The Method and System for Real-Time Detection of Specific Events Based on an Edge Module with Reduced Computational Load and Improved Detection Accuracy

Abstract

The present invention relates to a method and system for real-time detection of specific events based on an edge module with reduced computational load and improved detection accuracy, wherein a target embedding vector for each image frame is derived from a series of image frames of some images acquired from a camera module through a deep learning-based embedding model, a similarity between a reference embedding vector stored for each predefined event and the target embedding vector is calculated, and the occurrence of a specific event is finally determined based on the similarity, thereby drastically minimizing the computational load of the edge device and simultaneously performing high-precision real-time detection robust to environmental noise.

Description

Technology Field

[0001] The present invention relates to a method and system for real-time detection of specific events based on an edge module with reduced computational load and improved detection accuracy, wherein a target embedding vector for each image frame is derived from a series of image frames of some images acquired from a camera module through a deep learning-based embedding model, a similarity between a reference embedding vector stored for each predefined event and the target embedding vector is calculated, and the occurrence of a specific event is finally determined based on the similarity, thereby drastically minimizing the computational load of the edge device and simultaneously performing high-precision real-time detection robust to environmental noise. Background Technology

[0003] Recently, intelligent video detection systems have established themselves as a core infrastructure that goes beyond simple video recording to detect dangerous situations (falls, assaults, fires, etc.) in real time and support immediate initial response. Traditional object detection algorithms (such as YOLO and R-CNN) require complex computational power and vast amounts of data, making implementation possible only in server-intensive architectures; however, this presents limitations such as network latency and increased deployment costs.

[0004] As prior art for solving these problems, US Patent No. 12,367,677 (US 12,367,677 B1) proposes a technology that converts the features of an input image into high-dimensional numerical vector embeddings through a video embedding network within an edge device, and identifies events by comparing them with a stored reference embedding using cosine similarity, etc.

[0005] However, such existing embedding-based detection methods expose the following technical limitations. First, performing real-time inference on edge devices with performance constraints still incurs computational load; in particular, there is an aspect of resource waste due to the continuous computation performed even on static frames without significant movement. Second, relying on similarity results from a single frame or short intervals leads to a problem where real-time detection accuracy is degraded, as false positives are prone to occur when similarity temporarily increases due to environmental noise such as rain, snow, or changes in lighting.

[0006] Therefore, there is a need to develop technology that can reduce computational load while simultaneously improving event detection accuracy. Prior art literature

[0008] US Registered Patent Publication US 12,367,677 (July 22, 2025) US Published Patent Publication US 2025 / 0191372 (June 12, 2025). US Registered Patent Publication US 11,074,791 (July 27, 2021) US Published Patent Publication US 2025 / 0310585 (October 2, 2025) US Registered Patent Publication US 12,079,272 (September 3, 2024) US Published Patent Publication US 2021 / 0117697 (April 22, 2021) US Published Patent Publication US 2026 / 0011149 (January 8, 2026) US Registered Patent Publication US 12,067,762 (2024.08.20.) U.S. Patent Publication US 12,563,053 (2026.02.24.) U.S. Patent Publication US 12,340,582 (2025.06.24.) The problem to be solved

[0009] The present invention aims to provide a method and system for real-time detection of specific events based on an edge module with reduced computational load and improved detection accuracy, wherein the method derives a target embedding vector for each image frame through a deep learning-based embedding model from a series of image frames of some images acquired from a camera module, calculates the similarity between a reference embedding vector stored for each predefined event and the target embedding vector, and finally determines the occurrence of a specific event based on the similarity, thereby drastically minimizing the computational load of the edge device and simultaneously performing high-precision real-time detection robust to environmental noise. means of solving the problem

[0011] In order to solve the above problems, an embodiment of the present invention provides a method for detecting a specific event in real time performed in an image detection system, wherein the image detection system includes a camera module and an edge module comprising one or more processors and one or more memories; an image capturing step in which an image is captured by the camera module and transmitted to the edge module; a vector generation step in which a target embedding vector for each image frame is generated by the edge module by inputting each of a predetermined number of consecutive image frames among the image into a deep learning-based embedding model; a similarity calculation step in which a similarity between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame is calculated by the edge module; and an event judgment step in which, when the ratio of image frames having a similarity exceeding a predetermined threshold among the consecutive image frames exceeds a predetermined ratio standard, it is determined that a specific event corresponding to the similarity has occurred within the image.

[0012] In one embodiment of the present invention, the reference embedding vector may be characterized by being derived as the center point of the learning embedding vectors derived from a plurality of learning image frames for each specific event included in the learning data for the embedding model.

[0013] In one embodiment of the present invention, the threshold value may be characterized by being adaptively set to correspond to the environment based on a similarity distribution that appears differently depending on the shooting environment of the camera module.

[0014] In one embodiment of the present invention, the threshold value may be calculated by summing the average value of a similarity distribution derived based on a plurality of similarity data obtained in a normal period in which no event occurs, and the value obtained by multiplying the standard deviation of the similarity distribution by a preset constant.

[0015] In one embodiment of the present invention, the threshold value may be characterized by being set differently from each other by independently reflecting the similarity distribution that appears differently for each individual event when there are multiple predefined events.

[0016] In one embodiment of the present invention, the vector generation step may include: a change area extraction step that extracts a change area where movement occurs based on image difference operations between adjacent image frames among the consecutive image frames, or a pre-set region-specific motion vector for the consecutive image frames; and a local vector derivation step that derives the target embedding vector by inputting only the image of the change area into the embedding model.

[0017] In one embodiment of the present invention, the vector generation step may omit the derivation of the target embedding vector and the calculation of the similarity for image frames in which the magnitude of the movement is less than a preset change standard.

[0018] In one embodiment of the present invention, the reference embedding vector is stored in advance along with a reference difference average value calculated by determining the difference between a plurality of learning embedding vectors derived from a plurality of learning image frames for each specific event included in the learning data for the embedding model, and the similarity calculation step further comprises: an embedding difference calculation step that calculates a delta embedding difference, which is the difference between target embedding vectors between adjacent image frames among the consecutive image frames; and a pattern analysis step that calculates a target difference average value, which is the average value of a plurality of delta embedding differences calculated from the consecutive image frames, and calculates pattern similarity by comparing the target difference average value with the reference difference average value; and the event judgment step may further determine whether the pattern similarity exceeds a preset pattern threshold value to finally determine whether the event occurs.

[0019] In one embodiment of the present invention, the embedding model may be characterized as an artificial neural network model trained to output an embedding vector value having a vector similarity greater than or equal to a mutually preset standard when an event labeled in an image frame is the same specific event, even if an image frame included in a video captured in a different camera environment is input.

[0020] In one embodiment of the present invention, the preset number may be determined as the number of image frames acquired during a preset time interval in the image.

[0021] In one embodiment of the present invention, the similarity can be calculated through cosine similarity.

[0022] In one embodiment of the present invention, the detection method may be characterized by being performed solely by similarity comparison operations between embedding vectors without object detection operations that classify objects within the image and track bounding boxes, thereby minimizing the computational load of the edge module.

[0023] In order to solve the above-mentioned problem, an embodiment of the present invention provides an image detection system that performs a method for detecting a specific event in real time, comprising: a camera module including an image capturing unit that captures an image and transmits it to an edge module; and an edge module including one or more processors and one or more memories; wherein the edge module comprises: a vector generation unit that inputs each of a predetermined number of consecutive image frames among the image into a deep learning-based embedding model to generate a target embedding vector for each image frame; a similarity calculation unit that calculates a similarity between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame; and an event judgment unit that determines that a specific event corresponding to the similarity has occurred within the image when the ratio of image frames having a similarity exceeding a predetermined threshold among the consecutive image frames exceeds a predetermined ratio standard.

[0024] In one embodiment of the present invention, the vector generation unit may include: a change area extraction unit that extracts a change area where movement occurs based on image difference operations between adjacent image frames among the consecutive image frames, or a pre-set area-specific motion vector for the consecutive image frames; and a local vector derivation unit that derives the target embedding vector by inputting only the image of the change area into the embedding model.

[0025] In one embodiment of the present invention, the vector generation unit may omit the derivation of the target embedding vector and the calculation of the similarity for image frames in which the magnitude of the movement is less than a preset change standard.

[0026] In one embodiment of the present invention, the reference embedding vector is stored together with a reference difference average value calculated by determining the difference between a plurality of learning embedding vectors derived from a plurality of learning image frames for each specific event included in the learning data for the embedding model, and the similarity calculation unit further includes: an embedding difference calculation unit that calculates a delta embedding difference, which is the difference between target embedding vectors between adjacent image frames among the consecutive image frames; and a pattern analysis unit that calculates a target difference average value, which is the average value of a plurality of delta embedding differences calculated from the consecutive image frames, and calculates pattern similarity by comparing the target difference average value with the reference difference average value, and the event judgment unit can finally determine whether the event occurs by additionally determining whether the pattern similarity exceeds a preset pattern threshold value.

[0027] In order to solve the above problems, an embodiment of the present invention provides a method for detecting a specific event in real time performed in an image detection system, wherein the image detection system includes an edge module comprising one or more processors and one or more memories, and the edge module includes: an image reception step of receiving an image captured by a camera module; a vector generation step of generating a target embedding vector for each image frame by inputting each of a predetermined number of consecutive image frames among the image into a deep learning-based embedding model by the edge module; a similarity calculation step of calculating a similarity between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame by the edge module; and an event judgment step of determining that a specific event corresponding to the similarity has occurred within the image when the ratio of image frames having a similarity exceeding a predetermined threshold among the consecutive image frames exceeds a predetermined ratio standard by the edge module.

[0028] In order to solve the above problem, in one embodiment of the present invention, a computer-readable storage medium for implementing a method for detecting a specific event in real time in an image detection system comprising a camera module and an edge module comprising one or more processors and one or more memories, wherein the computer-readable storage medium comprises computer-executable instructions that cause the image detection system to perform the following steps, and the following steps include: an image capturing step in which an image is captured by the camera module and transmitted to the edge module; a vector generation step in which a predetermined number of consecutive image frames among the image are input into a deep learning-based embedding model to generate a target embedding vector for each image frame by the edge module; and a similarity calculation step in which a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame are calculated by the edge module. The present invention provides a computer-readable storage medium comprising: an event determination step in which, by means of an edge module, if the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard, a specific event corresponding to the similarity has occurred within the image. Effects of the invention

[0030] In one embodiment of the present invention, heavy object detection operations that classify objects in an image and track bounding boxes are excluded, and events are detected using only similarity comparison operations between embedding vectors, thereby drastically reducing the computational load of the edge module and enabling real-time independent operation without delay even on low-spec camera terminals.

[0031] In one embodiment of the present invention, by applying a time-series verification logic that determines whether the ratio of frames in which the similarity exceeds a threshold among consecutive image frames exceeds a preset ratio standard, without relying solely on the similarity result of a single frame, it is possible to fundamentally block false detections caused by temporary shaking or environmental noise and maximize detection accuracy.

[0032] In one embodiment of the present invention, the edge module compares and calculates a reference embedding vector and a target embedding vector for each of a plurality of predefined events (falling, assault, fire, etc.) in parallel, thereby enabling the simultaneous and independent detection of various dangerous situations that may occur in the field using only a single lightweight deep learning model, which can enhance the scalability and versatility of the system.

[0033] In one embodiment of the present invention, by performing a double verification process that calculates the average value of the delta embedding difference, which is the difference between target embedding vectors between adjacent image frames, and compares it with a reference pattern, it is possible to precisely analyze dynamic change patterns such as speed and acceleration beyond simple morphological similarity of objects, thereby significantly improving the ability to distinguish between falling or assault acts of a person.

[0034] In one embodiment of the present invention, by variably setting an adaptive threshold value by reflecting the normal similarity distribution (mean and standard deviation) that appears differently depending on the camera's shooting environment, it is possible to achieve the effect of ensuring uniform and stable detection performance even in poor installation environments, such as rapid changes in illumination between day and night or adverse weather conditions (snow, rain, etc.).

[0035] In one embodiment of the present invention, instead of performing heavy deep learning operations on the entire image, only local change areas where macroscopic movement occurs are lightly extracted by utilizing motion vectors calculated during the image encoding process or simple image difference operations between frames, and then input into an embedding model, and the deep learning inference operations themselves are completely omitted for static frames where no significant movement is detected, thereby minimizing unnecessary power consumption and memory occupation and maximizing the computing resource efficiency of edge devices. Brief explanation of the drawing

[0037] FIG. 1 schematically illustrates the overall operation concept of an image detection system according to one embodiment of the present invention. FIG. 2 schematically illustrates the internal configuration of an image detection system and the execution steps of a specific event detection method according to an embodiment of the present invention. FIG. 3 schematically illustrates integrated and separate implementation forms of a camera module and an edge module according to one embodiment of the present invention. FIG. 4 schematically illustrates the process of deriving a reference embedding vector for each specific event according to one embodiment of the present invention. FIG. 5 schematically illustrates a method for calculating the similarity between a target embedding vector and a reference embedding vector according to an embodiment of the present invention. FIG. 6 schematically illustrates a concept in which a threshold value is applied variably according to changes in the shooting environment according to an embodiment of the present invention. FIG. 7 schematically illustrates the process of calculating a threshold value according to the shooting environment based on a normal similarity distribution according to an embodiment of the present invention. FIG. 8 schematically illustrates a threshold value determined according to the mean value and standard deviation of a similarity distribution according to one embodiment of the present invention. FIG. 9 schematically illustrates a threshold table that is independently set for a plurality of specific events according to one embodiment of the present invention. FIG. 10 schematically illustrates an embedding model learned for high robustness in different camera environments according to one embodiment of the present invention. FIG. 11 schematically illustrates the process of extracting a change area according to one embodiment of the present invention. FIG. 12 schematically illustrates the process of deriving a local target embedding vector using a change region according to one embodiment of the present invention. FIG. 13 schematically illustrates the conditions under which embedding operations and similarity calculations are omitted for an image frame according to an embodiment of the present invention. FIG. 14 schematically illustrates the effect of reducing computational resources through the extraction of change regions and the omission of computation according to one embodiment of the present invention. FIG. 15 schematically illustrates the process of deriving a reference difference average value according to one embodiment of the present invention. FIG. 16 schematically illustrates the process of deriving the average difference value according to one embodiment of the present invention. FIG. 17 schematically illustrates a process of finally determining whether an event occurs by further comparing pattern similarity according to an embodiment of the present invention. FIG. 18 schematically illustrates a predetermined number of consecutive image frames acquired during a predetermined time interval according to one embodiment of the present invention. FIG. 19 schematically illustrates a principle for reducing the computational load of an edge module according to one embodiment of the present invention. FIG. 20 illustrates, in an exemplary manner, the internal configuration of a computing device according to one embodiment of the present invention. Specific details for implementing the invention

[0038] Hereinafter, various embodiments and / or aspects are disclosed with reference to the drawings. For illustrative purposes, numerous specific details are disclosed in the following description to aid in a general understanding of one or more aspects. However, it will also be recognized by those skilled in the art that these aspects may be practiced without such specific details. The following description and the accompanying drawings describe specific exemplary aspects of one or more aspects in detail. However, these aspects are exemplary, and some of the various methods in the principles of the various aspects may be used, and the description is intended to include all such aspects and their equivalents.

[0040] In addition, various aspects and features will be presented by a system that may include multiple devices, components and / or modules, etc. It should also be understood and recognized that various systems may include additional devices, components and / or modules, etc., and / or may not include all of the devices, components, modules, etc. discussed in relation to the drawings.

[0041] Terms such as “embodiment,” “example,” “aspect,” “example,” etc. as used herein may not be interpreted as implying that any aspect or design described is superior or more advantageous than other aspects or designs. Terms used below, such as “part,” “component,” “module,” “system,” “interface,” etc., generally refer to computer-related entities and may, for example, refer to hardware, a combination of hardware and software, or software.

[0042] Additionally, the terms “comprising” and / or “comprising” should be understood to mean that the relevant feature and / or component is present, but not to exclude the presence or addition of one or more other features, components and / or groups thereof.

[0043] Additionally, terms including ordinal numbers, such as first, second, etc., may be used to describe various components, but said components are not limited by said terms. Such terms are used solely for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be named the second component, and similarly, the second component may be named the first component. The term "and / or" includes a combination of a plurality of related described items or any of a plurality of related described items.

[0044] Furthermore, in the embodiments of the present invention, all terms used herein, including technical or scientific terms, unless otherwise defined, have the same meaning as generally understood by those skilled in the art to which the present invention pertains. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in the embodiments of the present invention.

[0046] FIG. 1 schematically illustrates the overall operation concept of an image detection system (1) according to one embodiment of the present invention.

[0048] Specifically, an image detection system (1) according to one embodiment of the present invention may be configured to include a camera module (1000) installed at a specific location to acquire an image, and an edge module (2000) that analyzes image data received from the camera module (1000) in real time to determine whether a specific event has occurred.

[0049] In one embodiment, the camera module (1000) can capture a surveillance target area and generate image data composed of continuous image frames. More specifically, the camera module (1000) can be implemented as various types of imaging devices, such as commonly used security CCTVs, IP cameras, and webcams.

[0050] Meanwhile, the present invention excludes heavy object detection algorithms that track minute object shapes at the pixel level and compares the overall similarity and time-series patterns between embedding vectors, thereby providing strong compatibility that enables stable and seamless event detection not only in high-resolution images but also in low-resolution, outdated CCTV environments installed in existing industrial sites. The captured video data is streamed in real-time to an edge module (2000) via a wired or wireless network or is transmitted directly through a physical interface.

[0052] The edge module (2000) is an independent computing hardware device based on edge computing technology that can be deployed at a site adjacent to the camera module (1000) or integrated inside the camera module (1000). Looking at the overall data flow of FIG. 1, the system of the present invention has an on-device AI architecture in which the analysis and judgment of an image are completed at the edge computing-based edge module (2000) without the intervention of a cloud or a central server. The edge module (2000) passes the image frames of the received image through an internal deep learning-based embedding model to convert them into a numeric vector (target embedding vector of the present invention) in which the visual features of the image are compressed to a high dimension.

[0053] Subsequently, the edge module (2000) retrieves reference embedding vectors for specific events (e.g., falling, assault, fire, etc.) that are pre-built in an internal or external database (DB), and performs a comparison operation with the extracted target embedding vector. As a result of the comparison, if it is determined that the distance between the embedding vectors is close and the similarity is high, an alarm indicating the occurrence of the corresponding event is immediately output.

[0054] The image detection system (1) of the present invention fundamentally breaks away from the existing server-centric structure that transmits all high-capacity image data to a central server to perform heavy object classification operations. By having an innovative architecture that immediately recognizes the context of a situation with only very light embedding operations at the field level, it can drastically reduce network bandwidth usage and provide a systemic foundation for monitoring dangerous situations in real time without communication delay.

[0056] Meanwhile, although not illustrated in FIG. 1, a central management server capable of receiving and managing event result information determined from edge modules according to an embodiment is connected to a plurality of edge modules (2000) installed at multiple sites via a wired or wireless network, and can perform the function of collecting event occurrence alarms determined independently by each edge module (2000) and transmitting them to a control agent's terminal or storing them in a separate log database.

[0058] A detailed explanation for implementing this will be provided below.

[0060] FIG. 2 schematically illustrates the internal configuration of an image detection system (1) and the execution steps of a specific event detection method according to one embodiment of the present invention.

[0062] As illustrated in FIG. 2, a method for detecting a specific event in real time performed by an image detection system (1) comprises: an image detection system (1) including a camera module (1000) and an edge module (2000) including one or more processors and one or more memories; an image capturing step (S100) in which an image is captured by the camera module (1000) and transmitted to the edge module (2000); a vector generation step (S200) in which a predetermined number of consecutive image frames among the image are input into a deep learning-based embedding model to generate a target embedding vector for each image frame by the edge module (2000); and a similarity calculation step (S300) in which a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame are calculated by the edge module (2000). The method includes an event judgment step (S400) in which, by the edge module (2000), if the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard, it is determined that a specific event corresponding to the similarity has occurred within the image.

[0063] Additionally, as an image detection system (1) that performs a method for detecting specific events in real time, it includes a camera module (1000) comprising an image capturing unit (1100) that captures an image and transmits it to an edge module (2000); and an edge module (2000) comprising one or more processors and one or more memories; wherein the edge module (2000) comprises: a vector generation unit (2100) that inputs each of a predetermined number of consecutive image frames among the image into a deep learning-based embedding model to generate a target embedding vector for each image frame; a similarity calculation unit (2200) that calculates the similarity between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame; and an event judgment unit (2300) that determines that a specific event corresponding to the similarity has occurred within the image when the ratio of image frames having a similarity exceeding a predetermined threshold among the consecutive image frames exceeds a predetermined ratio standard.

[0065] Specifically, the camera module (1000) in the present invention is implemented in hardware by including at least one image sensor or lens for capturing an image, and according to an embodiment, may further include at least one processor and memory for converting a signal obtained from the image sensor into image data and transmitting it to an edge module (2000).

[0066] Meanwhile, the camera module (1000) may be configured to include an image capturing unit (1100) that captures a surveillance area and generates real-time image data. According to an embodiment, the image capturing unit (1100) may be a physical component that acquires an image within the camera module (1000), or a software module or a form combined with hardware that performs the function of processing and transmitting the acquired image as digital data, and becomes the entity that performs the image capturing step (S100) illustrated in (b) of FIG. 2.

[0068] In addition, the edge module (2000) in the present invention is implemented in hardware by including at least one processor and memory, and logically implements each functional unit described later through program instructions loaded into said memory. As shown in FIG. 2 (a), the edge module (2000) may be configured to include a vector generation unit (2100), a similarity calculation unit (2200), and an event judgment unit (2300). Meanwhile, each component may be a software module that is loaded into the memory of the edge module (2000) and executed by a processor, or may be in a form combined with hardware, and becomes the entity that performs the vector generation step (S200), similarity calculation step (S300), and event judgment step (S400) respectively shown in FIG. 2 (b).

[0070] The video recording step (S100) is performed by the video recording unit (1100) of the camera module (1000). The video recording step (S100) captures the surveillance area to generate a real-time video stream and transmits it to the edge module (2000) via a wired / wireless network or an internal interface. As described above, since the algorithm of the present invention vectorizes and analyzes the visual features of the image, feature extraction for event detection can be performed stably even when the resolution of the image acquired by the video recording unit (1100) is low or contains noise.

[0072] The vector generation step (S200) is performed by the vector generation unit (2100) of the edge module (2000). The vector generation step (S200) inputs each of a predetermined number of consecutive image frames from the received image data into a deep learning-based embedding model. Through this, a 'target embedding vector' is generated in which the visual context and features within each image frame are quantified. This process consumes extremely little computing resources of the edge module (2000) because it adopts a method of immediately converting the features of the entire frame or the area where movement is detected into coordinates in a high-dimensional space, without heavy computational processes such as separating individual objects within the image or forming bounding boxes to track them.

[0074] The similarity calculation step (S300) is performed by the similarity calculation unit (2200). The similarity calculation step (S300) calculates the similarity between a reference embedding vector for a specific event, which is stored in a database (DB) or the like, and a target embedding vector for each image frame. In this case, as an example embodiment, the similarity calculation may utilize algorithms such as Cosine Similarity or Euclidean Distance.

[0076] The event judgment step (S400) is performed by the event judgment unit (2300). The event judgment step (S400) does not immediately trigger an alarm simply because the similarity value at a specific point in time is high, but applies a time-series verification logic. Specifically, it calculates the ratio of frames in which the calculated similarity value exceeds a threshold value among a preset number of consecutive image frames, and finally determines that the specific event has actually occurred only when this ratio exceeds a preset ratio standard.

[0078] This step-by-step execution structure can effectively filter out one-time noise that may exhibit characteristics similar to events, such as temporary shaking of the subject due to wind influence or changes in lighting. In other words, the present invention can simultaneously achieve computational efficiency and improved detection precision through the execution of the vector generation step (S200) to the event judgment step (S400).

[0080] FIG. 3 schematically illustrates integrated and separated implementation forms of a camera module (1000) and an edge module (2000) according to one embodiment of the present invention.

[0082] As illustrated in FIG. 3, a method for detecting a specific event in real time performed by an image detection system (1) comprises: an edge module (2000) including one or more processors and one or more memories, wherein the edge module (2000) receives an image captured by a camera module (1000); wherein the edge module (2000) inputs each of a predetermined number of consecutive image frames among the image into a deep learning-based embedding model to generate a target embedding vector for each image frame; wherein the edge module (2000) calculates a similarity between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame; The method includes an event judgment step (S400) in which, by the edge module (2000), if the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard, it is determined that a specific event corresponding to the similarity has occurred within the image.

[0084] Specifically, the image detection system (1) of the present invention can be flexibly implemented as a hardware integrated structure as shown in (a) of FIG. 3 or a hardware separated structure as shown in (b) depending on the installation environment and requirements of the site.

[0086] FIG. 3(a) illustrates a hardware-integrated (AI camera) structure in which a camera module (1000) and an edge module (2000) are integrated within a single physical housing. In this case, a computational processor (NPU, CPU, etc.) and a memory chipset constituting the edge module (2000) are directly mounted on the main board of the camera terminal. In this integrated structure, image data acquired through the lens is immediately transmitted to the edge module (2000) via an internal data bus for analysis, so there is almost no network delay due to data transmission.

[0087] In addition, when establishing a system in a new control area, an intelligent detection system can be built using only a single camera unit, without the need for separate space for external computing devices or complex wiring work, offering the advantage of excellent convenience in installation and maintenance.

[0089] Meanwhile, FIG. 3(b) illustrates a hardware separation structure in which the camera module (1000) and the edge module (2000) are physically separated and connected via a wired / wireless network or a local cable. In this case, the camera module (1000) may be composed of a general industrial CCTV, IP camera, etc., without a separate AI computing function, and the edge module (2000) may be implemented in the form of a small edge computer, a set-top box, or an edge-type NVR (Network Video Recorder) that accommodates multiple cameras, installed in a communication terminal box adjacent to the camera.

[0090] This separable structure offers high economic efficiency in terms of public infrastructure construction. Without the need to completely replace the aging, outdated camera infrastructure already installed in large quantities at existing sites, the entire system can be easily upgraded to a state-of-the-art intelligent control system simply by additionally connecting the edge module (2000) of the present invention. In other words, the system of the present invention provides strong interoperability that can maintain stable detection performance even in environments combined with low-resolution outdated cameras through the embedding-based analysis logic described later.

[0092] In particular, in the separated structure illustrated in Fig. 3(b), the edge module (2000) can directly receive and analyze video streams transmitted from the outside via a wired or wireless network, without being limited to the manufacturer or model of the camera module (1000). Through this independent structure, the edge module (2000) can function as an independent AI computing terminal capable of independently performing a video reception step of receiving video according to a preset standard protocol, free from hardware dependency of a specific video detection system (1).

[0093] Therefore, the present invention provides unrivaled interoperability and business flexibility by separately distributing only the edge module (2000) to a demand center where camera infrastructure has already been established, thereby enabling the immediate conversion of the existing general control environment into an intelligent automatic control environment.

[0095] Consequently, the specific event real-time detection method of the present invention can achieve the technical effect of minimizing the computational load of the central control server by performing independent embedding operations at the edge and time-series ratio-based judgment logic identically regardless of whether it is implemented in the integrated or separated hardware architecture.

[0097] FIG. 4 schematically illustrates the process of deriving a reference embedding vector for each specific event according to one embodiment of the present invention.

[0099] As illustrated in FIG. 4, the reference embedding vector is characterized by being derived as the center point of the learning embedding vectors derived from a plurality of learning image frames for each specific event included in the learning data for the embedding model.

[0101] The image detection system (1) according to the present invention performs real-time image analysis by undergoing a pre-learning process that defines the visual characteristics of individual events to be detected as numerical representative values, or by using an artificial neural network-based embedding model learned through the process.

[0103] Specifically, training data consisting of a large number of pre-labeled video or image frames is prepared for each specific individual event targeted for detection, such as falling, assault, and fire. For example, in the case of an 'assault' event, multiple training data including actions of pushing or hitting each other are utilized, while in the case of a 'falling' event, training data featuring a person lying on the floor from various angles and postures is used.

[0104] The prepared training data for specific events is input into a deep learning-based embedding model. The embedding model extracts visual feature information, such as the appearance, pose, and relationship with the background of an object, from the input image frames and converts it into an embedding vector, which is a specific coordinate in a high-dimensional feature space or latent space.

[0106] At this time, it is preferable that the embedding model of the present invention be trained to apply a metric learning technique in which data with the same label (event) are grouped close to each other in the vector space, and data with different labels are moved far apart. As shown at the bottom of Fig. 4, through this process, embedding vectors form a certain cluster for each specific event.

[0107] The edge module (2000) or a separate learning device calculates the center point or statistical representative value of the cluster of embedding vectors formed for each specific event and derives it as the reference embedding vector for the corresponding event. For example, a single representative vector having the average directionality of numerous embedding vectors labeled 'assault' is defined as the reference embedding vector for 'assault'.

[0109] The reference embedding vectors for specific events derived in this way are stored in the memory within the edge module (2000) or in a linked database (DB). Subsequently, during real-time monitoring, the edge module (2000) can immediately determine what kind of dangerous situation has occurred without heavy object classification operations by comparing the similarity between the target embedding vector extracted from the field video and the reference embedding vector for each event stored in the DB in parallel.

[0110] As a result, the reference embedding vector constructed through the process of Fig. 4 functions as a key judgment indicator that enables the present invention to quickly and accurately detect various multiple events even in low-spec edge environments.

[0112] However, the image background (e.g., urban street) or optical conditions (e.g., clear daytime) of the training data used in the process of training the above embedding model may differ from the specific background (e.g., snow-covered Antarctic base, complex indoor factory, etc.) and the constantly changing illumination environment of the actual site where the edge module (2000) of the present invention is installed and operated. In order to overcome the basis similarity error with the reference embedding vector caused by such environmental discrepancies between the training domain and the actual analysis domain, the present invention includes an adaptive threshold calculation logic that is autonomously adjusted to the installation site, as shown in FIGS. 6 to 8 to be described later.

[0114] FIG. 5 schematically illustrates a method for calculating the similarity between a target embedding vector and a reference embedding vector according to an embodiment of the present invention.

[0116] As shown in Fig. 5, the similarity is calculated through cosine similarity.

[0118] Specifically, the similarity calculation unit (2200) of the edge module (2000) obtains a target embedding vector (V) that compresses the features of a real-time image frame from the vector generation unit (2100). target ) is received. In parallel, a predefined reference embedding vector (V) for each event, which has been pre-learned and stored from the database (DB) through the process described in FIG. 4, is received. ref Load )

[0119] The core similarity calculation method adopted as an embodiment of the present invention may include a cosine similarity algorithm that utilizes the angle (θ) between two vectors located in a multidimensional space. Preferably, in FIG. 5, two vectors are depicted on a two-dimensional plane (X-axis and Y-axis) for convenience of explanation, but the embedding vectors output by the actual embedding model exist in a high-dimensional latent space of hundreds to thousands of dimensions.

[0120] Cosine similarity is not significantly affected by the magnitudes (scalar values) of the two vectors or changes in the physical size of the subjects within the image; instead, it determines the similarity of situations based solely on the directionality pointed to by the vectors, that is, the angle between the two vectors.

[0122] Specifically, the target embedding vector V extracted in real time target ) and the reference embedding vector V of a specific event refThe smaller the angle θ between them (i.e., the closer the value of cosθ is to 1), the more geometrically similar the global context of the image currently being captured by the camera module (1000) is to a dangerous situation (e.g., falling, assault, etc.) stored in the database (DB). Conversely, the closer the angle θ is to 90 degrees or greater (i.e., the closer the value of cosθ is to 0 or negative), the more the current image frame is to an everyday normal situation completely unrelated to the event.

[0124] In particular, the cosine similarity operation between these high-dimensional vectors is performed solely through simple vector dot product and normalization processes. This offers the advantage of consuming extremely low computing resources compared to complex object-unit post-processing operations, such as bounding box regression or Non-Maximum Suppression (NMS), which are adopted by existing intelligent CCTVs.

[0125] Consequently, the present invention achieves a significant technical effect by applying a lightweight mathematical measure of directional comparison in vector space to an edge computing environment instead of analyzing pixels of an image, thereby enabling real-time event similarity determination at tens of frames per second without communication delay, even within a low-spec edge module (2000) equipped with a low-power CPU or a small NPU.

[0127] FIG. 6 schematically illustrates a concept in which a threshold value is applied variably according to changes in the shooting environment according to an embodiment of the present invention.

[0129] As illustrated in FIG. 6, the threshold value is characterized by being adaptively set to correspond to the environment based on a similarity distribution that appears differently depending on the shooting environment of the camera module (1000).

[0131] Conventional video surveillance systems use a fixed single threshold value, which limits their ability to flexibly reflect spatial characteristics (background, field of view, etc.) of the location where the camera is installed and temporal characteristics (illumination, weather noise, etc.) that change from moment to moment. To overcome this limitation, the present invention dynamically calculates an optimal adaptive threshold value by statistically analyzing the unique shooting environment and temporal changes of each individual camera module (1000).

[0133] Specifically, since the deep learning embedding model analyzes global features of the entire screen, if the background of the learned data (e.g., an urban area) and the unique background of the actual site where the camera is installed (e.g., an empty snowfield in Antarctica) are different, the starting point of the similarity value derived under normal conditions changes solely due to visual differences in background pixels. In addition, even for the same location, the proportion of objects occupied by objects varies depending on whether the camera's field of view (FOV) is wide (a) or narrow (b), as shown in FIG. 6, and the volatile amplitude of the similarity value fluctuates drastically depending on whether it is a clear daytime environment or a noisy nighttime environment, as shown in FIG. 7.

[0135] Since the proportion of pixels occupied by the subject in the image varies depending on the lens characteristics or installation distance of the camera module (1000), an adaptive threshold value must be applied differently accordingly. As illustrated in FIG. 6 (a), in a camera shooting environment with a wide field of view (Wide FOV), when an event such as an assault occurs from a distance, the proportion of the corresponding act in the overall image context appears relatively small. In this case, since the similarity peak of the target embedding vector may be formed somewhat low when the event occurs, it is desirable to apply an adaptive threshold value that is relatively lowered in order not to miss minute feature changes.

[0136] On the other hand, in a camera environment with a narrow field of view (Narrow FOV) as shown in Fig. 6 (b) (e.g., when a zoom lens is applied or installed in close proximity to the subject), the object's behavioral pattern occupies most of the screen. In this environment, since the similarity value is calculated to be very clear and large when an event occurs, the adaptive threshold is raised to increase the reliability of detection so that large movements of ordinary passersby are not mistaken for dangerous events.

[0138] FIG. 7 schematically illustrates the process of calculating a threshold value according to the shooting environment based on a normal similarity distribution according to an embodiment of the present invention.

[0140] As illustrated in FIG. 7, the threshold value is characterized by being adaptively set to correspond to the environment based on a similarity distribution that appears differently depending on the shooting environment of the camera module (1000).

[0141] In addition, the above threshold is calculated by summing the average value of the similarity distribution derived based on multiple similarity data obtained in a normal period where no event occurs, and the value obtained by multiplying the standard deviation of the similarity distribution by a preset constant.

[0143] Specifically, FIG. 7 illustrates the process of calculating an adaptive threshold value according to changes in illumination environments, such as time of day or weather, as well as spatial (angle of view) installation environments. Preferably, in other embodiments, the threshold value may vary according to different camera environments in addition to the aforementioned environments. In the graph of FIG. 7, the horizontal axis (X-axis) represents time (t), and the vertical axis (Y-axis) represents the similarity value (0.0 to 1.0) between the reference embedding vector and the target embedding vector. As an embodiment of the present invention, the edge module (2000) can continuously extract the target embedding vector during normal periods when no event occurs, and learn the average similarity and the degree of similarity fluctuation (variance / standard deviation) in the environment in the background.

[0144] It can be seen that a baseline of normal average similarity, which vibrates at a constant rate, exists at the bottom of the graph shown in Fig. 7. This represents the average similarity value between the target embedding vector and the reference embedding vector acquired in the background during normal monitoring situations where no specific risk event occurs. As environmental noise, such as the viewing angle or illumination of the camera module (1000), changes, the range of fluctuation in the similarity value—that is, the distribution of data (variance and standard deviation)—changes around this baseline of normal average similarity.

[0146] Consequently, the difference between the adaptive threshold (day) and the adaptive threshold (night) shown in Fig. 7 is not merely a margin arbitrarily designated by the administrator, but is the result of thoroughly reflecting the statistical characteristics of the normal similarity distribution collected in real-time in the environment.

[0147] Meanwhile, through the process described in FIGS. 6 and 7, fixed background characteristics of the space are corrected by an average value regardless of whether it is the middle of Seoul or an Antarctic base, or whether the field of view is wide or narrow, and dynamic environmental changes such as day and night changing or wind and rain are corrected by a margin based on the standard deviation, thereby providing stable detection performance in any adverse installation environment.

[0149] Hereinafter, with reference to FIG. 8, it will be specifically explained how the edge module (2000) of the present invention calculates an optimal adaptive threshold value using these normal average similarity and standard deviations.

[0151] FIG. 8 schematically illustrates a threshold value determined according to the mean value and standard deviation of a similarity distribution according to one embodiment of the present invention.

[0153] As shown in FIG. 8, the threshold value is calculated by summing the average value of a similarity distribution derived from multiple similarity data obtained in a normal period in which no event occurs, and the value obtained by multiplying the standard deviation of the similarity distribution by a preset constant.

[0155] The above threshold (T adapt ) is calculated by summing the average value (μ) of the similarity distribution derived from multiple similarity data obtained in a normal period where no event occurs, the standard deviation (σ) of the similarity distribution, and a preset weighting constant (K). This can be expressed as a formula as [Equation 1] below.

[0156] [Mathematical Formula 1]

[0157]

[0159] Specifically, the table in Fig. 8 exemplarily shows data distribution curves (normal distribution type) formed in four cases depending on whether the mean value (μ) and standard deviation (σ) of the normal similarity distribution are high and low, respectively.

[0161] First, the average value (μ) can represent the 'background similarity' inherent in the camera environment. For example, in the case of a gym where many people run around even during normal times, the average similarity (μ) for 'assault' events may be high. In this case, the basic threshold baseline must be set high so that everyday running is not mistaken for assault.

[0162] Secondly, the standard deviation (σ) can represent the variability caused by noise in the corresponding camera environment. As shown in the bottom row of the table in Figure 8, in environments with a high standard deviation (σ) (e.g., nighttime environments with low illumination and severe pixel noise, or outdoor environments with wind and rain), the similarity curve takes on a shape that spreads out widely and smoothly. Preferably, a widely spread distribution implies a high probability of outliers occurring that deviate significantly from the average range.

[0164] Accordingly, in accordance with the process described above in FIGS. 6 to 8, the edge module (2000) derives a final threshold value by adding a larger margin (K * standard deviation) value to the average value according to [Equation 1], as the environment is one in which the standard deviation is calculated to be larger. This is a logical mechanism to preemptively raise the threshold barrier so that even if the similarity of the target embedding vector accidentally spikes due to temporary changes in illumination or adverse weather conditions, it is not mistaken for an actual dangerous event. Conversely, in an environment where the lighting is bright and noise is low and the standard deviation is measured to be low, the similarity curve is formed narrowly and sharply, so the threshold is automatically adjusted tightly to minimize the added margin so that it can respond immediately and sensitively to the occurrence of minute events.

[0166] Meanwhile, the preset weight constant (K) is a parameter that can be tuned to various values ​​such as 1.5, 2.0, and 3.0 depending on the sensitivity of the field requirements or the purpose of control. The above weight constant (K) plays a role in optimizing the trade-off balance between the false positive rate and the false negative rate of the system based on a statistical confidence interval.

[0167] As a result, the present invention does not require additional heavy deep learning computations by introducing an adaptive threshold algorithm based on a lightweight and intuitive statistical formula as in [Equation 1] above. Through this, the edge module (2000) can implement highly intelligent unmanned control performance by actively responding to spatial (background, viewing angle) and temporal (illumination, weather) environmental factors of the installation site that change from moment to moment, and automatically readjusting the optimal judgment zero point.

[0169] FIG. 9 schematically illustrates a threshold table that is independently set for a plurality of specific events according to one embodiment of the present invention.

[0171] As illustrated in FIG. 9, the threshold value is characterized by being set differently from each other by independently reflecting the similarity distribution that appears differently for each individual event when there are multiple predefined events.

[0173] Generally, when an intelligent control system is introduced into actual public infrastructure or industrial sites, there is a requirement to simultaneously detect and monitor multiple complex risk situations rather than a single threat factor. Referring to the table shown in Fig. 9, the system of the present invention is configured to simultaneously detect various actions (events) such as 'falling,' 'assault,' 'fire,' 'cigarette (smoking),' 'failure to wear a helmet,' and 'illegal parking,' depending on the monitoring purpose.

[0174] One of the technical features of the present invention is that, when detecting multiple events, a single uniform similarity criterion is not applied to the entire system, but rather different threshold values ​​are independently calculated and applied for each activity (event). This is because the normal similarity distribution appearing in the embedding space varies significantly depending on the unique visual and physical characteristics of individual events.

[0176] For example, in the case of a 'fire' event, the dynamic changes on the screen are significant due to factors such as the flickering of flames or the irregular diffusion of smoke; consequently, the variance of similarity with the fire reference vector can appear wide and unstable even in normal environments. On the other hand, since the 'falling' event has strong static characteristics where the subject lies still on the ground, the range of variation (variance) in similarity is relatively narrow when compared to normal pedestrians. If events with such completely different characteristics are grouped and judged using a single fixed threshold, it leads to fatal errors, resulting in frequent false positives in fire detection and missed detections in fall detection.

[0177] Accordingly, when the edge module (2000) of the present invention collects target embedding vectors in the background, it completely separates and calculates 'normal statistics with reference vectors for assault', 'normal statistics with reference vectors for falling', etc. By substituting the average values ​​and standard deviations for each individual event, derived independently in this way, into the [Mathematical Formula 1] above, multiple customized defense walls (adaptive threshold values ​​A to F) perfectly optimized for the characteristics of each event are constructed as shown in FIG. 9.

[0179] Consequently, the present invention fundamentally prevents judgment interference and false detection phenomena between events that may occur during multi-event detection. Furthermore, even if new detection targets (e.g., worker collapse, specific equipment malfunction, etc.) are added according to the needs of the client, only a new independent threshold line needs to be added without affecting the threshold of the existing detection logic at all, thereby providing excellent system scalability that enables highly customized comprehensive control using only a single edge device.

[0181] FIG. 10 schematically illustrates an embedding model learned for high robustness in different camera environments according to one embodiment of the present invention.

[0183] As illustrated in FIG. 10, the embedding model is characterized as being an artificial neural network model trained to output an embedding vector value having a vector similarity greater than or equal to a mutually preset standard when an event labeled in an image frame is the same specific event, even if an image frame included in a video captured in a different camera environment is input.

[0185] As previously explained in FIGS. 6 to 8, the present invention can primarily resolve the problem of false positives caused by changes in the spatiotemporal environment through an algorithmic barrier called an adaptive threshold. In addition, it is desirable that the embedding model of the present invention, which is a feature extractor, fundamentally possesses high generalization performance that is not affected by various viewing angles and background changes.

[0187] Specifically, the two image frames shown at the top and bottom left of FIG. 10 represent two sites with completely different spatial and optical characteristics. The top image depicts a narrow and monotonous indoor corridor environment, whereas the bottom image exemplifies an industrial site (factory) environment where complex equipment is arranged. Furthermore, the two cameras differ in their installation height and field of view, and the size, clothing, and angle of fall of the subjects (people) captured in the images are also all different.

[0188] A general object detection-based AI control model has a fatal limitation in that if the background, field of view, and scale of the object differ drastically from the training data, it fails to extract features, resulting in missing detections or mistaking complex background textures for objects. However, the embedding model mounted on or connected to the edge module (2000) of the present invention is designed to exclude coordinate regression operations that rely on local contours of pixels within the image, and to compress global contexts such as the subject's 'overall posture and shape while lying down' and 'interaction with the background' as a whole.

[0190] In particular, the embedding model of the present invention can be trained by the manager of the image detection system (1) of the present invention through metric learning techniques such as contrastive learning, so that images containing semantically identical actions (e.g., falling) are clustered adjacently in a high-dimensional latent space even if the visual backgrounds at the pixel level are vastly different (preferably, the transformed vector values ​​of each of the images form a vector similarity greater than or equal to a preset standard).

[0191] As a result, as shown in Fig. 10, even if two image frames with completely different environmental factors pass through the same embedding model, the target embedding vectors that are finally output have a vector similarity greater than a mutually preset standard, and accordingly, both images can be labeled and identified as having the same 'falling' event.

[0193] The inherent strong domain invariance and noise robustness of this invention demonstrate overwhelming practical strengths in public and industrial infrastructure fields. This is because it guarantees a solution that allows for the immediate installation and operation of the system without the need for expensive model retraining tailored to site characteristics or massive additional data collection, regardless of whether a client organization introduces the system into any new, heterogeneous environment such as schools, hospitals, corridors, underpasses, or factories.

[0195] FIG. 11 schematically illustrates the process of extracting a change region according to one embodiment of the present invention. FIG. 12 schematically illustrates the process of deriving a local target embedding vector using the change region according to one embodiment of the present invention.

[0197] As illustrated in FIGS. 11 and 12, the vector generation step (S200) includes: a change area extraction step that extracts a change area where movement occurs based on image difference operations between adjacent image frames among the consecutive image frames, or a pre-set area-specific motion vector for the consecutive image frames; and a local vector derivation step that derives the target embedding vector by inputting only the image of the change area into the embedding model.

[0199] Specifically, according to an embodiment, the edge module (2000) of the present invention may be equipped with logic that performs embedding operations by selecting only the region of interest where meaningful movement occurred among the entire image frame (Full-frame) in order to maximize the efficiency of image analysis under limited hardware resources.

[0200] At this time, in order to filter meaningless environmental noise such as tree branches swaying in the wind or temporary lighting changes, the edge module (2000) preferably determines that the area of ​​the calculated image difference or the magnitude of the motion vector exceeds a preset motion threshold and determines that area as a meaningful macroscopic movement and confirms the area as a change area (A).

[0202] As illustrated in FIG. 11, the vector generation unit (2100) performs a change area extraction step of comparing image frames received in succession, such as a first frame (F#1) and a second frame (F#2), to extract a 'change area (A)' where movement occurs.

[0203] At this time, to identify the change area (A), a simple image difference operation between adjacent image frames can be performed, or the direction and magnitude of macroscopic movement can be determined based on motion vector data that has been calculated for each pre-set area (macroblock, etc.) during the video compression / encoding process. That is, high-load operations for deriving precise pixel-unit contours of an object are excluded, and only the local area (A) where clustered shape changes occur within the entire screen area (B) is screened very quickly and lightly.

[0204] Meanwhile, the above-mentioned image difference operation refers to an ultra-lightweight preprocessing process performed prior to the feature extraction operation through the embedding model. Specifically, this can be performed using a simple matrix operation method that mutually subtracts the brightness or color values ​​of each corresponding pixel constituting two adjacent image frames (e.g., F#1 and F#2).

[0205] This image difference operation does not extract fine outlines of objects or analyze complex correlations between pixels, but simply serves to identify the existence of 'clustered areas' where changes in pixel values ​​have occurred in the current screen compared to the previous screen.

[0207] Next, as illustrated in FIG. 12, when a change region (A) is identified through the change region extraction step, the vector generation unit (2100) separates and extracts (crops) only the image corresponding to the change region (A) from the entire original image frame (B). Subsequently, a local vector derivation step is performed in which only the extracted image of the change region (A) is input into a deep learning-based embedding model to derive a target embedding vector for the corresponding region.

[0209] This local region-based change region extraction and embedding technique provides the following overwhelming practical advantages when establishing the image detection system (1) of the present invention in public and industrial sites.

[0210] First, a dramatic reduction in computational resources is achieved. Since the size of the resolution tensor of the input image that the deep learning model must process is significantly reduced, comfortable real-time AI inference at tens of frames per second (FPS) can be easily performed even on low-power CPUs without GPUs or inexpensive edge modules (2000) based on lightweight NPUs. This can be a driving force for maximizing performance while significantly lowering the unit cost.

[0211] Secondly, it improves precision and blocks the source of background noise. If the original image with a wide field of view is fed directly into the embedding model, there is a risk that extensive background information, such as unnecessary walls, roads, and stationary objects, will be mixed into the numerical values ​​of the target embedding vector, thereby hindering the calculation of similarity. However, if only the change region (A) is extracted and analyzed as in the present invention, interference from background noise is fundamentally blocked, and only pixels representing dynamic actions such as a person's 'assault' or 'falling down' are vectorized in a highly concentrated state. Consequently, a remarkable effect is achieved in which the accuracy of similarity matching with the reference embedding vector increases dramatically.

[0213] FIG. 13 schematically illustrates conditions under which embedding operations and similarity calculations are omitted for an image frame according to an embodiment of the present invention. In addition, FIG. 14 schematically illustrates the effect of reducing computational resources through the extraction of change regions and the omission of operations according to an embodiment of the present invention.

[0215] As illustrated in FIGS. 13 and 14, the vector generation step (S200) omits the derivation of the target embedding vector and the calculation of the similarity for image frames in which the magnitude of the movement is less than a preset change standard.

[0217] As previously described in FIGS. 11 and 12, the edge module (2000) of the present invention calculates motion vectors between consecutive frames and extracts a change area where movement of the subject occurs. Meanwhile, the edge module (2000) according to the present invention fundamentally blocks subsequent deep learning calculations for static frames in which no macroscopic movement change is detected as a result of the judgment in the preceding change area extraction step.

[0219] Specifically, as illustrated in FIG. 13, a fifth image frame (Frame #5) and a sixth image frame (Frame #6) are shown, which are captured sequentially in time at the same angle of view where a parked vehicle exists. As a result of the edge module (2000) performing a difference operation or motion vector between the two image frames, the amount of change in motion may fall short of the previously mentioned motion threshold, so a significant change area may not be extracted. In this way, if it is confirmed that there is no dynamic intervention of a person or object within the screen in a static state, the edge module (2000) skips the inference process of passing the image data of the sixth image frame (Frame #6) through a deep learning embedding model.

[0221] That is, as illustrated by the bottom arrow flow of Fig. 13, the 'omission of deriving target embedding vectors' through the embedding model is performed. Furthermore, since the target embedding vector to be analyzed has not been generated at the current time, there is no need to compare it with the reference embedding vector of the database (DB), so the subsequent logic of 'omission of similarity calculation' is also dynamically linked and performed.

[0222] The groundbreaking hardware advantages that this 'selective driving based on the presence or absence of a change area' algorithm brings to edge computing environments are clearly contrasted in Fig. 14.

[0224] The conventional method of unconditionally inputting the entire area into the embedding model for every image frame without considering whether the subject is moving, as shown in Fig. 14 (a), results in an extreme waste of computational resources of the NPU or CPU (increase in computational resources) due to analyzing meaningless backgrounds.

[0225] On the other hand, if a significant change area is detected according to an embodiment of the present invention as in Fig. 14 (b), only that area (hatched part) is input into the embedding model, or if the embedding operation itself is omitted when there is no movement at all as in Fig. 13, the consumption of computational resources of the edge module (2000) is dramatically reduced.

[0227] As a result, the present invention blocks unnecessary AI inference on meaningless static frames, thereby minimizing power consumption and heat generation of the edge module (2000), and can achieve optimal economy and efficiency by comfortably parallel monitoring of multiple CCTV channels with just a single small edge module (2000).

[0229] FIG. 15 schematically illustrates the process of deriving a reference difference average value according to one embodiment of the present invention. FIG. 16 schematically illustrates the process of deriving a target difference average value according to one embodiment of the present invention. FIG. 17 schematically illustrates the process of finally determining whether an event occurs by additionally comparing pattern similarity according to one embodiment of the present invention.

[0231] As illustrated in FIGS. 15 to 17, the reference embedding vector is stored along with a reference difference average value calculated by determining the difference between multiple learning embedding vectors derived from multiple learning image frames for specific events included in the learning data for the embedding model, and the similarity calculation step (S300) further includes: an embedding difference calculation step that calculates a delta embedding difference, which is the difference between adjacent image frames among the consecutive image frames; and a pattern analysis step that calculates a target difference average value, which is the average value of multiple delta embedding differences calculated from the consecutive image frames, and calculates pattern similarity by comparing the target difference average value with the reference difference average value; and the event judgment step (S400) further determines whether the pattern similarity exceeds a preset pattern threshold value to finally determine whether the event occurs.

[0233] Conventional video surveillance systems analyze only morphological features on a single frame (still image). However, threat events such as 'assault' or 'falling down' are difficult to accurately distinguish based solely on posture at a specific point in time. For example, an emergency patient who suddenly collapses while walking down the street and a worker who has been lying on the ground resting from the beginning may have geometrically very similar embedding vectors when looking only at the video at a specific point in time.

[0234] In order to minimize false positives caused by the limitations of such static analysis, the present invention verifies the time-series behavioral trajectory by utilizing the difference vector (ΔV), which is the amount of vector change between consecutive frames.

[0236] As illustrated in FIG. 15, during the pre-training process, training image frames (F#1 to F#5) included in sequential training data labeled with a specific event (e.g., falling) are sequentially input into an embedding model and converted into respective reference training embedding vectors (V1 to V5). At this time, the system calculates the difference between vectors of adjacent frames (e.g., V2-V1, V3-V2, etc.) over time to produce a series of difference vectors (ΔV1 to ΔV4). The average value of these difference vectors is stored in a database as a reference difference average value that compresses the flow of the unique speed and shape change of the subject when the corresponding event occurs.

[0238] Subsequently, as illustrated in FIG. 16, which describes a real-time monitoring process, the similarity calculation unit (2200) further includes an embedding difference variance calculation unit that performs an embedding difference variance calculation step and a pattern analysis unit that performs a pattern analysis step, and sequentially extracts target embedding vectors (V'1 to V'5) for a preset number of real-time continuous image frames (F#1' to F#5') among the images received from the image capturing unit (1100) (depending on the embodiment, after undergoing the lightweighting process of FIG. 11 and FIG. 12 above). Then, by calculating the vector change amount between them equally, a target difference average value representing the real-time dynamic behavior trajectory of the current subject is derived.

[0240] Meanwhile, as described above in FIG. 2, the edge module (2000) selects a preset number (e.g., 10 frames) of consecutive image frames from a video received in real time and derives target embedding vectors from each. Subsequently, the target embedding vectors are individually compared to see how much they satisfy a similarity condition with a specific reference embedding vector in a database. As a result of the comparison, the first condition is determined as to whether the ratio of image frames satisfying the similarity condition among the preset number exceeds a preset ratio standard (e.g., 70%).

[0241] If the ratio satisfying the similarity condition is 70% or less, it is considered a normal situation where the subject briefly stumbled or momentary noise occurred, so the alarm is dismissed and the operation is terminated immediately.

[0242] In addition, if the ratio satisfying the similarity condition exceeds the ratio standard of 70%, it is determined that a specific event corresponding to the reference embedding vector being compared has occurred.

[0244] At this time, as illustrated in the judgment flowchart of FIG. 17 according to an embodiment, the event judgment unit (2300) of the present invention further performs the embedding difference variance calculation step and the pattern analysis step to generate a final alarm through a very strict 2-step cross-verification logic as follows.

[0245] Only when the first condition regarding the similarity between the reference embedding vector and the target embedding vector is satisfied (Yes), that is, when it is first determined that the object in the image frame is reliably maintaining a lying posture above a certain percentage, the edge module (2000) can enter into a judgment through a pre-set second condition. Here, the target difference average value (real-time behavior trajectory), which represents the vector change amount between the previously collected image frames, is filtered again to see if it satisfies the condition regarding pattern similarity corresponding to the similarity with the reference difference average value (normal falling trajectory) stored in the database.

[0247] Meanwhile, the similarity used for determining the first condition refers to the degree of static spatial and morphological correspondence between the target embedding vector extracted from a single image frame and the reference embedding vector, and the preset threshold refers to the first reference value that determines whether such static poses are similar.

[0248] On the other hand, the pattern similarity used for determining the second condition is a comparison measure between the target difference average value and the reference difference average value accumulated over time, and refers to a value calculated for the degree of agreement of dynamic transitions, such as the time-series behavioral trajectory, speed, and directionality of the subject, rather than the posture at a specific point in time. In addition, the preset pattern threshold refers to a second reference value that is set independently of the preset threshold to determine whether the pattern similarity passes.

[0249] In this way, the present invention provides a technical advantage in that it allows a manager to finely tune the sensitivity of static verification and the sensitivity of dynamic verification individually according to field conditions by separating and operating the similarity / threshold value for evaluating spatial features and the pattern similarity / threshold value for evaluating time-series dynamic trajectories as mutually independent parameters.

[0251] That is, according to the embodiment, the present invention can determine the situation as an actual dangerous event only when both of the above two conditions are satisfied. Even if it is a lying position (satisfying the first condition), if the previous trajectory of behavior is a normal resting pattern of gradually lying down and thus does not satisfy the second condition, this is treated as a normal, problem-free situation and suppresses the occurrence of unnecessary alarms.

[0252] Consequently, the present invention provides a significant technical effect of achieving a high level of situational awareness reliability while minimizing computational load by removing single-frame noise through majority ratio verification within a certain interval and tracking the causal relationship (trajectory) of behavior through differential mean value verification.

[0254] FIG. 18 schematically illustrates a predetermined number of consecutive image frames acquired during a predetermined time interval according to one embodiment of the present invention.

[0256] As illustrated in FIG. 18, the preset number is determined by the number of image frames acquired during a preset time interval in the image.

[0258] As previously described in the first condition of FIG. 17, the event judgment unit (2300) of the present invention calculates a ratio satisfying a similarity condition within a preset number of frame groups to filter out temporary noise in a single image frame. At this time, according to the embodiment, the preset number of image frames from which the target embedding vector is extracted is not limited to a simple constant (e.g., 10 frames), but can be dynamically defined based on a preset time series T interval, which is a physical time standard.

[0259] The edge module (2000) sets a time series T interval from an arbitrary starting point to an ending point, having a preset time interval (e.g., 3 seconds, 5 seconds, etc.) from the real-time video. Then, the edge module (2000) derives target embedding vectors by using the entire number of image frames captured and received during the preset T interval as parameters.

[0260] That is, the event judgment unit (2300) calculates the ratio of the number of target embedding vectors satisfying the similarity condition with the reference embedding vector to the total number of image frames included within the preset time interval, and determines whether the event occurs by determining whether this ratio exceeds a preset ratio standard.

[0262] This method of frame grouping and ratio calculation based on a preset time interval provides significant technical advantages in a control environment where heterogeneous camera modules (1000) with various specifications are mixed. For example, in an environment where a 30 FPS camera and a 15 FPS camera are installed, if only the number of frames is used as a fixed standard, even with the same 10 frames, the actual observed physical time (0.3 seconds vs. 0.6 seconds) may differ, causing a serious deviation in detection sensitivity.

[0263] However, if the frame parameter is set based on a preset time interval (e.g., the last 2 seconds) as in the present invention, even if the frame reception rate temporarily fluctuates due to differences in the camera's frames per second (FPS) setting or network conditions, the ratio standard can be evaluated using the same standard without shaking based on the absolute physical flow of time.

[0264] Consequently, the present invention supports managers in flexibly setting time intervals and ratio standards according to on-site conditions, thereby ensuring uniform and highly reliable intelligent control performance at all times, regardless of the camera's hardware specifications or communication environment.

[0266] FIG. 19 schematically illustrates a principle for reducing the computational load of an edge module (2000) according to one embodiment of the present invention.

[0268] As illustrated in FIG. 19, the detection method is characterized by minimizing the computational load of the edge module by performing only similarity comparison operations between embedding vectors without object detection operations that classify objects within the image and track bounding boxes.

[0270] As previously described, the intelligent image detection system (1) according to the present invention is characterized by minimizing complex operations at the pixel level within an image in order to be optimized for an edge computing environment having limited hardware resources.

[0271] Specifically, conventional video surveillance methods require heavy object detection operations that involve individually classifying the types of objects (people, things, etc.) within the video and continuously drawing bounding boxes along their movement coordinates, thereby causing a severe load on the terminal.

[0272] On the other hand, the edge module (2000) of the present invention does not perform such heavy object classification and bounding box coordinate tracking operations at all. Instead, as described above in FIG. 5, it immediately determines whether the current situation is dangerous by only performing a simple cosine similarity comparison operation based on the angle between the high-dimensional target embedding vector extracted from the image frame and the reference embedding vector.

[0273] Consequently, the present invention provides significant economic efficiency and practical efficiency, which can process multi-channel CCTV video in parallel in real time without frame drops or delays, even in low-spec environments without a GPU, while minimizing the computational load and heat generation of the edge module (2000) to an extreme degree.

[0275] FIG. 20 illustrates the internal configuration of a computing device (11000) according to one embodiment of the present invention.

[0277] The edge module (2000) mentioned in the description of FIG. 1 may include components of the computing device (11000) illustrated in FIG. 20, which will be described later.

[0279] As illustrated in FIG. 20, the computing device (11000) may include at least one processor (11100), memory (11200), peripheral interface (11300), input / output subsystem (I / O subsystem) (11400), power circuit (11500), and communication circuit (11600).

[0281] Specifically, the memory (11200) may include, for example, high-speed random access memory, magnetic disk, SRAM, DRAM, ROM, flash memory, or non-volatile memory. The memory (11200) may include software modules, instruction sets, or various other data required for the operation of the computing device (11000).

[0282] At this time, access to the memory (11200) from other components, such as the processor (11100) or the peripheral device interface (11300), can be controlled by the processor (11100). The processor (11100) may be composed of a single or multiple units and may include processors in the form of GPUs and TPUs to improve computational processing speed.

[0283] The above peripheral device interface (11300) can connect input and / or output peripheral devices of the computing device (11000) to the processor (11100) and the memory (11200). The processor (11100) can perform various functions for the computing device (11000) and process data by executing a software module or instruction set stored in the memory (11200).

[0284] The input / output subsystem (11400) may connect various input / output peripheral devices to the peripheral device interface (11300). For example, the input / output subsystem (11400) may include a controller for connecting peripheral devices such as a monitor, keyboard, mouse, printer, or, if necessary, a touchscreen or sensor to the peripheral device interface (11300). According to another aspect, the input / output peripheral devices may be connected to the peripheral device interface (11300) without passing through the input / output subsystem (11400).

[0285] The power circuit (11500) may supply power to all or part of the components of the terminal. For example, the power circuit (11500) may include one or more power sources such as a power management system, a battery or alternating current (AC), a charging system, a power failure detection circuit, a power converter or inverter, a power status indicator, or any other components for power generation, management, and distribution.

[0286] The communication circuit (11600) may enable communication with another computing device using at least one external port. Alternatively, as described above, the communication circuit (11600) may enable communication with another computing device by including an RF circuit and transmitting and receiving an RF signal, also known as an electromagnetic signal, as needed.

[0288] The embodiment of FIG. 20 is merely an example of the computing device (11000), and the computing device (11000) may have some components shown in FIG. 20 omitted, additional components not shown in FIG. 20 added, or a configuration or arrangement that combines two or more components. For example, a computing device for a communication terminal in a mobile environment may include a touchscreen or sensors in addition to the components shown in FIG. 20, and the communication circuit (1160) may include a circuit for RF communication of various communication methods (Wi-Fi, 3G, LTE, 5G, 6G, Bluetooth, NFC, Zigbee, etc.). The components that can be included in the computing device (11000) may be implemented as hardware, software, or a combination of both hardware and software, including one or more integrated circuits specialized for signal processing or applications.

[0289] Methods according to embodiments of the present invention may be implemented in the form of program instructions that can be executed through various computing devices and recorded on a computer-readable medium. In particular, the program according to the present embodiment may be configured as a PC-based program or an application dedicated to a mobile terminal. An application to which the present invention is applied may be installed on a user terminal through a file provided by a file distribution system. For example, the file distribution system may include a file transmission unit (not shown) that transmits the file upon a request from the user terminal.

[0291] 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 multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. In addition, other processing configurations, such as parallel processors, are also possible.

[0292] 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 command the processing unit independently or collectively. Software and / or data may be permanently or temporarily embodied in any type of machine, component, physical device, virtual equipment, computer storage medium or device, or transmitted signal wave 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 computing devices and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0293] 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. The hardware devices described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.

[0295] In one embodiment of the present invention, heavy object detection operations that classify objects in an image and track bounding boxes are excluded, and events are detected using only similarity comparison operations between embedding vectors, thereby drastically reducing the computational load of the edge module and enabling real-time independent operation without delay even on low-spec camera terminals.

[0296] In one embodiment of the present invention, by applying a time-series verification logic that determines whether the ratio of frames in which the similarity exceeds a threshold among consecutive image frames exceeds a preset ratio standard, without relying solely on the similarity result of a single frame, it is possible to fundamentally block false detections caused by temporary shaking or environmental noise and maximize detection accuracy.

[0297] In one embodiment of the present invention, the edge module compares and calculates a reference embedding vector and a target embedding vector for each of a plurality of predefined events (falling, assault, fire, etc.) in parallel, thereby enabling the simultaneous and independent detection of various dangerous situations that may occur in the field using only a single lightweight deep learning model, which can enhance the scalability and versatility of the system.

[0298] In one embodiment of the present invention, by performing a double verification process that calculates the average value of the delta embedding difference, which is the difference between target embedding vectors between adjacent image frames, and compares it with a reference pattern, it is possible to precisely analyze dynamic change patterns such as speed and acceleration beyond simple morphological similarity of objects, thereby significantly improving the ability to distinguish between falling or assault acts of a person.

[0299] In one embodiment of the present invention, by variably setting an adaptive threshold value by reflecting the normal similarity distribution (mean and standard deviation) that appears differently depending on the camera's shooting environment, it is possible to achieve the effect of ensuring uniform and stable detection performance even in poor installation environments, such as rapid changes in illumination between day and night or adverse weather conditions (snow, rain, etc.).

[0300] In one embodiment of the present invention, instead of performing heavy deep learning operations on the entire image, only local change areas where macroscopic movement occurs are lightly extracted by utilizing motion vectors calculated during the image encoding process or simple image difference operations between frames, and then input into an embedding model, and the deep learning inference operations themselves are completely omitted for static frames where no significant movement is detected, thereby minimizing unnecessary power consumption and memory occupation and maximizing the computing resource efficiency of edge devices.

[0302] 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 may be achieved even if the described techniques are performed in a different order than described, and / or if 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. Therefore, other implementations, other embodiments, and equivalents to the claims below are also within the scope of the claims.

Claims

Claim 1 A method for detecting a specific event in real time performed in an image detection system, wherein the image detection system comprises a camera module and an edge module including one or more processors and one or more memories; an image capturing step in which an image is captured by the camera module and transmitted to the edge module; a vector generation step in which a predetermined number of consecutive image frames among the image are input into a deep learning-based embedding model to generate a target embedding vector for each image frame by the edge module; and a similarity calculation step in which a similarity is calculated between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame by the edge module. A detection method comprising: an event determination step in which, by means of an edge module, when the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard, a specific event corresponding to the similarity has occurred within the image; wherein the embedding model is an artificial neural network model trained to output an embedding vector value having a vector similarity greater than or equal to a preset standard when the event labeled in the image frame is the same specific event, even if image frames included in images captured in different camera environments are input. Claim 2 A detection method according to claim 1, characterized in that the reference embedding vector is derived from the center point of the learning embedding vectors derived from a plurality of learning image frames for each specific event included in the learning data for the embedding model. Claim 3 A detection method according to claim 1, characterized in that the threshold value is adaptively set to correspond to the environment based on a similarity distribution that appears differently depending on the shooting environment of the camera module. Claim 4 A detection method according to claim 1, wherein the threshold value is calculated by summing the average value of a similarity distribution derived based on a plurality of similarity data obtained in a normal period in which no event occurs, and the value obtained by multiplying the standard deviation of the similarity distribution by a preset constant. Claim 5 A detection method according to claim 1, wherein the threshold value is set differently from each other by independently reflecting the similarity distribution that appears differently for each individual event when there are multiple predefined events. Claim 6 A method for detecting a specific event in real time performed in an image detection system, wherein the image detection system comprises a camera module and an edge module including one or more processors and one or more memories; an image capturing step in which an image is captured by the camera module and transmitted to the edge module; a vector generation step in which a predetermined number of consecutive image frames among the image are input into a deep learning-based embedding model to generate a target embedding vector for each image frame by the edge module; and a similarity calculation step in which a similarity is calculated between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame by the edge module. A detection method comprising: an event judgment step in which, by means of an edge module, if the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard, it is determined that a specific event corresponding to the similarity has occurred within the image; wherein the embedding model is an artificial neural network model trained to output an embedding vector value having a vector similarity greater than or equal to a preset standard when the event labeled in the image frame is the same specific event, even if image frames included in images captured in different camera environments are input; and wherein the vector generation step comprises: a change area extraction step in which a change area where movement has occurred is extracted based on image difference operations between adjacent image frames among the consecutive image frames, or based on a preset region-specific motion vector for the consecutive image frames; and a local vector derivation step in which only the image of the change area is input into the embedding model to derive the target embedding vector. Claim 7 A detection method according to claim 6, wherein the vector generation step omits the derivation of the target embedding vector and the calculation of the similarity for image frames in which the magnitude of the movement is less than a preset change standard. Claim 8 A method for detecting a specific event in real time performed in an image detection system, wherein the image detection system comprises a camera module and an edge module including one or more processors and one or more memories; an image capturing step in which an image is captured by the camera module and transmitted to the edge module; a vector generation step in which a predetermined number of consecutive image frames among the image are input into a deep learning-based embedding model to generate a target embedding vector for each image frame by the edge module; and a similarity calculation step in which a similarity is calculated between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame by the edge module. The method includes an event determination step in which, by means of an edge module, if the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard, it is determined that a specific event corresponding to the similarity has occurred within the image; wherein the reference embedding vector is stored in advance along with a reference difference average value calculated by determining the difference between a plurality of training embedding vectors derived from a plurality of training image frames for each specific event included in the training data for the embedding model; and the similarity calculation step includes an embedding difference calculation step that calculates a delta embedding difference, which is the difference between target embedding vectors between adjacent image frames among the consecutive image frames. A method for detecting, further comprising: a pattern analysis step of calculating a target difference average value, which is the average value of a plurality of delta embedding differences calculated from the consecutive image frames, and calculating pattern similarity by comparing the target difference average value with the reference difference average value; wherein the event judgment step further determines whether the pattern similarity exceeds a preset pattern threshold value to finally determine whether the event occurs. Claim 9 delete Claim 10 A method for detecting a specific event in real time performed in an image detection system, wherein the image detection system comprises a camera module and an edge module including one or more processors and one or more memories; an image capturing step in which an image is captured by the camera module and transmitted to the edge module; a vector generation step in which a predetermined number of consecutive image frames among the image are input into a deep learning-based embedding model to generate a target embedding vector for each image frame by the edge module; and a similarity calculation step in which a similarity is calculated between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame by the edge module. The detection method comprises: an event determination step in which, by means of an edge module, if the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard, it is determined that a specific event corresponding to the similarity has occurred within the image; wherein the embedding model is an artificial neural network model trained to output an embedding vector value having a vector similarity greater than or equal to a preset standard when the event labeled in the image frame is the same specific event, even if image frames included in images captured in different camera environments are input; and wherein the preset number is determined by the number of image frames acquired during a preset time interval in the image. Claim 11 A detection method according to claim 1, wherein the similarity is calculated through cosine similarity. Claim 12 A detection method according to claim 1, wherein the detection method is performed solely by similarity comparison operations between embedding vectors without object detection operations that classify objects within the image and track bounding boxes, thereby minimizing the computational load of the edge module. Claim 13 A video detection system for performing a method of detecting a specific event in real time, comprising: a camera module including a video capturing unit that captures an image and transmits it to an edge module; and an edge module including one or more processors and one or more memories; wherein the edge module comprises: a vector generation unit that inputs each of a predetermined number of partial consecutive image frames among the image into a deep learning-based embedding model to generate a target embedding vector for each image frame; and a similarity calculation unit that calculates the similarity between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame. An image detection system comprising: an event determination unit that determines that a specific event corresponding to the similarity has occurred within the image when the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard; wherein the embedding model is an artificial neural network model trained to output an embedding vector value having a vector similarity greater than or equal to a preset standard when the event labeled in the image frame is the same specific event, even if image frames included in images captured in different camera environments are input. Claim 14 An image detection system according to claim 13, wherein the reference embedding vector is derived from the center point of learning embedding vectors derived from a plurality of learning image frames for each specific event included in the learning data for the embedding model. Claim 15 An image detection system according to claim 13, wherein the threshold value is adaptively set to correspond to the environment based on a similarity distribution that appears differently depending on the shooting environment of the camera module. Claim 16 An image detection system according to claim 13, wherein the threshold value is calculated by summing the average value of a similarity distribution derived based on a plurality of similarity data obtained in a normal period in which no event occurs, and the value obtained by multiplying the standard deviation of the similarity distribution by a preset constant. Claim 17 An image detection system according to claim 13, wherein the threshold value is set differently from each other by independently reflecting the similarity distribution that appears differently for each individual event when there are multiple predefined events. Claim 18 A video detection system for performing a method of detecting a specific event in real time, comprising: a camera module including a video capturing unit that captures an image and transmits it to an edge module; and an edge module including one or more processors and one or more memories; wherein the edge module comprises: a vector generation unit that inputs each of a predetermined number of partial consecutive image frames among the image into a deep learning-based embedding model to generate a target embedding vector for each image frame; and a similarity calculation unit that calculates the similarity between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame. The image detection system comprises: an event judgment unit that determines that a specific event corresponding to the similarity has occurred within the image when the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard; wherein the embedding model is an artificial neural network model trained to output an embedding vector value having a vector similarity greater than or equal to a preset standard when the event labeled in the image frame is the same specific event, even if image frames included in images captured in different camera environments are input; and wherein the vector generation unit comprises: a change area extraction unit that extracts a change area where movement has occurred based on image difference operations between adjacent image frames among the consecutive image frames or a preset region-specific motion vector for the consecutive image frames; and a local vector derivation unit that derives the target embedding vector by inputting only the image of the change area into the embedding model. Claim 19 An image detection system according to claim 18, wherein the vector generation unit omits the derivation of the target embedding vector and the calculation of the similarity for an image frame in which the magnitude of the movement is less than a preset change standard. Claim 20 A video detection system for performing a method of detecting a specific event in real time, comprising: a camera module including a video capturing unit that captures an image and transmits it to an edge module; and an edge module including one or more processors and one or more memories; wherein the edge module comprises: a vector generation unit that inputs each of a predetermined number of partial consecutive image frames among the image into a deep learning-based embedding model to generate a target embedding vector for each image frame; and a similarity calculation unit that calculates the similarity between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame. The image detection system comprises: an event judgment unit that determines that a specific event corresponding to the similarity has occurred within the image when the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard; wherein the reference embedding vector is stored together with a reference difference average value calculated by determining the difference between a plurality of learning embedding vectors derived from a plurality of learning image frames for each specific event included in the learning data for the embedding model; and the similarity calculation unit further comprises: an embedding difference calculation unit that calculates a delta embedding difference, which is the difference between adjacent image frames among the consecutive image frames; and a pattern analysis unit that calculates a target difference average value, which is the average value of a plurality of delta embedding differences calculated from the consecutive image frames, and calculates pattern similarity by comparing the target difference average value with the reference difference average value; and wherein the event judgment unit further determines whether the pattern similarity exceeds a preset pattern threshold to finally determine whether the event has occurred. Claim 21 delete Claim 22 A video detection system for performing a method of detecting a specific event in real time, comprising: a camera module including a video capturing unit that captures an image and transmits it to an edge module; and an edge module including one or more processors and one or more memories; wherein the edge module comprises: a vector generation unit that inputs each of a predetermined number of partial consecutive image frames among the image into a deep learning-based embedding model to generate a target embedding vector for each image frame; and a similarity calculation unit that calculates the similarity between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame. The image detection system comprises: an event judgment unit that determines that a specific event corresponding to the similarity has occurred within the image when the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard; wherein the embedding model is an artificial neural network model trained to output an embedding vector value having a vector similarity greater than or equal to a preset standard when the event labeled in the image frame is the same specific event, even if image frames included in images captured in different camera environments are input; and wherein the preset number is determined by the number of image frames acquired during a preset time interval in the image. Claim 23 An image detection system according to claim 13, wherein the similarity is calculated through cosine similarity. Claim 24 A method for detecting a specific event in real time performed in an image detection system, wherein the image detection system includes an edge module comprising one or more processors and one or more memories, and wherein the edge module performs an image reception step of receiving an image captured by a camera module; wherein the edge module performs a vector generation step of generating a target embedding vector for each image frame by inputting each of a predetermined number of consecutive image frames among the image into a deep learning-based embedding model; and wherein the edge module performs a similarity calculation step of calculating the similarity between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame. A detection method comprising: an event determination step in which, by means of an edge module, when the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard, a specific event corresponding to the similarity has occurred within the image; wherein the embedding model is an artificial neural network model trained to output an embedding vector value having a vector similarity greater than or equal to a preset standard when the event labeled in the image frame is the same specific event, even if image frames included in images captured in different camera environments are input. Claim 25 A computer-readable storage medium for implementing a method for real-time detection of a specific event performed in an image sensing system comprising a camera module and an edge module comprising one or more processors and one or more memories, wherein the computer-readable storage medium comprises computer-executable instructions that cause the image sensing system to perform the following steps, the following steps comprising: an image capturing step of capturing an image by the camera module and transmitting it to the edge module; a vector generation step of generating a target embedding vector for each image frame by inputting each of a predetermined number of consecutive image frames in the image into a deep learning-based embedding model by the edge module; and a similarity calculation step of calculating a similarity between a reference embedding vector stored for each predetermined event and the target embedding vector for each image frame by the edge module. A computer-readable storage medium comprising: an event determination step in which, by means of an edge module, when the ratio of image frames having a similarity exceeding a preset threshold among the consecutive image frames exceeds a preset ratio standard, it is determined that a specific event corresponding to the similarity has occurred within the image; wherein the embedding model is an artificial neural network model trained to output an embedding vector value having a vector similarity greater than or equal to a preset standard when the event labeled in the image frame is the same specific event, even if image frames included in images captured in different camera environments are input.

Images