CN105976365A - Nocturnal fire disaster video detection method - Google Patents

Abstract

The invention relates to a night fire video detection method, comprising: converting the format of the night fire video shot in infrared mode from RGB color space to HSV color space; performing morphological filtering processing; using a neural network model to train night flame pixels The brightness feature value of the point, and judge whether the pixel in the picture belongs to the suspected flame area according to the training results; perform binarization on the area that meets the brightness feature threshold to obtain the highlighted area, and then use the canny algorithm to find its edge and calculate The perimeter and area of the edge contour; calculate the circularity of the highlighted area; calculate the nighttime flame flashing features; calculate the number of full-screen flashes per second; establish an SVM classifier, and use each feature value to classify the highlighted area suspected of fire . The invention can be used to search the video in the initial stage of the fire occurrence, and locate the flame occurrence object intuitively and accurately.

Description

A nighttime fire video detection method

technical field

The invention belongs to the field of digital image and video processing, and specifically relates to a video detection method for fire at night.

Background technique

All over the world, fire is one of the major disasters faced by human beings, causing a large number of human casualties and property losses. With the development of science and technology, fire early warning technology and fire protection technology have made great progress in recent years, but serious fire accidents still occur from time to time. Therefore, the precise location of the fire point and the accurate analysis of the cause of the fire after the fire is very important. On the one hand, accurate fire location can be used as evidence to clarify the responsibility of fire accidents; on the other hand, accurate fire cause analysis is conducive to early warning and prevention of the same type of fire in the future.

At present, in the investigation of fire causes, the main method is on-site inspection, and the auxiliary means are fire investigation visits and fire technical appraisal. However, this traditional fire cause investigation method has drawbacks and limitations. First of all, the traditional methods are likely to lead to the destruction of the fire scene and the loss of relevant evidence during the investigation process, thus wrongly identifying the cause of the fire. Secondly, the traditional method sometimes fails to form a chain of evidence for determining the cause of the fire, and has no definite probative and persuasive force, resulting in problems such as unclear determination of the cause of the fire, insufficient evidence for the determination of the fire, and inability to corroborate each other between evidences.

In recent years, more and more public places have installed complete video surveillance systems. This makes the fire detection algorithm based on the video surveillance platform emerge as the times require. Compared with the traditional smoke-, temperature- and light-sensitive point fire detectors, the video-type fire detection system has a large detection range, can be used in tall spaces and outdoor environments, is low in cost, and can provide information on the development trend and spread speed of fires. information and so on. However, all current video-based fire detection systems are used for real-time monitoring of fires and are used to send out fire alarm signals when a fire occurs, but there is no video-based fire detection algorithm for post-event cause analysis of fires and location of fire points.

In many fire accidents, due to the unattended situation at night, most of the out-of-control fire accidents often occur at night. Compared with daytime fires, the analysis of the cause of fires at night is particularly difficult. There are often fewer witnesses to fires at night, and it is difficult for fire investigation and visit methods to work. Moreover, after the fire was extinguished, most of the objects on the scene were burned up, and it was sometimes difficult to find strong evidence for on-site inspections. At this time, the video left by the on-site monitoring system has become the main means of analyzing the cause of the fire. However, it is impossible to quickly complete the search task in a large number of videos by using human eyes to find the fire point and the cause of the fire. Moreover, the video at night is very dark, and it is often impossible to see anything other than the flame in the video. Even if the human eye finds the fire process, it is still impossible to determine the burning object and the cause of the fire. It can only be compared with the video taken by the same camera during the day. right. If this kind of work is done manually, it is time-consuming and laborious, and the accuracy is low.

Contents of the invention

The invention provides a fire video detection method at night, which can be used to analyze and process the night video captured by a camera in an infrared mode, thereby judging whether a fire occurs at night, and helping fire investigators to analyze the cause of the fire and the burning material. Technical scheme of the present invention is as follows:

A fire video detection method at night, comprising the following steps:

1) Detection of nighttime fire videos taken in infrared mode;

2) Carry out format conversion to night fire video, convert from RGB color space to HSV color space;

3) Perform morphological filtering to reduce the impact of noise on the detection effect;

4) Use the neural network model to train the luminance feature value of the flame pixels at night, and judge whether the pixels in the picture belong to the suspected flame area according to the training results, and exclude the pixels whose luminance value is lower than the threshold;

5) Binarize the area that meets the brightness feature threshold, and then perform erosion and expansion operations on the binary image to eliminate the influence of noise, obtain the highlighted area, then use the canny algorithm to find its edge, and calculate the perimeter of the edge contour length and area;

6) According to the characteristic that the shape of the flame at night is similar to a circle, the circularity of the highlighted area is calculated to exclude the highlighted area with a lower circularity;

7) According to the characteristics that the area of the flame burning process changes drastically at night and the frequency fluctuates around 10 Hz, the nighttime flame flickering characteristics are calculated, and it is considered that the difference in the area of the highlighted area of two consecutive frames is more than 50%, which meets the flickering characteristic. The number of times the bright area satisfies the flickering feature;

8) According to the characteristics of full-screen flickering caused by indoor flames at night, the fire detection method is as follows: use the frame difference method to detect the number of pixels whose brightness values change greater than a certain threshold in two adjacent frames. 60% of the total number is considered to be a full-screen flash; calculate the number of full-screen flashes per second;

9) Use the parameter values calculated in the above steps 4) to 8) as feature values, use the method of machine learning to establish an SVM classifier, and use each feature value to classify the highlighted areas of suspected fires, thereby judging suspected fires Whether the highlighted area of the camera matches the characteristics of the night flames captured by the camera in infrared mode, draw a conclusion, and record the time and position when the highlighted area first appeared.

The invention can be used to search the video at the initial stage of the fire, intuitively and accurately locate the object where the flame occurred, and can assist in finding out the cause of the fire.

Description of drawings

Figure 1 is a screenshot of a fire video at night.

Figure 2 is a high-brightness area extracted by the canny algorithm.

Figure 3 is the nighttime fire image detected by the program.

detailed description

Taking a specific example as an example, briefly describe the process of realizing auxiliary analysis of the cause of fire at night

Use the interface program to select the night fire video captured by the camera in infrared mode for detection, and then select the storage path of the detection results. Convert the format of the input video code stream, and convert the video from RGB color space to HSV color space, which is convenient for subsequent processing. Scale the picture according to the predetermined scaling ratio to reduce the calculation amount of the subsequent algorithm. Morphological filtering is performed on the picture to reduce the impact of noise on the detection algorithm. Screenshots of nighttime fire videos after filtering are shown in Figure 1.

Use the neural network model to train the brightness feature value of the flame pixels at night, and judge whether the pixels in the picture belong to the suspected flame area according to the training results, and exclude the pixels whose brightness value is lower than the threshold. Binarize the regions that meet the brightness feature threshold. Then perform an erosion and a dilation operation on the binary image to eliminate the influence of noise. Among them, a 3*3 rectangle is selected as the structural element of corrosion and expansion. For the area that reaches the flame brightness threshold in infrared mode, use the canny algorithm to find its edge, and calculate the perimeter and area of the edge contour. The high brightness area is shown in Figure 2.

According to the feature that the shape of the flame at night is approximately circular, according to the formula Computes the circularity of highlighted areas to exclude interference from non-fire objects with complex shapes. where e represents circularity, s represents the area of the contour, and l represents the perimeter of the contour. According to the characteristics that the area of the flame burning at night will change drastically and the frequency fluctuates around 10 Hz, the characteristics of the flame flickering at night are calculated. It is considered that the area of the highlight area in two consecutive frames differs by more than 50%, which meets the flicker feature once, and the number of times the highlight area satisfies the flicker feature within 25 frames per second is calculated. Fire detection is based on the fact that indoor flames at night cause full-screen flickering. The frame difference method is used to detect the number of pixels whose luminance value changes greater than a certain threshold in two adjacent frames. If the number of changed pixels is greater than 60% of the total number of pixels, it is considered that a full-screen flicker has occurred. Count the number of full-screen flashes that occur within 25 frames per second. Using the method of machine learning, a SVM classifier is used to classify the suspected highlighted areas according to all the eigenvalues calculated above, so as to judge whether the suspected areas meet the characteristics of night flames captured by the camera in infrared mode, and draw a conclusion. Note when and where the highlighted area first appeared. The picture of the fire at night identified by the program is shown in Figure 3, where the box represents the identified flame area.

Claims (1)

1. fire at a night video detecting method, including following step:

1) under infrared mode, the fire video at night of shooting detects；

2) fire video at night is carried out form conversion, be transformed into HSV color space from rgb color space；

3) morphologic filtering process is carried out, to reduce the noise impact on Detection results；

4) utilize the brightness value of neural network model training flame pixels point at night, and whether judge the pixel in picture according to training result Belong to doubtful flame region, get rid of the brightness value pixel less than threshold value；

5) region meeting brightness threshold value is carried out binary conversion treatment, then bianry image is corroded and expansive working with eliminate noise shadow Ringing, obtain highlight regions, recycling canny algorithm is sought its edge, and is calculated girth and the area of edge contour；

6) according to the feature that flame profile at night is approximately round, the circularity of highlight regions is calculated, in order to the highlight regions that circularity is relatively low to be given To get rid of；

7) according to the area during night flame combustion can the feature that fluctuate near 10Hz of acute variation and frequency, calculate flame sudden strain of a muscle at night Bright feature, it is believed that two continuous frames highlight regions area difference more than 50% then meets a blinking characteristics, calculates highlight regions per second and meets blinking characteristics Number of times；

8) can cause the feature detection fire of full frame flicker according to interior flame at night, method is as follows: utilize adjacent two frames of frame difference method detection to occur bright Angle value change is more than the pixel number of certain threshold value, if the pixel number of change is more than the 60% of pixel sum, think there occurs the most full frame Flicker；Calculate the number of times of the full frame flicker of generation per second；

9) using above-mentioned steps 4) to 8) calculated parameters value as eigenvalue, the method utilizing machine learning, set up svm classifier Device, utilizes each eigenvalue to classify the highlight regions of doubtful fire, thus judges whether the highlight regions of doubtful fire meets photographic head red The flame feature at night of shooting under external schema, it was therefore concluded that, record time and position that this highlight regions starts most to occur.