Deep Learning-based Real-Time Object Recognition Fire Detection Method and System using Relative Change Rate

Abstract

A deep learning-based real-time object recognition fire detection method and system utilizing relative fluctuation rates are presented. The deep learning-based real-time object recognition fire detection method utilizing relative fluctuation rates proposed in the present invention includes a step of detecting and learning the shapes of objects in real time by performing deep learning-based real-time object recognition through an object recognition unit, and a step of detecting a fire using relative fluctuation rates of the learned objects, rather than fluid shapes including smoke and flames generated by the fire when a fire occurs, through an analysis unit.

Description

Technology Field

[0001] The present invention relates to a deep learning-based real-time object recognition fire detection method and system utilizing relative fluctuation rates. Background Technology

[0002] Over the past five years, an average of more than 40,000 fire incidents have occurred, resulting in 600 billion won in property damage. While the proportion of fire incidents is decreasing, property damage caused by fires continues to increase. Out of 400,103 dispatches for 119 life safety services, 38,110 were false alarms, accounting for 9.53% of the total. This is leading to a waste of firefighting resources and a decline in confidence in fire detectors.

[0003] As of 2020, the number of dispatches caused by malfunctioning fire detectors was 32,685, which accounted for 0.2% of actual fires. In the case of smoke detectors, they detect other types of smoke as fire smoke and activate, so false alarms can occur due to various causes such as dust and humidity.

[0004] With the Fourth Industrial Revolution, the smart city industry is growing rapidly as countries around the world have begun investing in smart cities. In particular, research is being conducted to develop fire notification services by building user interfaces using IoT technology to utilize information that previously relied on sensors to detect heat and smoke in existing fire alarms and alarms. Additionally, research is being carried out in various fields on technology that enables real-time remote simultaneous alerting for fires automatically detected by AI via CCTV.

[0005] Object recognition methods of object detection models according to conventional technology include R-CNN and YOLO.

[0006] R-CNN detects objects through two stages of object classification and bounding box regression for objects in CNN-based models, which is a 2-stage method. Although it is somewhat slower, it enables more accurate object detection.

[0007] YOLO is a 1-stage method, and YOLO-based models can detect objects by simultaneously performing object classification and localization, enabling fast real-time object detection, but it has the problem of being somewhat less accurate compared to CNN-based models, which are 2-stage methods.

[0008] In object detection technology for fire detection according to conventional technology, objects with clear shapes, such as safety helmets and people, can be detected with relatively high accuracy. However, since smoke or flames have unclear shapes and are fluid, fire detection performance can be significantly degraded depending on combustible materials and surrounding environmental conditions. Furthermore, there is a problem in that real-time object detection of smoke or flames is very slow and has low accuracy performance.

[0009] In object detection technology for fire detection according to conventional technology, AI-based automatic fire detection via CCTV has primarily been developed through object detection algorithms and has been applied in various fields, including not only smoke and flames but also face recognition and object recognition.

[0010] However, unlike ordinary objects and entities, fires possess fluid characteristics, and since the ignition forms of smoke or flames vary greatly depending on the location and time of occurrence, problems of performance degradation have mostly occurred.

[0011] Therefore, a new solution for AI-based fire image processing is needed to enhance the performance of real-time fire detection algorithms. Prior art literature

[0012] Korean Registered Patent No. 10-2045871 (2019.11.12) The problem to be solved

[0013] The technical problem that the present invention aims to solve is to provide a real-time intelligent object tracking fire detection method and system based on relative fluctuation rates to improve upon the problem of conventional fire detection methods that recognize actual fires and smoke as objects to issue fire alarms, which face difficulties in early determination or recognition of a fire due to low object recognition rates caused by the amorphous and fluid characteristics of fires and smoke. The fire detection method and system according to an embodiment of the present invention is a method that detects surrounding objects in real time rather than recognizing the fire itself; if the relative fluctuation rate of a performance indicator for an object decreases due to smoke or flames, etc., the current situation is recognized as a fire, and an alarm and the area suspected of being the fire location can be tracked. means of solving the problem

[0014] In one aspect, the fire detection method utilizing relative fluctuation rates for deep learning-based real-time object recognition proposed in the present invention includes the step of detecting and learning the shapes of objects in real time by performing deep learning-based real-time object recognition through an object recognition unit, and the step of detecting a fire using relative fluctuation rates of the learned objects, rather than fluid shapes including smoke and flames generated by the fire when a fire occurs, through an analysis unit.

[0015] The step of detecting and learning the shapes of objects in real time by performing deep learning-based real-time object recognition through the object recognition unit described above performs real-time object recognition in advance before a fire occurs in order to analyze changes in the object recognition rate due to fluid shapes including smoke and flames caused by a fire when a fire occurs.

[0016] The step of detecting a fire using the relative fluctuation rate of the learned objects, rather than the fluid form including smoke and flames generated by the fire through the analysis unit, recognizes that a fire has occurred when the relative fluctuation rate of the performance indicator representing the object recognition rate due to the fluid form including smoke and flames falls below a predetermined standard, and tracks the origin of the fire by analyzing the severity of the relative fluctuation rate of surrounding objects.

[0017] The step of detecting a fire using the relative variation rate of the learned objects, rather than the fluid form including smoke and flames generated by the fire when a fire occurs through the analysis unit described above, involves obtaining a performance indicator representing the object recognition rate to evaluate the performance of the object recognition model through True positive, True negative, False positive, and False negative in order to detect a fire using the relative variation rate of a performance indicator representing the object recognition rate, and tracking the fire origin by analyzing the severity of the relative variation rate of the performance indicator using Precision, which represents the ratio of actual correct answers among those predicted as correct by the object recognition model, and Recall, which represents the ratio of actual correct answers accurately detected by the object recognition model.

[0018] The step of detecting a fire using the relative variation rate of the learned objects, rather than the fluid form including smoke and flames generated by the fire when a fire occurs through the analysis unit described above, uses Average Precision (AP) to evaluate the performance of the object recognition model; calculates precision and recall through differential recall and converts them into a Precision-Recall curve (PR curve), calculates Average Precision (AP) through the area under the PR curve, and calculates Mean Average Precision (mAP) by summing the APs of each class of the object recognition model and dividing by the number of classes.

[0019] In one aspect, the fire detection system utilizing relative fluctuation rates for deep learning-based real-time object recognition proposed in the present invention includes an object recognition unit that performs deep learning-based real-time object recognition to detect and learn the shapes of objects in real time, and an analysis unit that detects a fire using relative fluctuation rates of the learned objects, rather than fluid shapes including smoke and flames generated by the fire when a fire occurs. Effects of the invention

[0020] According to embodiments of the present invention, the real-time intelligent object tracking fire detection method and system based on relative fluctuation rates can improve upon the problem that, in the case of conventional fire detection methods that recognize actual fires and smoke as objects to issue fire alarms, the object recognition rate is low and it is difficult to determine or recognize a fire early because fires and smoke have amorphous and fluid characteristics. The fire detection method and system according to the embodiments of the present invention detects surrounding objects in real time rather than recognizing the fire itself; if the relative fluctuation rate of a performance indicator for an object decreases due to smoke or flames, etc., the current situation is recognized as a fire, and an alarm and the area suspected of being the fire location can be tracked. Brief explanation of the drawing

[0021] Figure 1 is a diagram illustrating a method for classifying whether a fire has occurred through a fire image according to the prior art. Figure 2 is a diagram illustrating a real-time fire detection method based on a CNN algorithm according to the prior art. FIG. 3 is a flowchart illustrating a fire detection method utilizing a relative fluctuation rate for deep learning-based real-time object recognition according to an embodiment of the present invention. FIG. 4 is a diagram showing the configuration of a fire detection system utilizing a relative fluctuation rate for deep learning-based real-time object recognition according to an embodiment of the present invention. Figure 5 is a diagram showing a conceptual diagram of a fire detection method according to the prior art. FIG. 6 is a diagram showing a conceptual diagram of a fire detection method according to one embodiment of the present invention. Specific details for implementing the invention

[0022] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

[0024] Figure 1 is a diagram illustrating a method for classifying whether a fire has occurred through a fire image according to the prior art.

[0025] FIG. 1(a) is a diagram showing the configuration of a CNN classification model according to the prior art, and FIG. 1(b) is a diagram showing the result of determining whether a fire occurred through a CNN classification model according to the prior art.

[0026] The CNN classification model according to the prior art detects objects through two stages of object classification and bounding box regression, and although the speed is somewhat slow, it enables more accurate object detection.

[0027] The CNN classification model according to the prior art illustrated in Fig. 1(a) uses a Convolutional Neural Network (CNN) algorithm that takes fire, non-fire, and smoke images as input, extracts image features, and automatically classifies whether there is a fire. While this method has high accuracy in cases where there are clear differences between images, it has the disadvantage of reduced classification accuracy in situations where it is difficult to determine whether there is a fire in the image.

[0028] Referring to the example (110) in which object classification was successful as illustrated in FIG. 1(b), the original images (111, 112, 113) match the case where a human recognized them and the CNN classification model match.

[0029] On the other hand, referring to the example (120) where object classification failed as illustrated in FIG. 1(b), the original images (121, 122, 123) do not match the CNN classification model as perceived by a human.

[0031] Figure 2 is a diagram illustrating a real-time fire detection method based on a CNN algorithm according to the prior art.

[0032] An object detection model developed based on a CNN algorithm takes a fire image as input, extracts features from the input image in the backbone stage, and finds the location of objects based on the features in the neck stage. Then, it generates bounding boxes in the dense prediction stage to detect input objects in real time.

[0033] However, unlike ordinary objects and entities, fires possess fluid characteristics, and because the form of ignition varies greatly depending on the location and time of occurrence, problems of performance degradation have mostly arisen.

[0034] As such, while the conventional CNN algorithm-based real-time fire detection method can detect objects with clear shapes, such as safety helmets and people, with relatively high accuracy, smoke or flames lack clear shapes and are fluid; therefore, fire detection performance can be significantly degraded depending on combustible materials and surrounding environmental conditions. Furthermore, there is a problem in that real-time object detection for smoke or flames is very slow and has low accuracy.

[0036] FIG. 3 is a flowchart illustrating a fire detection method utilizing a relative fluctuation rate for deep learning-based real-time object recognition according to an embodiment of the present invention.

[0037] The proposed fire detection method utilizing relative fluctuation rates for real-time object recognition based on deep learning includes a step (310) of detecting and learning the shapes of objects in real time by performing real-time object recognition based on deep learning through an object recognition unit, and a step (420) of detecting a fire using relative fluctuation rates of the learned objects, rather than fluid shapes including smoke and flames caused by the fire, through an analysis unit when a fire occurs.

[0038] In step (310), real-time object recognition based on deep learning is performed through the object recognition unit to detect and learn the shapes of objects in real time.

[0039] According to an embodiment of the present invention, real-time object recognition is performed in advance before a fire occurs in order to analyze changes in the object recognition rate due to fluid forms including smoke and flames generated by a fire when a fire occurs.

[0040] In step (320), the fire is detected through the analysis unit using the relative fluctuation rate of the learned objects, rather than a fluid form including smoke and flames generated by the fire when a fire occurs.

[0041] According to an embodiment of the present invention, if the relative fluctuation rate of a performance indicator representing the object recognition rate falls below a predetermined standard due to a fluid form including smoke and flames, it is recognized that a fire has occurred, and the origin of the fire is tracked by analyzing the severity of the relative fluctuation rate of surrounding objects.

[0042] According to an embodiment of the present invention, a performance indicator representing an object recognition rate is obtained to evaluate the performance of an object recognition model through true positive, true negative, false positive, and false negative in order to detect a fire using the relative fluctuation rate of a performance indicator representing an object recognition rate.

[0043] According to an embodiment of the present invention, the degree of severe relative variation of the performance indicator is analyzed using Precision, which represents the ratio of actual correct answers among those predicted as correct answers by the object recognition model, and Recall, which represents the ratio of actual correct answers accurately detected by the object recognition model, to track the fire origin.

[0044] According to an embodiment of the present invention, Average Precision (AP) is used to evaluate the performance of an object recognition model. Precision and recall are calculated through differential recall and converted into a Precision-Recall curve (PR curve), and Average Precision (AP) is calculated through the area under the PR curve. The APs of each class of the object recognition model are all added together and then divided by the number of classes to calculate Mean Average Precision (mAP).

[0046] FIG. 4 is a diagram showing the configuration of a fire detection system utilizing a relative fluctuation rate for deep learning-based real-time object recognition according to an embodiment of the present invention.

[0048] The fire detection system (400) according to the present embodiment may include a processor (410), a bus (420), a network interface (430), memory (440), and a database (450). The memory (440) may include an operating system (441) and a fire detection routine (442) utilizing a relative variation rate for deep learning-based real-time object recognition. The processor (410) may include an object recognition unit (411) and an analysis unit (412). In other embodiments, the fire detection system (400) may include more components than those of FIG. 4. However, it is not necessary to clearly illustrate most of the prior art components. For example, the fire detection system (400) may include other components such as a display or a transceiver.

[0049] Memory (440) is a computer-readable recording medium and may include a non-perishable permanent mass storage device such as RAM (random access memory), ROM (read only memory), and a disk drive. Additionally, program code for an operating system (441) and a fire detection routine (442) utilizing relative fluctuation rates for deep learning-based real-time object recognition may be stored in memory (440). These software components may be loaded from a computer-readable recording medium separate from memory (440) using a drive mechanism (not shown). This separate computer-readable recording medium may include computer-readable recording media (not shown), such as a floppy drive, disk, tape, DVD / CD-ROM drive, or memory card. In another embodiment, software components may be loaded into memory (440) via a network interface (430) rather than a computer-readable recording medium.

[0050] The bus (420) can enable communication and data transmission between components of the fire detection system (400). The bus (420) can be configured using a high-speed serial bus, a parallel bus, a Storage Area Network (SAN), and / or other suitable communication technology.

[0051] The network interface (430) may be a computer hardware component for connecting the fire detection system (400) to a computer network. The network interface (430) may connect the fire detection system (400) to a computer network via a wireless or wired connection.

[0052] The database (450) can serve to store and maintain all information necessary for fire detection using the relative fluctuation rate of deep learning-based real-time object recognition. Although FIG. 4 illustrates the database (450) being built and included inside the fire detection system (400), it is not limited thereto and may be omitted depending on the system implementation method or environment, or it is also possible for all or part of the database to exist as an external database built on a separate system.

[0053] The processor (410) may be configured to process instructions of a computer program by performing basic arithmetic, logic, and input / output operations of the fire detection system (400). Instructions may be provided to the processor (410) via memory (440) or a network interface (430) and via a bus (420). The processor (410) may be configured to execute program code for an object recognition unit (411) and an analysis unit (412). Such program code may be stored in a recording device such as memory (440).

[0054] The object recognition unit (411) and the analysis unit (412) may be configured to perform the steps (310–320) of FIG. 1.

[0055] The fire detection system (400) may include an object recognition unit (411) and an analysis unit (412).

[0056] The object recognition unit (411) according to an embodiment of the present invention performs real-time object recognition based on deep learning to detect and learn the shapes of objects in real time.

[0057] The object recognition unit (411) according to an embodiment of the present invention performs real-time object recognition in advance before a fire occurs in order to analyze changes in the object recognition rate due to fluid forms including smoke and flames generated by a fire when a fire occurs.

[0058] The analysis unit (412) according to an embodiment of the present invention detects a fire using the relative fluctuation rate of the learned objects, rather than a fluid form including smoke and flames generated by the fire when a fire occurs.

[0059] The analysis unit (412) according to an embodiment of the present invention recognizes that a fire has occurred when the relative fluctuation rate of a performance indicator representing an object recognition rate due to a fluid form including smoke and flames falls below a predetermined standard, and tracks the origin of the fire by analyzing the degree of severe relative fluctuation rate of surrounding objects.

[0060] The analysis unit (412) according to an embodiment of the present invention obtains a performance indicator representing an object recognition rate to evaluate the performance of an object recognition model through true positive, true negative, false positive, and false negative in order to detect a fire using the relative fluctuation rate of a performance indicator representing an object recognition rate.

[0061] The analysis unit (412) according to an embodiment of the present invention tracks the fire origin by analyzing the degree of severe relative fluctuation of the performance indicator using precision, which represents the ratio of the actual correct answer among the correct answers predicted by the object recognition model, and recall, which represents the ratio of the actual correct answers accurately detected by the object recognition model.

[0062] The analysis unit (412) according to an embodiment of the present invention uses AP (Average Precision) to evaluate the performance of an object recognition model, calculates precision and recall through differential recall for the recall, converts it into a PR curve (Precision-Recall curve), calculates AP (Average Precision) through the area under the PR curve, and calculates mAP (Mean Average Precision) by adding all the APs of each class of the object recognition model and dividing by the number of classes.

[0064] Figure 5 is a diagram showing a conceptual diagram of a fire detection method according to the prior art.

[0065] Referring to FIG. 5, an image (520) showing the fluid appearance of smoke over time is shown in the initial image (510) of the fire outbreak.

[0066] In conventional fire detection methods, fire is detected by identifying fire and smoke within an image. However, since smoke, flames, and fire are not physical objects, they exhibit a fluid appearance, which can lead to a decrease in detection rates.

[0067] Accordingly, the present invention proposes a method for detecting surrounding objects in real-time prior to the occurrence of a fire, and recognizing a fire when the detection rate of objects relatively drops due to smoke, flames, fire, etc. Additionally, the present invention can track the origin of a fire by analyzing the severity of the relative fluctuation rate of the detection rate for surrounding objects.

[0069] FIG. 6 is a diagram showing a conceptual diagram of a fire detection method according to one embodiment of the present invention.

[0070] Referring to FIG. 6, images (620, 630, 640) showing the fluid appearance of smoke over time are shown in the order of arrows, starting from the initial image (610) of the fire outbreak.

[0071] According to an embodiment of the present invention, instead of a conventional method of directly detecting fire, the method determines whether a fire exists by utilizing the real-time object recognition error of surrounding objects.

[0072] According to an embodiment of the present invention, not only can a fire be detected, but the ignition origin where the fire starts can also be estimated. That is, a new method for detecting a fire based on the relative variation rate of a performance indicator is proposed, based on the performance indicator of an object recognition model.

[0073] According to an embodiment of the present invention, a performance indicator is provided that can determine the performance of an object recognition model through true positive, true negative, false positive, and false negative to detect a fire using the relative variation rate of a performance indicator representing an object recognition rate.

[0074] Precision represents the ratio of actual correct answers among those predicted as correct by the model, and Recall represents the ratio of actual correct answers accurately detected by the model. The calculation process for precision and recall is expressed in Equation (1) and Equation (2).

[0075] (1)

[0076] (2)

[0077] While precision and recall are useful for understanding model performance, it is difficult to quantitatively evaluate model performance due to their inverse relationship.

[0078] According to an embodiment of the present invention, Average Precision (AP) is used to quantitatively evaluate the performance of an object recognition model. AP is calculated by calculating precision and recall through differential recall, converting them into a Precision-Recall curve, and then calculating the Average Precision (AP) through the area under the PR curve.

[0079] Subsequently, the APs of each class of the object recognition model are all added together and divided by the number of classes to calculate mAP (Mean Average Precision). Equation (3) represents AP, and Equation (4) represents mAP.

[0080] (3)

[0081] (4)

[0082] According to an embodiment of the present invention, it is determined that it can be very usefully utilized in places where people do not usually reside but where property damage from fire is likely to be relatively high, such as data centers, factories, ESS systems, and outdoor hazardous material storage facilities, and is characterized by being able to predict the fire ignition point and track it through video equipment in areas where the relative fluctuation rate of performance indicators of surrounding objects changes relatively severely.

[0083] According to an embodiment of the present invention, the model can be configured in various ways depending on the size of each fire room, the amount of combustible material, the fire risk, etc., and is characterized by being applicable to various object detection models that detect objects more accurately.

[0084] According to an embodiment of the present invention, it is applicable to various video transmission equipment such as existing CCTV, PTZ equipment, and cameras, and is characterized by being able to transmit an alarm or video images of the situation at the site to an administrator or occupant by linking with a smartphone app and alarm facilities when a fire is detected.

[0086] The device described above may be implemented as a hardware component, a software component, and / or a combination of a hardware component and a software component. For example, the device and components described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include a plurality of processing elements and / or a plurality of types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are also possible.

[0087] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or instruct the processing unit independently or collectively. Software and / or data may be embodied in any type of machine, component, physical device, virtual equipment, computer storage medium, or device so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0088] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiment, or they may be those known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

[0089] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results can be achieved even if the described techniques are performed in a different order than described, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0090] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

Claim 1 The method comprises: a step of performing deep learning-based real-time object recognition before a fire occurs through an object recognition unit including an object recognition model, detecting and learning the shapes of objects, which are non-fire objects existing within the space, from an image of the space through bounding box prediction, thereby obtaining a reference object recognition rate for the objects, which are non-fire objects; and a step of detecting a fire through an analysis unit using the relative fluctuation rate of the object recognition rate that decreases as the objects are obscured by a fluid shape including smoke and flames generated by the fire when a fire occurs, wherein the step of detecting a fire does not directly detect smoke and flames, but instead performs real-time object recognition on the objects through the object recognition unit to determine whether a fire has occurred, and calculates a performance indicator representing the object recognition rate for the objects in real time based on the precision and recall of the current time point; and a step of recognizing that a fire has occurred if, as the objects are obscured by the fluid shape including smoke and flames, the relative fluctuation rate of the performance indicator representing the real-time object recognition rate compared to the reference object recognition rate falls below a predetermined threshold. A fire detection method comprising the step of comparing the relative fluctuation rate of the performance indicator for each of the objects, analyzing the location of the object with the most severe fluctuation in the performance indicator among the objects, and tracking the fire origin within the space. Claim 2 delete Claim 3 delete Claim 4 A fire detection method according to claim 1, wherein the step of calculating the performance indicators in real time includes the step of calculating the precision, which is the ratio of the actual correct answer among the correct answers predicted by the object recognition model, and the recall, which is the ratio of the actual correct answers accurately detected by the object recognition model, respectively, using true positives, false positives, and false negatives according to the prediction results of the object recognition model. Claim 5 A fire detection method according to claim 4, wherein the performance indicator is mAP (Mean Average Precision), and the step of calculating the performance indicator in real time comprises: generating a PR curve (Precision-Recall curve) using the calculated precision and recall rate, and calculating the area under the PR curve through differential recall for the recall rate to obtain AP (Average Precision); and calculating mAP by summing the APs for each class of the object recognition model and dividing by the number of classes. Claim 6 An object recognition unit that performs deep learning-based real-time object recognition before a fire occurs through an object recognition model, detects and learns the shapes of objects, which are non-fire objects existing within the space, from an image of the space through bounding box prediction, and obtains a reference object recognition rate for the objects, which are non-fire objects; A fire detection system comprising an analysis unit that detects a fire by utilizing the relative fluctuation rate of an object recognition rate that decreases as the objects are obscured by a fluid shape including smoke and flames generated by the fire when a fire occurs, wherein the analysis unit detects the fire without directly detecting the smoke and flames, and performs real-time object recognition of the objects through the object recognition unit to determine whether a fire has occurred, calculates a performance indicator representing the object recognition rate of the objects in real time based on the precision and recall of the current time point, and recognizes that a fire has occurred when the relative fluctuation rate of the performance indicator representing the real-time object recognition rate relative to the reference object recognition rate falls below a predetermined standard as the objects are obscured by the fluid shape including the smoke and flames, and tracks the origin of the fire within the space by analyzing the location of the object with the most severe fluctuation in the performance indicator among the objects by comparing the relative fluctuation rate of the performance indicator for each of the objects. Claim 7 delete Claim 8 delete Claim 9 A fire detection system according to claim 6, wherein the analysis unit calculates the precision, which is the ratio of the actual correct answer among the correct answers predicted by the object recognition model, and the recall, which is the ratio of the actual correct answers accurately detected by the object recognition model, respectively, using true positives, false positives, and false negatives based on the prediction results of the object recognition model. Claim 10 A fire detection system according to claim 9, wherein the performance indicator is mAP (Mean Average Precision), and the analysis unit generates a PR curve (Precision-Recall curve) using the calculated precision and recall rate, calculates the area under the PR curve through differential recall for the recall rate to obtain AP (Average Precision), and calculates mAP by summing the APs for each class of the object recognition model and dividing by the number of classes.

Images