2. 1117, Budapest, Hungary e-mail: [email protected]

Gábor Vámos Department of Control Engineering and Information Technology, Budapest University of Technology and Economics, Magyar tudósok körútja 2. 1117, Budapest, Hungary e-mail: [email protected]

Summary The Aggtelek National Park is one of the largest woodlands in Hungary having important forest reserves. A semi-automated early fire detection system is being installed in this area. The system uses a visual light camera sensor mounted on a commutation tower situated at the centre of the protected area. A rotating mechanism allows an almost full round view completing a rotation in 5 minutes. The acquired images are processed on a tower based industrial PC to detect smoke. The image processing algorithm has to be adapted to changing atmospheric conditions and to the relief of the area which impedes direct visual contact with the fire itself. Depending on the likelihood of the presence of smoke on the image, an alarm is transmitted to the regional fire fighting unit together with the image and the location of the possible fire source. Each alarm is subject to human visual verification and affirmation before any deployment. Hence the cost of a false alarm is kept considerably lower than the cost of a missed fire. The paper presents the architecture of the early fire detection system and the image processing algorithm. Test results are presented on images recorded from the communication tower.

1 Introduction Vegetation fires have a significant long-lasting effect on the environment. In order to reduce damages caused by vegetation fires, the early recognition of fires is essential. Remote sensing provides a quick and economic means of gaining fire information over large areas. The autonomous fire detection system under development at the Budapest University of Technology and Economics in collaboration with the Professional Fire Department of the Municipality of the City of Szendrı, and funded by the Ministry of Education, is a ground-based system with one scanning colour camera. Notice that existing systems [5] may use satellite [4], airborne, and multiple ground tower [2] based remote sensing techniques, too. The advantages of this system are its high spatial and relatively high temporal resolution and reduced cost due to its simple structure, still providing high reliability in fire detection. This paper is organized as follows. Section 2 is about the autonomous fire detection system. Subsection 2.1 describes the architecture of the system. The fire recognition part and a new smoke-detection algorithm are presented in 2.2. A decision network is responsible for the validation of the output of the image processing part. The network is introduced in 2.3. The whole system is currently in test phase, the pilot system is described in Section 3. Section 4 presents the test results. Section 5 concludes this paper with further outlook.

2 The fire detection and early warning system Launched in June 2004, the project aims to realize an environment monitoring and early fire detection system in the Aggtelek National Park with one single tower and one single alarm centre. The functioning costs of the sensing devices are to be kept at low level. The system is expected to be cost and human resource effective before the fire event occurs, but its post-alarm task is to increase the efficiency of the deployment of the eventually costly equipment, fire-fighter men and women. Notice that, since the alarm is directly sent to the Fire Department on permanent duty, the cost

of a false alarm is considerably less than the cost of a missed alarm. The final decision about the validity of an alarm and about the deployment of fire-fighters remains in the hand of the professional human operators. 2.1 The system architecture The sensor, a visual light camera, is situated in a transmission tower at a height of about 60 m. Rotated around by a mechanical system, the camera takes pictures of the area at given angles. A cycle is completed (the camera gets back to the initial point) every 5 minutes. The required detection distance is about 18 km in good weather conditions (in daylight). The smoke detecting algorithm works on the images from the camera. It is followed by a decision network that validates the output of the algorithm and specifies the location of the fire. The detection results, the fire location and features, together with visual information are transmitted to the fire department. 2.2 Image processing based fire detection The colour images coming from the camera require a larger space to store and longer time to process compared to greyscale images and do not contain significant colour information regarding smoke plumes. In order to reduce space and time requirements, the images are converted into greyscale, but this way the colour information is lost. Because of the simple architecture of the system –it contains only one sensor- the recognition should be done on a series of images. Since during the time interval between two consecutive images, taken from same region, the shape of the smoke plume may change considerably, motion detection and characterization methods cannot be used. The smoke is recognized by the changes between two consecutive images. These changes can easily be detected on the difference of the two images. The detection algorithm uses the methodology proposed by Breejen et al. in 1998 [1],[3]. Once the images are converted into greyscale and the image differences are produced, the first step of the detection is the partition of the image into smaller rectangular parts (the so-called “detection bins”). The partitioning is motivated by the fact that image characteristics might be

very different from each other in distant zones of the image. Let B k denote the k-th detection bin within an image. The image is divided into M bins with the same size, so that

K, M . The values of the pixels of a B = {p , K p }, k = 1,K , M . B i = B j , ∀i, j = 1, k

1

grey-level bin

are

n

For each bin in the image, the average and the standard deviation of the luminance-values of the pixels within the bin is calculated. The average for the k-th bin is:

Ik =

1 n

∑ n

j=1

p j , p j ∈ Bk

(1)

The deviation for the kth bin is:

σk =

1 n

∑ (p − I ) , p ∈ B n

j =1

2

j

k

j

k

(2)

The bin intensities have normal variations caused by changes in illumination (e.g. clouds). In order to eliminate this effect, a moving average and a moving standard deviation is used on the image sequence. The moving average is calculated by the following expression:

MA k = α ⋅ MA prev ,k + (1 − α ) ⋅ I k ,

(3)

where MA k is the current moving average of the k-th bin, α is a scalar factor, MA prev , k is the previous moving average and I k is the average of the bin computed in (1). The formula for calculating the moving standard deviation is similar to (3) and uses the same scalar factor:

MD k = α ⋅ MD prev ,k + (1 − α ) ⋅ σ k ,

By changing the parameter α the sensitivity of the algorithm can be set to track “fast” changes. (A “fast” change is caused by a smoking fire while a “slow” change is caused by fog or clouds.) Two criteria are used for detection: a temporal and a spatial criterion. The bases of the temporal criterion are two thresholds: an upper threshold and a lower threshold determined by the moving average and by the moving standard deviation. There is smoke in the k-th bin if the following expression holds true:

I k − MA k > c1 ⋅ MD k ,

(5)

where c1 is a positive scalar. Assuming that the temporal changes of grey levels follow normal distribution, the possibility of a grey-level belonging to the distribution but lying outside the triple of the deviation is less than 1%, so the value of c1 is set to 3. Because the algorithm works on difference images and averages are calculated, it may happen (especially in case of a heavily smoking fire) that the averages of the bins are not below or under the thresholds. To overcome this difficulty, a spatial criterion is determined for the image based on the standard deviation of the detection bins. The basis of the criterion is the observation that in the cases described above the standard deviation of the bin significantly increases compared to the rest of the image. In order to examine the bin-deviations within the image, we need to determine the average and the standard deviation of the grey-level deviations. The average of the deviations is:

σ=

1 k

(4)

And the deviation is:

∑σ k

j=1

j

(6)

σD =

1 k

∑ (σ − σ ) k

j=1

2

j

(7)

In the expressions, σ j stands for the bindeviations calculated in (2). There is a detection in the k-th bin if:

σk − σ > c2 ⋅ σD

(8)

The value of the detection constant c 2 was set to 3 based on the initial tests. If either (5) or (8) are true for a bin, that bin becomes a candidate for detection. If the bin is a candidate for detection on three out of five consecutive images, then that bin is selected as a “valid detection bin”. In order to reduce the number of false detections caused by non-fire events (e.g. changes in illumination, moving objects) the position of bins with valid detections is passed to the decision network. 2.3 The decision network The presence of fire can be predicted more accurately if several features are simultaneously investigated including information gathered from outside the image. Geographic information is provided using the camera angle and surface models; the image recording times and a library of previous valid and invalid detections are also considered. This extra information may enforce or weaken the likelihood of fire. Relationships between features and the presence of fire are modelled using belief networks. The background we use to construct knowledge based expert systems for the prediction of fire among given circumstances with uncertainty is provided by the theory of Bayesian networks [6]. These networks are used to model a given problem domain by means of a joint probability distribution function (PDF) over a set of probabilistic variables as given by:

P : V1 × V2 ×

K× V

n

Since these networks have a wellinterpretable graphical (qualitative) component and a numerical (quantitative) component, they fit the cases where the existing knowledge in the domain comes partly from human experts and partly from databases. In fact, human expert knowledge can be coded using mainly the graphical component, whereas data extracted from databases are used to train the quantitative component. Definition [Bayesian networks]: A Bayesian network over a set of variables U = {V1 , V2 , , Vn } consists of a graphical and a quantifying component:

K

1. Graphical component: directed acyclic graph: G . Each node in the graph represents a variable in U. The set of parents of a variable V (i.e. the nodes from which there is an arc pointing to V) is denoted by π V . 2. Quantifying component: each variable V in U (i.e. each node in G) is quantified with a conditional probability distribution function denoted by P( V π V ) .

The Bayesian network encodes a joint PDF over U:

P(U ) = ∏i =1 P(Vi π Vi ) . n

Let us suppose now that for some variables of the network we are able to obtain either their deterministic value or their distribution and we wish to calculate the distribution of a given query variable Vq knowing the evidences, i.e.

P(Vq E ) , where E ⊂ U is the set of evidence variables. This is corresponds to the

P(Vq E ) =

(9)

called

∑∏

U \{Vq }

→ [0,1]

(10)

n

i =1

inference

P(Vi π Vi )

and

(11)

marginalization where all V ∈ E is replaced by the values corresponding to the evidence. The resulting probability distribution combines the expert knowledge encoded in the network and the collected evidence. 3 The pilot system The fire detection system is currently in test phase. In the pilot system, the image capture and the smoke recognition parts are realized. Test images were taken from a height of about 40m. A registry file contains the names of a sequence of images. This file is passed to the detection software. The output of this software is a vector made from the image features. There is also a visual output where the detection results are shown: bins with valid detection are denoted by white rectangles in the output image.

Figure 2. The pre-processed image (image difference)

4 Results Let us present some examples showing the functioning of the detection algorithm. After preprocessing, images with a resolution of 1500 x 1200 pixels are used. Several detection bin sizes have been tested; best results were obtained using 70 x 50 pixel size. Different stages of the recognition are shown below.

Figure 1. The greyscale image

Figure 3. The output of the algorithm On Figure 1 a section of the converted greyscale image shows a heavily smoking fire. This image goes under a pre-processing procedure in which the image difference is made and grey-levels are shifted to the middle of the grey-level interval (in order to reveal differences with negative signs as well). A pre-processed image is represented on Fig. 2. The algorithm works on these pre-processed images; it produces a visual output alongside the output vector to the decision network. Fig. 3 shows the visual output of the algorithm. In the output image, valid detection bins are replaced by white rectangles. Note that when choosing the right detection bin size alongside the size of the image the expected size of the fire should be taken into consideration as well. This depends on different geographical and architectural features (the terrain, the height of the camera, the distance of the fire from the tower, etc.) therefore optimal parameter settings should be determined at the place of installation only.

5 Conclusion and future work During the tests, the algorithm proved to be reliable in the detection of fires and produced a small number of false detections. In further work, the decision network should be implemented in the system; the complete system should be tested thoroughly and as a last step it should be set up in the National Park. Acknowledgement This research was partially funded by the Hungarian Ministry of Education under grant KMFP 00025/2004. The authors are also deeply indebted to the Antenna Hungária Inc. for granting access to the transmission tower. References [1] AFFIRM: Autonomous forest fire detection and prevention aid system integrating remote multisensor terminals, Science Research and Development Program of the European Union May 96 − May 98 http://gps-tsc.upc.es/imatge/_Philippe/AFFIRM.html [2] Autonomous Early Warning System for Forest Fires Tested in Brandenburg (Germany) (IFFN No. 22 April 2000, p. 84-90) http://www.fire.uni-freiburg.de/iffn/country/de/ /de_6.htm [3] E. Breejen, M. Breuers, F. Cremer, R. Kemp, M. Roos, K. Schutte, J. Vries: Autonomous Forest Fire Detection; 3rd International Conference on Forest Fire Research, Vol.II. (1998) 2003-2012, [4] F.J. Ahern, J.G. Goldammer, C.O. Justice: Global and Regional Vegetation Fire Monitoring from Space: Planning a Coordinated International Effort, Hague, SPB Academic Pub., 2001, ch. 11. [5] Multimedia Applications Section of Canada Centre for Remote Sensing: Fundamentals of Remote Sensing, http://www.ccrs.nrcan.gc.ca [6] S.J. Russel, P. Norvig, : Artificial Intelligence. A modern approach, Prentice Hall, 1995