Video Based Smoke Detection Using Smoke Growth Analysis

Fires are one of the main causes of death and destruction of properties which cost billions of lose every year. Smoke is the early stage of starting fire. If smoke can be detected accurately in its initial stage, massive fire incident can be prevented. This research proposes a video-based smoke detection approach based on smoke-growth analysis. In this proposed model, the preprocessing stage; Image frames are extracted from the videos. From the extracted frames, background subtraction method has been applied to subtract the background to achieve the objects using Gaussian’s mixture model (GMM). Now, the subtracted objects are gone through a process to extract its feature. To extract the features, we proposed a method that determines the growth region of the object using feature vectors. Finally, these features are given to a support vector machine (SVM) for discriminating and clustering our data more expendably which leads the way to provide accurate detection of smoke. In addition, histogram of the growth analysis has been provided to observe the movements of the objects. The proposed approach can work with greater effectiveness with low rate of false alarm.

Introduction

Detecting fire in its early stage can prevent mass destructions and can save thousands lives and valuable assets. Smoke components are the represent the beginning or the early stage of a fire incident. That is why detecting smoke at the early stage can provide an effective approach the alarm the incident of a fire event. Fire incidents occur economical degradation and environmental 2 impairment, so detection of early stage of fire can avert fire and minimize the damage and save lives and properties. Video based fire detection system has been developed as the newly technique for the past few years. It’s effectiveness in the detection of fire made it popular quickly. More Accuracy and low rate of false alarm made it more robust and dynamic. Using of surveillance in the smoke detection is the strength to serve the requirements and essentiality in large rooms and tall buildings or the woods.

There are many methods for detection of smoke in the field of computer image processing sector. And most of them even use combined approaches of several methods to improve performance and reliability. Some of the methods are similar in most smoke detection approach; they are motion detection, region of particles analysis, dynamic volume analysis and smoke classification stages. Growth rate is calculated from increasing number of pixel of the moving particles area. The differences are observed by algorithms used in those different stages. All of the methods and techniques use mainly three stages, preprocessing stage, feature extraction and Classification. Researchers from all over the world have been working on this new technique and have done a lot of works.

R. Islam Used Optical Flow Characteristics for fire Alarm Systems. Y. Cappellini built an Intelligent System For Automatic Fire Detection in Forests. G. Healey they made a System for Real-Time Fire Detection. H. Yamagishi worked on an algorithm Using a Color Camera for fire flame detection. But fire detection which are based on flame detection were vey inaccurate. The traditional fire detection system uses electric sensors which work in the manner of heat transaction through the sensors. Any kind of interference in the sensor could delay the time of alarming or raise the false alarm. So new methods are developed to detect fire and smoke. Motion and color are very important feature of fire and smoke.

Y. Chunyu has worked on such a video-based fire and smoke detection system using motion and color features. I. Kolesov worked with optimal mass transport based on optical flow and neural networks for Fire and smoke detection in video. Wavelet in smoke, image processing for automatic smoke detection and adoptive background modeling for real time tracking have been developed for building a very effective and accurate fire alarm system. In this paper, we proposed an algorithm that utilizes image processing and works growth of smokes its region. The smoke has the growth in the region which is different from non-smoke particles growth against a certain time. The features are classified by a Support Vector Machine or shortly called SVM.

The proposed smoke-growth smoke detection model

This paper presents a technique for motion detection that organize various innovative mechanisms. In this paper, preprocessing stage combines background subtraction and image segmentation. We used support vector machine (SVM) algorithm to classify moving objects which are supposed to be smoke. The proposed approach characteristics features of smoke focused is the growth area of the smokes and growth rate against time. A flow chart of the proposed smoke detecting method is shown in Fig. 2.1 and they are further discussed with more details below.

Frame extraction

Several videos are taken from the surveillance. Some videos are filmed with camera and phones igniting smokes intentionally for the experiments as dataset. Then we extracted 4 the frames from the videos with an existing algorithm. In our case, a video of 10 seconds span has been divided into several frames at 1 second of interval, among which 5 frames at 2, 4, 6, 8 & 10 seconds have been selected for further analysis.

Background subtraction

Extracted frames are needed to be subtracted for the foreground objects. There are several existing methods for background subtraction. Gaussian mixture model (GMM) is used in this experiment for the subtraction of background of the frames. The GMM subtracts the background from the foreground. It uses a threshold value determining the grey level. And normalizes the image into grey image depending on the threshold value as depth of color.

Frame segmentation block

Frame Segmentation is an important part of frame analysis. In our case, we have segmented each frame into 16 density block which are later used for further analysis of smoke growth. Notation of each segmentation of the frame which are called density blocks are shown in the table below with their respective percentage growth values

Feature extraction

In this section, on the basis of Binary pattern, Local binary patterns are introduced to classified and characterize the smoke. All of these determined features are extracted based on a block method. In study of machine learning, pattern recognition and image processing, extraction of features starts from an initial set of measured data and builds derived values which are actually the features, intended to be informatory and necessary, simplifying the future learning and generalization steps, and in some cases, leads to better human interpretations. Feature extraction is related to dimensional reduction differences. When the input data to an algorithm is too large to be processed and it is suspected to be unnecessary (e.g. the same measurement in both feet and meters, or the repetitive presentation of image as pixels), then it can be simplified as a reduced set of features (also called feature vector). Determining a necessary subset of the initial features is called feature selection. The selected features contain the relevant information from the input data, so that the expected action can be performed by using this reduced simplified representation instead of the whole initial data.

The following frames of smoke, their feature tables and smoke-growth diagrams are shown according to the corresponding letter a, b, c. The frames extracted from video are used to calculate the growth values and these values are subtracted from the following frame to get a percentage growth value in each density blocks.

Feature vector

Feature vector is the average feature of five frames’ features. For achieving more accuracy and error handling we not only extracted the smoke frame, but also the objects which are similar to smokes like car lights and light bulbs on the wall. Car lights and light bulbs give similar view of smoke and thus create confusion. In this manner, classification becomes more difficult. Hence the accuracy may reduce which results to high rate of false alarm. In this paper, we have attempted to reduce this kind of disturbance with experimenting non-smoke objects also. Feature vectors are the density of smoke particles containing in a segment. The smoke particles fill the block gradually. With certain amount of smoke particles almost all the blocks of the segmented are covered.

SVM based classification

In machine learning, support vector machines are called the supervised models associated with particular learning algorithms that analyze data used for clustering and classification and analysis of regressions. Given a set of training data set, each of them classified as one or more certain categories. An SVM model is a representation of the dataset as points in space and mapped so that the dataset of the different categories. New examples are then given into that with same space and predicted them to a category, based on which side of the gap they fall. Not only it can perform linear classification approach but also SVM can perform a non-linear classification task using the kernel trick method. supervised learning is not possible if the data are no labeled. In that case, an unsupervised learning approach is required, which intends to find natural clustering of the data to various categories. Then it maps new data to these formed categories. The clustering algorithm provides an improvement to the support vector machines which is called support vector clustering. which is often used in industrial applications either when data are not labeled or some data are labeled as a preprocessing for a classification. We have examined those video datasets and each training and testing was performed with cross validation the better and improved results.

Experimental results

Experimental environment

For this experiment, we needed to come up with some setup. We used Matlab 2017a version to build a prototype. We set fire and filmed it for collecting the experimental dataset. We also used some videos from the internet. The resolution of every frame was 40*32 for achieving the most simplified frames of smoke and non-smoke. Figure 6,7 and 8 shows the screenshot of the smoke, car and light bulb respectively. Phone cameras are used to filmed videos. Figure 6 shows about frames where we can see the smoke growing against the time. For evaluating the proposed model, we construct a dataset which contain 3 signs. For each signs, there are 20 videos. Total number of videos is therefore 60. For each video, there are 5 number of frames. Each frame type was Binarized image. Where there are smoking class and non-smoking class. The dataset is branched into three sets. One for feature selection, another for training and last one for testing data.

Analysis of performance

The dataset of this experiment is only the feature vector which are extracted from the videos and which reflects the characteristic of smoke. As saying, the characteristic of car lights floats in the horizontal of its direction. Where the light bulb objects never go anywhere. Which we can mathematically explain with their corresponding percentage density block. From the percentage of density values tables of smoking, non-smoking frame, the density values of different times at the density block are plotted side by side as histograms which are given below: As we can see the histograms of the smoke frames, it is gradually taking almost all the block of segmented particles. It means, the area where the smoke particles are being detected increasingly with time. Smoke particles generally floats in the vertical direction. With time, the volume area of smoke objects expands. We observe these phenomena with the histograms of smoke and non-smoke objects. From figure 10, the variation of percentage growth of smoking frame has been shown. Here we can vividly see that the variations in all density blocks are not uniform and increases at first then decreases with time. This nonlinear and rapid changes indicate towards smoke growth as we expected.

From figure 11, the variation of percentage growth has been shown. Here it can be easily see that the variations are not uniform and increases at first then decreases with time but they are limited in certain region of frames. From this observation, it can be concluded that it a non-smoking growth, more specifically a car light. From figure 12, the variation of percentage growth has been shown. Here it can be easily see that the variations are uniform and there is no increase or decreases with time and they are limited in only in a specific region of frames. From this observation, it can be concluded that it a non-smoking growth, more specifically a bulb light.

Conclusion

Nowadays fire incidents are taking lives more than other accidents. Wild fire causes death of thousands of lives and wild properties. It also causes imbalance to the natural and wild life cycle. In this paper we proposed a more effective and robust approach for the detection of smoke based on smoke-growth analysis. The presence of smoke in a target area is determined by the growth of smoke in the region. The flow of the working process starts with the preprocessing, frame extractions from the videos, background subtraction feature extraction of the frames, growth analysis with individual histograms. And then a SVM based decision making is applied to identify the smokes of the video frames. The proposed model is experimented with standard datasets and at the end it showed significant improvement with smoke growth characteristics.