Fire detection method and fire detection device

Abstract

The invention relates to methods for fire detection by means of an evaluation of a scene (12) captured by a camera (10) via a machine learning system, characterized in that in the event that a fire is suspected based on a first evaluation of the scene (12), a fire warning is only issued if an expected value for the smoke density and / or a change in smoke density exceeds a predetermined expected value for the smoke density.

Description

State of the art

[0001] A smoke detection device and a method for detecting a fire are already known from DE 10 2016 207 705 A1. The movement of a moving object is determined from at least two individual images of an image sequence, with a distinction between smoke and non-smoke being made based on the movement of the moving object. Disclosure of the inventionAdvantages of the invention

[0002] The fire detection method according to the invention has the advantage that, following an initial computer-assisted analysis to determine whether a fire is present, a warning is only issued if a determined expected value for a smoke density exceeds a predetermined expected value for the smoke density. This can ensure that false alarms are avoided in cases where the smoke density is low, for example as a result of steam or dust development. The expected value for a smoke density is understood to be the p-value, which is assigned to the respective classes for different smoke densities in the machine learning system, which is designed, for example, as a neural network. During classification, the class assigned the highest p-value is selected.The machine learning system thus assigns the captured image data to different groups of smoke densities.

[0003] If the captured image of a scene is assigned to a smoke density class that exceeds a predefined smoke density limit, an alarm is triggered. In other cases where the assignment does not assign a smoke density class that exceeds the predefined smoke density limit, no alarm is triggered. The individual classes are assigned to corresponding smoke density ranges from a lower to an upper value, i.e. a minimum and a maximum smoke density are specified for each class. By suppressing alarms when the smoke density is insufficient, the robustness and detection speed for fire detection can be improved. By assigning signals to different classes using the machine learning system, a more reliable result can be achieved than with an absolute measurement.

[0004] Further advantages arise from the dependent claims. It is advantageous for a scene to be assigned to at least one class of scenes without detected smoke development or to at least one of two classes with different smoke density ranges. This clearly makes it possible to classify not only the presence of smoke, but also a quantitative measure of smoke density.

[0005] Classification is preferably performed by evaluating all or a predefined subset of the pixels of a captured image of a scene. If necessary, smoke density can also be determined at different points in time. By limiting the analysis to a specific pixel area, or even to a single pixel, it is possible to easily determine smoke density.

[0006] In one embodiment, a selected image area is determined by means of the upstream fire detection using a machine learning system. In particular, image areas that exhibit high dynamic range or high optical flow can be selected. A smoke density determination can then be limited to such an image area, so that the probability of detection and thus a correct assignment can be improved by means of the machine learning system. In particular, the probability of correctly assigning an image to one of the predefined classes of different smoke densities increases.

[0007] Furthermore, it is advantageous to derive a change in smoke density from a change in the color of the image pixels in the selected image area. Considering the hypothesis that the color of the smoke and the color of the background are constant in such an image area, a change in smoke density leads to a change in the transparency of the smoke relative to the background and thus to a color change in the area between the background and the smoke color. Such a color change can be easily evaluated and can then be used for analysis using the machine learning system.

[0008] Furthermore, it is advantageous to store smoke density information assigned by the machine learning system together with an image for advanced analysis and / or for training the machine learning system.

[0009] Corresponding advantages result for a fire detection device according to the invention. drawing

[0010] Embodiments of the invention are illustrated in the drawing and explained in more detail in the following description.

[0011] They show: Fig. 1 an embodiment of a fire detection device with a camera and an alarm unit when observing a scene, Fig. 2a, Fig. 2b and Fig. 2c Example for different smoke densities, Fig. 3 an embodiment of a method sequence according to the invention. Embodiments of the invention

[0012] In the Fig. 1 shows a camera 10, in whose field of view 11 there is a scene 12 that is being observed by the camera 10. Images from the camera 10 are subsequently forwarded to a fire detection device 20, in which a computing device 21 evaluates the image data captured by the camera 10 by accessing a stored machine learning system 22. A plurality of images with various fire situations and non-fire situations are learned in the machine learning system 22, wherein the fire detection device 20, in particular by evaluating a sequence of images of the scene 12 captured by the camera 10, assigns the captured image data to either a fire situation or a non-fire situation.If a recorded scene is associated with a fire scene, the fire scene is forwarded to an evaluation device 25, where a computing device 26, accessing a provided machine learning system, evaluates the recorded image information with respect to smoke density. For this purpose, individual images or image sequences can be forwarded.

[0013] In the Fig. 2a, Fig. 2b and Fig. 2c shows examples of image representations 31, 32, 33 in which an object 34 is shown in a scene. In the scene according to the Fig. 2a no smoke is visible, in the Fig. 2b light smoke 35 can be seen and in the Fig. 2c shows a sealed space 36. In one embodiment, the fire detection device 20 specifies a frame area for a possible fire area to the evaluation device 25, in which Fig. 2b for example the frame 37 around the light smoke and in the Fig. 2c the frame 38 around the dense smoke. According to the Fig. 2a, either no frame at all or a frame above the object 34 is specified as frame 39. In the case of frame 39, this can be a specified area above the object in which warning signals may be recorded. In the case of Fig. 2b and Fig. 2c, these are areas with dynamic image movement that have a high probability of fire or smoke developing in these areas.

[0014] After an evaluation by the evaluation device 25, this now provides for the image recording according to the Fig. 2a, that this is assigned to the class “no smoke”. The recording according to Fig. 2b is assigned to the class “low smoke density” and the intake according to Fig. 2c of class 2c is assigned to the class "high smoke density." The respective classes are preferably defined with specific limit values ​​for the respective smoke densities in a memory 27. These assignment values ​​are stored in memory fields 40, 41, 42, assigned to the respective image data.

[0015] If the smoke density is assigned to the low smoke density or "no smoke" class, no warning is subsequently issued. However, if the smoke density is assigned to the "high smoke density" class, an alarm device 50 is activated via a warning line 28, and a fire warning is issued to the user via an output device 51, for example, a visual and / or acoustic warning device.

[0016] In one embodiment, image information captured by camera 10 of the scene is displayed to the user. Via an input device 52, the user can then provide feedback as to whether the fire situation was correctly or incorrectly detected. This feedback, together with the information stored in memory fields 40, 41, 42, can be used for further training of the machine learning system.

[0017] A mathematical determination of the smoke density can preferably be carried out by estimating the smoke density from the image data. One possibility for this is to describe the color of a pixel in the image as x=(1−α)*s / −α*b, where α is the transparency of the smoke, s the color of the smoke, and b the color of the background. Color is the mathematical representation of a color based on a color mapping table, the color captured by the camera to a screen color, where the color is assigned a value between 0 and 255, for example. With a larger color palette, other color values ​​are possible. A change in transparency between two points in time i and j can be written as xi−xj=(αi−αj)*(s−b), where it is assumed that the color of the background and the color of the smoke in the image do not change significantly between time points i and j. Thus, an approximate change in smoke density can be written as (αi−αj)=(s−b) / (xi−xj).

[0018] If one now sets a background color of a reference image and uses a smoke density determination based on the machine learning system described above, especially with regard to a large number of different smoke density classes, a change in smoke density can also be determined.

[0019] In the Fig.3 shows an exemplary embodiment of a method sequence according to the invention. Starting from an initialization step 60, fire detection 61 initially takes place. In a subsequent first test step 62, it is checked whether a fire has been detected. If this is not the case, fire detection 61 is continued. If this is the case, a smoke density assignment takes place in an assignment step 63. In a subsequent test step 64, it is checked whether a minimum smoke density is reached. If this is not the case, the system branches back to fire detection 61. In a subsequent assignment step, a change in smoke density is determined from a plurality of consecutive image data according to the method described above. If a critical measure for a change in smoke density is reached, an alarm is issued in a warning step 66. Otherwise, the system branches back to fire detection 61. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2016 207 705 A1

[0001]

Claims

[1] Method for fire detection by means of an evaluation of a scene (12) captured by a camera (10) via a machine learning system, characterized by that in the event that a fire is suspected based on an initial evaluation of the scene (12), a fire warning is only issued if an expected value for the smoke density and / or a change in smoke density exceeds a predetermined expected value for the smoke density. [2] Method according to claim 1, characterized by that, in order to derive the expected value for the smoke density, a scene is assigned to at least one class of scenes without determined smoke development or to one of at least two classes of scenes, each with different smoke density ranges. [3] Method according to one of the preceding claims, characterized by that a classification of the images is carried out across all or a given subset of the pixels of the images. [4] Method according to claim 3, characterized in that a selected image area of ​​an image of the scene is evaluated for a classification of the images. [5] Method according to claim 4, characterized by that a selected image area is defined for fire detection based on an automated analysis of the pixels of the image in such a way that an image area adjacent to a suspected fire is selected. [6] Method according to one of claims 4-5, characterized by that an expected value for a change in smoke density is derived from a change in the color of the image pixels in the selected image area over time. [7] Method according to one of the preceding claims, characterized by that the determined smoke density information is stored together with the image for advanced analysis and / or for training the machine learning system. [8] Fire detection device for carrying out a method according to one of the preceding claims, with an interface to a camera (10) for capturing images of a scene (12) and with an interface to a warning unit (50) for issuing a warning.

Images