Adaptive object recognition method using environmental characteristics for each area of interest, and object detection system therewith

Abstract

The present invention relates to an adaptive object recognition method using environmental characteristics by region of interest and an object detection system using the same, wherein learning models (MN) constituting an AI algorithm are configured in N numbers according to environmental characteristic information, and when learning the learning models (MN), learning is performed using learning data according to the corresponding environmental characteristic information, and simultaneously, when analyzing images, at least one image of a region of interest (ROI) within a frame that is pre-set or set by a manager (User) is analyzed to detect environmental characteristic information, and then a learning model (MN) corresponding to the detected environmental characteristic information is selected to recognize an object and generate object attribute information, thereby enabling the recognition of an object using a suitable learning model (MN) according to an environment with various variables, which can not only dramatically increase the object recognition rate and accuracy but also improve the uncertainty of object attribute information (type, speed, direction, etc.).

Description

Technology Field

[0001] The present invention relates to an adaptive object recognition method using environmental characteristics by region of interest and an object detection system using the same. More specifically, the invention relates to an adaptive object recognition method using environmental characteristics by region of interest and an object detection system using the same, wherein N learning models (MN) constituting an AI algorithm are configured according to environmental characteristic information, and when learning the learning models (MN), learning is performed using learning data according to the corresponding environmental characteristic information, and simultaneously, when analyzing images, at least one image of a region of interest (ROI) within a frame that is pre-set or set by a manager (User) is analyzed to detect environmental characteristic information, and then a learning model (MN) corresponding to the detected environmental characteristic information is selected to recognize an object and generate object attribute information. By using a suitable learning model (MN) according to an environment with various variables to recognize an object, the object recognition rate and accuracy can be significantly increased, and the uncertainty of object attribute information (type, speed, direction, etc.) can be improved. Background Technology

[0002] Recently, as vehicle ownership rates have increased and urban areas have expanded, traffic congestion has become a part of daily life, causing massive economic losses. Consequently, various efforts are being made to resolve this, such as constructing new roads or expanding existing ones; however, these methods not only require substantial financial resources but also have limitations in that they are difficult to resolve in the short term.

[0003] Therefore, various studies are being conducted to detect vehicles, generate traffic information, and utilize the generated traffic information.

[0004] Meanwhile, research on image analysis technology for performing object recognition, object attribute detection, and tracking using AI artificial intelligence techniques has been actively conducted recently. The application fields of such AI-based image analysis technology are increasing exponentially due to the advantage of significantly improving the accuracy and precision of object recognition by reducing the error rate through learning.

[0005] FIG. 1 is a block diagram showing a controller disclosed in Korean Registered Patent No. 10-2561656 (Title of Invention: Vehicle Flow Control System Utilizing Deep-Learning-Based Object Sensing), filed and patented by the applicant.

[0006] The controller of FIG. 1 (hereinafter referred to as the prior art) (100) includes a plurality of GPUs.

[0007] In addition, the prior art (100) includes, as illustrated in FIG. 1, an artificial intelligence learning unit (133) that learns an object analysis algorithm, an image classification unit (134) that classifies a GPU to analyze each image according to the capacity of each input image and an image classification table created / updated by an image classification table creation / updating module (143) described later, and an artificial intelligence-based image analysis module (135-1), ..., (135-M) that, when an image classified by the image classification unit (134) is input, analyzes the input images using an object analysis algorithm learned by the artificial intelligence learning unit (133) and outputs object information (recognition, location, type, size, etc.).

[0008] This prior art (100) has the advantage of not only increasing the object recognition rate by optimally classifying images acquired from the shooting of cameras by M GPUs and analyzing the images using an object analysis algorithm, but also enabling real-time processing and analysis of high-capacity images.

[0009] However, the conventional technology (100) has the disadvantage that when learning an object analysis algorithm, the learning proceeds without taking into account environmental characteristic information such as snow, rain, and light intensity, so the accuracy of object recognition rate and object attribute information (type, speed, direction, etc.) is reduced, and the learning efficiency is low.

[0010] Generally, as road surveillance cameras are installed outdoors, they have the characteristic that the quality of the image is determined by various variables such as snow, rain, and illumination.

[0011] For example, assuming that a learning model is trained using data that does not account for environmental characteristics, the model may demonstrate excellent object recognition rates and accuracy through image analysis in environments without snow, but a problem arises where object recognition rates and accuracy deteriorate in environments with heavy snowfall.

[0012] That is, in the case of cameras installed outdoors, the quality is determined according to environmental characteristic information (the degree (numerical value) of parameters such as snow, rain, and illuminance), but in the prior art (100), the learning model is trained without taking into account such environmental characteristic information at all, so a problem arises in which the object recognition rate and the accuracy of object attribute information (type, speed, direction, etc.) are reduced. The problem to be solved

[0013] The present invention aims to solve these problems. The objective of the present invention is to provide an adaptive object recognition method using environmental characteristics by region of interest and an object detection system using the same, which can significantly increase the object recognition rate, improve the accuracy of object attribute information, and reduce unnecessary computational processing. This is achieved by configuring the control center server to generate learning models (MNs) according to environmental characteristic information, and during the learning of each learning model (MN), to use images corresponding to the environmental characteristic information of the learning model (MN) as learning data for learning, and when the controller analyzes the image, to analyze the image to detect environmental characteristic information, and then to analyze the image using the learning model (MN) corresponding to the detected environmental characteristic information.

[0014] Furthermore, another objective of the present invention is to provide an adaptive object recognition method using environmental characteristics by region of interest and an object detection system using the same, which can further enhance the object recognition rate and the accuracy of object attribute information by configuring the controller to detect environmental characteristic information by analyzing only the images of the set detection ROIs after setting detection ROIs in response to a request from an administrator (User), and then analyzing the images of the set detection ROIs during image analysis.

[0015] Furthermore, another objective of the present invention is to provide an adaptive object recognition method utilizing environmental characteristics by region of interest and an object detection system utilizing the same, wherein the controller is configured to separately set a monitoring ROI and then analyze the set monitoring ROI image to determine whether an unexpected situation has occurred, thereby separately observing areas requiring continuous monitoring to quickly and accurately determine whether an unexpected situation has occurred and enabling rapid response to the unexpected situation.

[0016] Furthermore, another objective of the present invention is to provide an adaptive object recognition method using environmental characteristics by region of interest and an object detection system using the same, which not only enhances service convenience and efficiency by configuring detection ROIs to allow pixel location changes according to the request of an administrator (User), but also further improves the object recognition rate and the accuracy of object attribute information by setting an area suitable for detecting environmental characteristic information as a detection ROI according to the administrator's judgment. means of solving the problem

[0017] The solution means of the present invention for solving the above problem comprises at least one shooting device comprising a camera and a controller that uses one of a preset learning model (MN) to analyze an image input from the camera, recognizes an object, and simultaneously detects object attribute information including at least one of the object's type, speed, and direction, and proceeds with the learning of the learning models (MN) according to a preset period (T) or a request from an administrator. The controller matches different environmental characteristic information to the learning models (MN), and during learning, performs learning by utilizing an image corresponding to the environmental characteristic information of the corresponding learning model (MN) as learning data. When an image is input from the camera, the controller analyzes the input image to detect environmental characteristic information, selects a learning model (MN) corresponding to the detected environmental characteristic information, and uses the selected learning model (MN) to analyze the input image and generate object attribute information.

[0018] In addition, in the present invention, the AI ​​algorithm learning process (S1) of the controller preferably includes: a step 10 (S10) in which the controller collects images to be used as training data and correct data (object recognition and object attribute information); a step 20 (S20) in which the controller analyzes the images collected in step 10 (S10), detects environmental characteristic information of the images, and labels the images; a step 30 (S30) in which the controller classifies the training data collected in step 10 (S10) according to the environmental characteristic information labeled in step 20 (S20); and a step 40 (S40) in which the controller utilizes the training data classified in step 30 (S30) to proceed with the training of the corresponding learning model (MN).

[0019] In addition, in the present invention, step 20 (S20) comprises: step 21 (S21) in which the controller receives the training data collected in step 10 (S10); step 22 (S22) in which the controller analyzes the image of the training data input through step 21 (S21); step 23 (S23) in which the controller detects a pre-set degree (numerical value) for each parameter by referencing and utilizing the analysis data detected in step 22 (S22); step 24 (S24) in which the controller searches a parameter-specific reference table, which is data in which a grade (level) is matched according to the range of the degree (numerical value) of each parameter, and searches for a parameter-specific grade (level) corresponding to the degree (numerical value) of each parameter detected in step 23 (S23); and step 25 (S25) in which the controller sets the parameter-specific grade (level) searched in step 24 (S24) as the parameter-specific grade (level) of the corresponding image. It is preferable to further include step 26 (S26), in which the controller generates environmental characteristic information including the grade (level) of each parameter set in step 25 (S25); and step 27 (S27), in which the controller labels the environmental characteristic information generated in step 26 (S26) onto the corresponding input image.

[0020] In addition, another solution of the present invention comprises at least one shooting device including a camera and a controller that analyzes an image input from the camera using one of a preset learning model (MN) to recognize an object and simultaneously detect object attribute information including at least one of the object's type, speed, and direction; and a control center server that stores and monitors the image and object attribute information received from the shooting device and performs learning of the learning models (MN) according to a preset period (T) or a request from an administrator, wherein the control center server matches different environmental characteristic information to the learning models (MN), and during learning, performs learning by utilizing an image corresponding to the environmental characteristic information of the corresponding learning model (MN) as learning data, and the controller, when an image is input from the camera, analyzes the input image to detect environmental characteristic information, selects a learning model (MN) corresponding to the detected environmental characteristic information, and analyzes the input image using the selected learning model (MN) to generate object attribute information.

[0021] In addition, in the present invention, the AI ​​algorithm learning process (S1) of the control center server preferably includes: a step 10 (S10) in which the control center server collects video and correct answer data (object recognition and object attribute information) to be used as learning data; a step 20 (S20) in which the control center server analyzes the video collected in step 10 (S10), detects environmental characteristic information of the video, and labels the video; a step 30 (S30) in which the control center server classifies the learning data collected in step 10 (S10) according to the environmental characteristic information labeled in step 20 (S20); and a step 40 (S40) in which the control center server utilizes the learning data classified in step 30 (S30) to proceed with the learning of the corresponding learning model (MN).

[0022] In addition, in the present invention, step 20 (S20) comprises: step 21 (S21), in which the control center server receives the learning data collected in step 10 (S10); step 22 (S22), in which the control center server analyzes the image of the learning data input through step 21 (S21); step 23 (S23), in which the control center server detects a pre-set degree (numerical value) for each parameter by referencing and utilizing the analysis data detected in step 22 (S22); and step 24 (S24), in which the control center server searches a parameter-specific reference table, which is data in which a grade (level) is matched according to the range of the degree (numerical value) of each parameter, and searches for a parameter-specific grade (level) corresponding to the degree (numerical value) of each parameter detected in step 23 (S23). It is preferable to further include: a step 25 (S25) in which the control center server sets the parameter-specific grade (level) discovered in step 24 (S24) as the parameter-specific grade (level) of the corresponding image; a step 26 (S26) in which the control center server generates environmental characteristic information including the grade (level) of each parameter set in step 25 (S25); and a step 27 (S27) in which the control center server labels the environmental characteristic information generated in step 26 (S26) onto the corresponding input image.

[0023] In addition, in the present invention, the controller preferably comprises: an image input unit that receives an image acquired by shooting with the camera; an environmental characteristic information detection unit that analyzes the image input through the image input unit and detects environmental characteristic information; a learning model (MN) selection unit that selects a learning model (MN) among the learning models (MN) that corresponds to the environmental characteristic information detected by the environmental characteristic information detection unit by referring to the environmental characteristic information matched to each of the learning models (MN); an image analysis unit that uses the learning model (MN) selected by the learning model (MN) selection unit to analyze the image input through the image input unit to recognize an object, and then outputs at least one of the type, speed, and direction of the recognized object; and an object attribute information generation unit that generates object attribute information by matching the information output from the image analysis unit.

[0024] In addition, in the present invention, the controller further includes an ROI setting unit for setting detection ROIs defined as pixel areas within a frame for detecting environmental characteristic information, and the environmental characteristic information detection unit includes: a detection ROI extraction module that extracts detection ROI images from an image frame input through the image input unit by referring to pixel location information of the detection ROI set by the ROI setting unit; an image analysis module that analyzes the detection ROI images extracted by the detection ROI extraction module; a parameter degree (numerical value) detection module that detects the degree (numerical value) of each detection ROI by referencing and utilizing the analysis data detected by the image analysis module by referencing and utilizing the pre-set parameter degree (numerical value) of each detection ROI; and a filtering module that filters parameter degrees (numerical values) with large deviations, limited to the same parameter, by using the deviation of the parameter degree (numerical value) of each detection ROI detected by the parameter degree (numerical value) detection module. It is preferable to include a parameter-specific grade (level) detection module that detects a parameter-specific grade (level) corresponding to the degree (numerical value) of each parameter of each detection ROI selected through the filtering module by searching a reference table of each parameter using the average value of the degree (numerical value) of each parameter of the corresponding image; and an environmental characteristic information detection module that detects environmental characteristic information including the grade (level) of each parameter detected by the parameter-specific grade (level) detection module. Effects of the invention

[0025] According to the present invention having the above problem and means of solution, a control center server generates learning models (MNs) for each environmental characteristic information, and when learning each learning model (MN), the learning is performed by utilizing images corresponding to the environmental characteristic information of the learning model (MN) as learning data, and when the controller analyzes the image, it analyzes the image to detect environmental characteristic information, and then analyzes the image using the learning model (MN) corresponding to the detected environmental characteristic information, thereby significantly increasing the object recognition rate and simultaneously improving the accuracy of object attribute information and reducing unnecessary computational processing.

[0026] In addition, according to the present invention, after the controller sets detection ROIs in response to a request from the administrator (User), it is configured to analyze only the images of the set detection ROIs during image analysis to detect environmental characteristic information, thereby further increasing the object recognition rate and the accuracy of object attribute information.

[0027] In addition, according to the present invention, the controller is configured to separately set a monitoring ROI and then analyze the set monitoring ROI image to determine whether an unexpected situation has occurred. This allows for the separate observation of areas requiring continuous monitoring, thereby enabling not only a rapid and accurate determination of whether an unexpected situation has occurred but also a rapid response to the unexpected situation.

[0028] In addition, according to the present invention, detection ROIs are configured to allow pixel location changes upon the request of an administrator (User), thereby increasing service convenience and efficiency. Furthermore, since an area suitable for detecting environmental characteristic information is set as the detection ROI based on the administrator's judgment, the object recognition rate and the accuracy of object attribute information can be further improved. Brief explanation of the drawing

[0029] FIG. 1 is a block diagram showing a controller disclosed in Korean Registered Patent No. 10-2561656 (Title of Invention: Vehicle Flow Control System Utilizing Deep-Learning-Based Object Sensing), filed and patented by the applicant. FIG. 2 is a configuration diagram showing an object detection system using image-based environmental characteristics, which is an embodiment of the present invention. Figure 3 is a flowchart illustrating the AI ​​algorithm learning process of the control center server of Figure 2. Figure 4 is a flowchart showing the labeling step (S20) of Figure 3. Figure 5 is a block diagram showing the controller of Figure 2. Figure 6 is a block diagram showing the ROI setting section of Figure 5. Figure 7 is an example diagram showing a region of interest (ROI) set in the detection ROI setting unit of Figure 6. Figure 8 is a block diagram showing the environmental characteristic information detection unit of Figure 5. Figure 9 is a conceptual diagram illustrating the process of analyzing an image by selecting a learning model of the controller of Figure 5. Specific details for implementing the invention

[0030] Hereinafter, an embodiment of the present invention will be described with reference to the attached drawings.

[0031] FIG. 2 is a configuration diagram showing a vehicle detection system using image-based environmental characteristics, which is an embodiment of the present invention.

[0032] The object detection system (1) using environmental characteristics by region of interest, which is an embodiment of the present invention of FIG. 2, is configured such that N learning models (MN) constituting an AI algorithm are configured according to environmental characteristic information, and when learning the learning models (MN), learning is performed using learning data according to the corresponding environmental characteristic information, and at the same time, when analyzing images, at least one image of a region of interest (ROI) within a frame that is pre-set or set by a manager (User) is analyzed to detect environmental characteristic information, and then a learning model (MN) corresponding to the detected environmental characteristic information is selected to recognize an object and generate object attribute information. This is so that the object recognition rate and accuracy can be significantly increased by using a suitable learning model (MN) according to an environment with various variables, and the uncertainty of object attribute information (type, speed, direction, etc.) can be improved.

[0033] In addition, the object detection system (1) utilizing environmental characteristics of the area of ​​interest according to the present invention is installed at various locations on the road as shown in FIG. 2, and consists of a shooting device (3-1), ..., (3-N) that captures a preset area (S) to acquire an image, analyzes the acquired image to recognize an object, and simultaneously generates object attribute information (type, direction, speed, etc.); a control center server (5) that stores and monitors object attribute information received from the shooting devices (3-1), ..., (3-N) and simultaneously learns the learning models (MN) of each shooting device (3); and a communication network (10) that provides a data transfer path between the control center server (5) and the shooting devices (3-1), ..., (3-N).

[0034] In this invention, for convenience of explanation, the controller (300) provided in each shooting device (3) is configured to independently analyze the image acquired from the shooting device (3), recognize an object, generate object attribute information, and then transmit it to the control center server (5). However, the controller (300) may be configured to be connected to multiple shooting devices (3) via a wired or wireless communication network to analyze images captured by multiple shooting devices (3).

[0035] Meanwhile, for the convenience of explanation, the present invention is described as an example in which the learning of learning models (MNs) is performed at the control center server (5) and the shooting device (3) connects to the control center server (5) to download the learned learning models (MNs); however, it is obvious that the learning of learning models (MNs) can be configured to be performed independently by the controller of the shooting device (3).

[0036] The communication network (10) supports data communication between the shooting devices (3-1), ..., (3-N) and the control center server (5), and can be implemented in detail as a wide area network (WAN), a local area network (LAN), a value-added network (VAN), a wired communication network, etc.

[0037] The control center server (9) stores and monitors images and object attribute information received from controllers (300) of the shooting devices (3-1), ..., (3-N), and at the same time references and utilizes the object attribute information to generate traffic information.

[0038] In addition, the control center server (9) learns learning models (MN) to be used for image analysis of each shooting device (3) according to a preset period (T) or a request from an administrator, and downloads the learned learning models (MN) to the corresponding shooting device (3).

[0039] At this time, the learning model (MN) is an artificial intelligence algorithm that takes an image as input data and outputs object recognition and attribute information (type, speed, direction, etc.) of the recognized object. Specifically, various learning models based on arbitrary convolutional neural network algorithms such as CNN, Faster R-CNN, R-CNN, YOLO, MASK-FCN, etc. can be used.

[0040] In addition, the control center server (9) learns learning models (MNs) according to environmental characteristic information. For example, assuming that environmental characteristic information is classified into 9 types, the control center server (9) learns 9 learning models (MNs).

[0041] In this case, environmental characteristic information includes grades (levels) for parameters such as snow, rain, and illuminance.

[0042] For example, assuming that the parameters are snow, rain, and illuminance, and the grades of the snowfall parameters are composed of 'large / medium / small / no', the grades of the precipitation parameters are composed of 'large / medium / small / no', and the grades of the illuminance parameters are composed of 'bright / medium / dark', the control center server (5) can detect 48 environmental characteristic information as shown in Table 1 below, and accordingly, can learn 48 learning models (MN).

[0043] [Table 1]

[0044]

[0045] In addition, when the control center server (5) learns a learning model (MN) according to each environmental characteristic information, it labels the images to be used as learning data according to the environmental characteristic information, selects the images according to the labeling, and uses them as learning data for the corresponding learning model (MN) so that learning is performed.

[0046] The shooting devices (3-1), ..., (3-N) are installed at various points along the road to capture images of a preset area (S).

[0047] Additionally, the shooting devices (3-1), ..., (3-N) include a controller (300) that analyzes the acquired image to recognize an object and simultaneously generates object attribute information.

[0048] At this time, the controller (300) links with the control center server (5) to periodically download and store learning models (MN), and when an image is acquired through shooting, it analyzes the acquired image to detect environmental characteristic information.

[0049] Meanwhile, for convenience of explanation, the present invention is described as an example in which the learning of learning models (MNs) is performed at the control center server (5) and the controller (300) receives and stores the learned learning models (MNs) from the control center server (5); however, the controller (300) may be configured to perform the learning of learning models (MNs) independently.

[0050] Additionally, when environmental characteristic information is detected, the controller (300) selects a learning model (MN) that corresponds to the detected environmental characteristic information among the learning models (MN), and then analyzes the acquired image using the selected learning model (MN) to recognize an object and simultaneously generate object attribute information.

[0051] That is, the controller (300) of the shooting device (3) of the present invention recognizes an object through image analysis using a learning model (MN) suitable for current environmental information characteristics, and simultaneously generates object attribute information, thereby flexibly responding to an environment with various variables to recognize the object, and can dramatically increase the object recognition rate and the accuracy of object attribute information.

[0052] Figure 3 is a flowchart illustrating the AI ​​algorithm learning process of the control center server of Figure 2.

[0053] As illustrated in FIG. 3, the AI ​​algorithm learning process (S1) of the control center server (5) consists of a learning data collection step (S10), a labeling step (S20), a learning data classification step based on environmental characteristics information (S30), a first, ..., N learning model (MN) learning step (S40-1), ..., (S40-N).

[0054] The learning data collection step (S10) is a step in which the control center server (5) collects video and correct answer data (object recognition and object attribute information) to be used as learning data.

[0055] Figure 4 is a flowchart showing the labeling step (S20) of Figure 3.

[0056] The labeling step (S20) consists of a training data input step (S21), an image analysis step (S22), a parameter-specific degree (numerical value) detection step (S23), a reference table search step (S24), a parameter-specific grade (level) setting step (S25), an environmental characteristic information generation step (S26), and an environmental characteristic information labeling step (S27).

[0057] The training data input step (S21) is a step of receiving the training data collected in the training data collection step (S10) respectively.

[0058] The image analysis step (S22) is a step for analyzing the images of the training data input through the training data input step (S21).

[0059] The parameter-specific degree (numerical value) detection step (S23) is a step of detecting the pre-set parameter-specific degree (numerical value) by referring to and utilizing the analysis data detected in the image analysis step (S22).

[0060] For example, when the parameters include precipitation, the parameter-specific degree (numerical value) detection step (S23) can detect how much precipitation there is.

[0061] The reference table search step (S24) searches the reference table of each parameter to find the grade (level) of each parameter corresponding to the degree (numerical value) of each parameter detected in the parameter degree (numerical value) detection step (S23).

[0062] In this case, the reference table for each parameter is data in which grades (levels) are matched according to the range of the degree (numerical value) of the corresponding parameter.

[0063] The parameter-specific grade (level) setting step (S25) sets the parameter-specific grade (level) found in the reference table search step (S24) as the parameter-specific grade (level) of the corresponding input image.

[0064] The environmental characteristic information generation step (S26) generates environmental characteristic information including the grade (level) of each parameter set in the parameter-specific grade (level) setting step (S25).

[0065] The environmental characteristic information labeling step (S27) labels the environmental characteristic information generated in the environmental characteristic information generation step (S26) onto the corresponding input image.

[0066] Returning to Fig. 3, if we look at the learning data classification step (S30) based on environmental characteristic information, the learning data classification step (S30) based on environmental characteristic information is a step of classifying the learning data collected in the learning data collection step (S10) according to the environmental characteristic information labeled in the labeling step (S20).

[0067] The first, ..., N learning model (MN) learning step (S40-1), ..., (S40-N) is a step of learning a learning model (MN) corresponding to the corresponding environmental characteristic information by utilizing the learning data classified in the environmental characteristic information-based learning data classification step (S30).

[0068] In this way, the control center server (5) of the present invention configures M environmental characteristic information according to the type of preset parameters and simultaneously configures M learning models (MN) for each environmental characteristic information, and when learning the learning models (MN), it is possible to build learning models (MN) suitable for an environment with various variables by utilizing only the learning data corresponding to the environmental characteristic information of each learning model (MN) during the learning of the learning models (MN).

[0069] Additionally, when the training of the training model (MN) is completed, the training center server (5) downloads the completed training models (MN) to the controller (300) of each shooting device (3).

[0070] Figure 5 is a block diagram showing the controller of Figure 2.

[0071] The controller (300) of Fig. 5 is a controller installed in the aforementioned shooting device (3) of Fig. 2 and manages and controls the operation of the shooting device (3).

[0072] In addition, as shown in FIG. 5, the controller (300) is composed of a control unit (30), a memory (31), a communication interface unit (32), an AI algorithm update unit (33), an ROI setting unit (34), a shooting control unit (35), an image input unit (36), an environmental characteristic information detection unit (37), a learning model (MN) selection unit (38), an image analysis unit (39), an object recognition unit (40), an object attribute information generation unit (41), a sudden occurrence determination unit (42), and a sudden occurrence data generation unit (43).

[0073] The control unit (30) is the OS (Operating System) of the controller (300) and manages and controls the operation of the control targets (31), (32), (33), (34), (35), (36), (37), (38), (39), (40), (41), (42), (43).

[0074] In addition, the control unit (30) executes the AI ​​algorithm update unit (33) according to a preset period (T) or a request from the manager.

[0075] Additionally, when the control unit (30) receives a request from the administrator (User) to set the region of interest (ROI), it executes the ROI setting unit (34).

[0076] Additionally, when an image obtained by shooting with a camera is input to the control unit (30), the input image is output to the image input unit (36).

[0077] Additionally, when object attribute information is generated in the object attribute information generation unit (41), the control unit (30) controls the communication interface unit (32) so that the generated object attribute information and image are transmitted to the control center server (5).

[0078] In memory (31), the learning models (MN) of the previously learned AI algorithm and the environmental characteristic information of each learning model (MN) are matched and stored.

[0079] In addition, the memory (31) stores the identification information and communication address of the pre-set controller (3).

[0080] In addition, the memory (31) temporarily stores images obtained from the camera's shooting.

[0081] In addition, the memory (31) stores pixel location information within the frame of each region of interest (ROI) set in the ROI setting unit (34).

[0082] The communication interface unit (32) transmits and receives data with the control center server (5).

[0083] The AI ​​algorithm update unit (33) is executed according to the control of the control unit (30) at a preset period (T) or at the request of the administrator (User), and is linked with the control center server (5) to download the learning model (MN) of the AI ​​algorithm created / updated and the environmental characteristic information matched to the learning model (MN), and stores the downloaded learning model (MN) and environmental characteristic information in memory (31).

[0084] At this time, the learning models (MNs) are composed of a quantity equal to the quantity of environmental characteristic information.

[0085] FIG. 6 is a block diagram showing the ROI setting section of FIG. 5, and FIG. 7 is an example diagram showing the region of interest (ROI) set in the detection ROI setting section of FIG. 6.

[0086] The ROI setting unit (34) of Fig. 6 is executed under the control of the control unit (30) when a request to set an area of ​​interest (ROI) is received from a user.

[0087] In addition, the ROI setting unit (34) is composed of a first selection module (341), a detection ROI generation module (342), a detection ROI setting module (343), a second selection module (344), a detection target setting module (345), a monitoring ROI setting module (346), and a data storage module (347), as shown in FIG. 6.

[0088] The first selection module (341) is executed when a user requests the setting of a detection ROI, and the user selects a pixel location.

[0089] In this case, the detection ROI refers to a region of interest set to detect environmental characteristic information.

[0090] When a specific pixel is selected by an operator through the first selection module (341), the detection ROI generation module (342) generates a predetermined area based on the selected pixel as a detection ROI.

[0091] The detection ROI setting module (343) sets the detection ROI generated by the detection ROI generation module (342) as the detection ROI.

[0092] The second selection module (344) is executed when a user requests the setting of a monitoring ROI, and the user selects a pixel location.

[0093] In this context, the monitoring ROI refers to an area of ​​interest used to determine the occurrence of puddles, ice, sinkholes, road damage, etc.

[0094] The detection target setting module (345) receives the detection target of the monitoring ROI selected in the second selection module (344) from the operator.

[0095] In this case, the detection targets may be puddles, ice, sinkholes, road damage, etc.

[0096] The monitoring ROI setting module (346) sets a predetermined area based on the pixel location selected by the second selection module (344) as the monitoring ROI, and at the same time matches the pixel location information of the monitoring ROI with the detection target information set by the detection target setting module (345).

[0097] The data storage module (347) stores the detection ROI information set in the detection ROI setting module (343) in the memory (31).

[0098] Additionally, the data storage module (347) stores the monitoring ROI information set in the monitoring ROI setting module (346) in the memory (31).

[0099] Returning to Fig. 5 and looking at the shooting control unit (35), the shooting control unit (35) manages and controls the operation of the camera.

[0100] The video input unit (36) receives the video obtained by the camera's shooting.

[0101] Figure 8 is a block diagram showing the environmental characteristic information detection unit of Figure 5.

[0102] The environmental characteristic information detection unit (37) of Fig. 8 is a processor for detecting environmental characteristic information by analyzing an image input through the image input unit (36).

[0103] In addition, the environmental characteristic information detection unit (37) is composed of a detection ROI extraction module (371), an image analysis module (372), a parameter-specific degree (numerical value) detection module (373), a filtering module (374), a parameter-specific grade (level) detection module (375), and an environmental characteristic information detection module (376), as shown in FIG. 8.

[0104] The detection ROI extraction module (371) extracts detection ROI images from an image frame input through the image input unit (35) by referring to the pixel location information of a preset detection ROI.

[0105] The image analysis module (372) analyzes the detection ROI images extracted by the detection ROI extraction module (371).

[0106] The parameter-specific degree (numerical value) detection module (373) refers to and utilizes the analysis data detected by the image analysis module (372) to detect the preset parameter-specific degree (numerical value) of each detection ROI.

[0107] For example, assuming that the parameters include snowfall amount, the parameter-specific degree (numerical value) detection module (373) analyzes the detection ROI image and detects the degree (numerical value) of the snowfall amount of each detection ROI.

[0108] The filtering module (374) uses the deviation of the parameter-specific degree (numerical value) of each detection ROI detected by the parameter-specific degree (numerical value) detection module (373) to filter out parameter degrees (numerical values) with large deviations, limited to the same parameter.

[0109] The parameter-specific grade (level) detection module (375) uses the average value of the degree (numerical value) of each parameter of each detection ROI selected through the filtering module (374) to search the reference table of each parameter and detect the parameter-specific grade (level) corresponding to the degree (numerical value) of each parameter of the corresponding image.

[0110] In this case, the reference table for each parameter is data in which grades (levels) are matched according to the range of the degree (numerical value) of the corresponding parameter.

[0111] The environmental characteristic information detection module (376) detects environmental characteristic information including the grade (level) of each parameter detected by the parameter-specific grade (level) detection module (375).

[0112] Returning to Fig. 5, when looking at the learning model (MN) selection unit (38), the learning model (MN) selection unit (38) refers to and utilizes the environmental characteristic information of the learning models (MNs) stored in memory (31) to select a learning model (MN) corresponding to the environmental characteristic information detected by the environmental characteristic information detection unit (37).

[0113] The image analysis unit (39) uses a learning model (MN) selected from the learning model (MN) selection unit (38) to analyze the image input from the image input unit (36), recognizes the object, and simultaneously outputs object attribute information.

[0114] The object recognition unit (40) recognizes an object by referencing and utilizing the output data detected by the image analysis unit (39).

[0115] The object attribute information generation unit (41) generates object attribute information including the type, direction, and speed of an object by referencing and utilizing the output data detected by the image analysis unit (39).

[0116] At this time, the control unit (30) controls the communication interface unit (32) to transmit the object attribute information generated by the object attribute information generation unit (41) to the control center server (5).

[0117] The sudden occurrence determination unit (42) refers to and utilizes pre-set monitoring ROI information to extract a monitoring ROI image from an image input through the image input unit (36), and then analyzes the extracted monitoring ROI to determine whether a sudden occurrence has occurred.

[0118] For example, assuming the detection target is a puddle, the sudden occurrence determination unit (42) analyzes the image of the monitoring ROI and determines that a sudden occurrence has occurred if the size of the puddle is greater than or equal to a threshold.

[0119] When the sudden event data generation unit (43) determines that a sudden event has occurred in the sudden event determination unit (42), it generates sudden event data including the location of the sudden event, the content of the sudden event, etc.

[0120] At this time, the sudden data generated by the sudden data generation unit (43) is transmitted to the control center server (5) through the communication interface unit (32) under the control of the control unit (30).

[0121] Figure 9 is a conceptual diagram illustrating the process of analyzing an image by selecting a learning model of the controller of Figure 5.

[0122] Assuming the parameter is 'snowfall amount', as illustrated in FIG. 9, when the controller (5) obtains an image through the shooting of a camera, it extracts detection ROI images from the obtained image, and then analyzes the extracted detection ROI images to detect environmental characteristic information.

[0123] Additionally, when environmental characteristic information is detected, the controller (5) refers to the environmental characteristic information matched to the learning models (MNs) stored in the memory (31), selects a learning model (MN) corresponding to the detected environmental characteristic information, and then uses the selected learning model (MN) to analyze the image to recognize an object and simultaneously detect object attribute information (type, speed, direction, etc. of the object).

[0124] In this way, the object detection system (1) using environmental characteristics by region of interest, which is an embodiment of the present invention, is configured such that a control center server (5) generates learning models (MNs) according to environmental characteristic information, and when learning each learning model (MN), the learning is performed by utilizing images corresponding to the environmental characteristic information of the corresponding learning model (MN) as learning data, and when the controller analyzes the image, it analyzes the image to detect environmental characteristic information, and then analyzes the image using the learning model (MN) corresponding to the detected environmental characteristic information, thereby significantly increasing the object recognition rate and simultaneously improving the accuracy of object attribute information and reducing unnecessary computational processing.

[0125] In addition, the object detection system (1) using environmental characteristics by region of interest of the present invention is configured such that, after the controller (300) sets detection ROIs according to the request of the administrator (User), only the images of the set detection ROIs are analyzed during image analysis to detect environmental characteristic information, thereby further increasing the object recognition rate and the accuracy of object attribute information.

[0126] In addition, the object detection system (1) utilizing environmental characteristics of the area of ​​interest according to the present invention is configured such that the controller (300) separately sets a monitoring ROI and then analyzes the set monitoring ROI image to determine whether an unexpected situation has occurred, thereby allowing the area requiring continuous monitoring to be observed separately, so that not only can the occurrence of an unexpected situation be determined quickly and accurately, but a rapid response to the unexpected situation can also be made.

[0127] In addition, the object detection system (1) using environmental characteristics by region of interest according to the present invention is configured such that the detection ROIs can change pixel positions according to the request of the administrator (User), thereby increasing service convenience and efficiency. Furthermore, since the area that is easy to detect environmental characteristic information is set as the detection ROI according to the administrator's judgment, the object recognition rate and the accuracy of object attribute information can be further improved. Explanation of the symbols

[0128] 1: Object detection system using environmental characteristics by region of interest 3-1, ..., 3-N: Recording devices 5: Control center server 10: Communication network 30: Control unit 31: Memory 32: Communication interface section 33: AI Algorithm Update Section 34: ROI Setting Section 35: Shooting control unit 36: Video input unit 37: Environmental characteristic information detection unit 38: Learning model (MN) selection unit 39: Image Analysis Unit 40: Object Recognition Unit 41: Object attribute information generation unit 42: Sudden occurrence determination unit 43:Sudden data generation unit 300:Controller

Claims

Claim 1 An adaptive object recognition system utilizing environmental characteristics by region of interest comprising at least one capturing device including a camera and a controller configured to recognize an object by analyzing an image input from the camera using one of a plurality of preset learning models (MNs) and simultaneously detect object attribute information including at least one of the type, speed, and direction of the object, and to perform learning of the plurality of learning models (MNs) according to a preset period (T) or a request from an administrator, wherein the controller comprises: an image input unit receiving an image acquired by the capturing of the camera; an ROI setting unit setting at least one detection ROI defined as a pixel area within a frame for detecting environmental characteristic information; an environmental characteristic information detection unit analyzing the image input through the image input unit to detect environmental characteristic information; a learning model (MN) selection unit selecting a learning model (MN) corresponding to the environmental characteristic information detected by the environmental characteristic information detection unit by referring to the environmental characteristic information matched to each of the learning models (MNs); and an image analysis unit analyzing the image using the learning model (MN) selected by the learning model (MN) selection unit. The system includes an object attribute information generation unit that generates object attribute information using the output result of the image analysis unit; and an AI algorithm update unit that performs learning of the plurality of learning models (MN), wherein the environment characteristic information detection unit comprises: a detection ROI extraction module that extracts a detection ROI image from an image frame by referencing pixel location information of a detection ROI set in the ROI setting unit; an image analysis module that analyzes the detection ROI image; a parameter-specific degree (numerical value) detection module that calculates the degree (numerical value) of each of the plurality of image parameters of each detection ROI by referencing the analysis data of the image analysis module; and a filtering module that filters values ​​with large deviations for the same parameter using the deviation of the parameter-specific degree (numerical value).An adaptive object recognition system using environmental characteristics by region of interest, comprising: a parameter-specific grade (level) detection module that searches a parameter-specific reference table using the average value of the filtered parameter-specific degrees (numerical values) and determines the grade (level) of each parameter; and an environmental characteristic information detection module that generates complex environmental characteristic information including the determined grade (level) of each parameter, wherein the AI ​​algorithm update unit is configured to: ① collect images and ground truth data to be used as training data; ② calculate the parameter-specific degrees (numerical values) for the collected images and determine the grade (level) according to the parameter-specific reference table to generate and label environmental characteristic information; ③ classify the training data according to the environmental characteristic information; and ④ utilize the classified training data to train a learning model (MN) that matches the corresponding environmental characteristic information. Claim 2 delete Claim 3 delete Claim 4 An object detection system utilizing environmental characteristics by region of interest, comprising: at least one shooting device including a camera and a controller that analyzes an image input from the camera to recognize an object and simultaneously generates object attribute information including at least one of the object's type, speed, and direction; and a control center server connected to the shooting device via a communication network, which stores and monitors the image and object attribute information received from the shooting device and performs learning of a plurality of learning models (MN) according to a preset period (T) or a request from an administrator, wherein the controller comprises: an image input unit configured to receive and store an image frame acquired by the camera's shooting and transmit it to a subsequent processing module; an ROI setting unit configured to set at least one detection ROI defined as a pixel area within a frame for detecting environmental characteristic information, wherein the detection ROI is created based on a pixel location selected by a user and the pixel location information of the detection ROI is stored; and an environmental characteristic information detection unit configured to generate environmental characteristic information by extracting and analyzing an area corresponding to the detection ROI set by the ROI setting unit from an image frame input through the image input unit. A learning model (MN) selection unit configured to select a learning model (MN) corresponding to the environmental characteristic information generated by the environmental characteristic information detection unit by referring to environmental characteristic information matched to each of the plurality of learning models (MN) stored in memory; and an image analysis unit configured to recognize an object by analyzing an image input through the image input unit using the learning model (MN) selected by the learning model (MN) selection unit.The system includes an object attribute information generation unit configured to generate object attribute information including at least one of the type, speed, and direction of a recognized object by referring to the output result of the image analysis unit; and the environment characteristic information detection unit comprises: a detection ROI extraction module that extracts a detection ROI image from an image frame by referring to pixel location information of a detection ROI set by the ROI setting unit; an image analysis module that analyzes the detection ROI image; a parameter-specific degree (numerical value) detection module that calculates the degree (numerical value) of each of a plurality of image parameters of each detection ROI by referring to the analysis data of the image analysis module; a filtering module that filters values ​​with large deviations for the same parameter using the deviation of the parameter-specific degree (numerical value); and a parameter-specific grade (level) detection module that searches a reference table for each parameter and determines the grade (level) of each parameter using the average value of the filtered parameter-specific degree (numerical value). An object detection system utilizing environmental characteristics by region of interest, comprising an environmental characteristic information detection module that generates complex environmental characteristic information including grades (levels) of each determined parameter, wherein the control center server is configured to match different environmental characteristic information to each of the plurality of learning models (MNs) and to train each learning model (MN) using only the learning data corresponding to the environmental characteristic information during training, wherein in the learning process, images and ground truth data to be used as training data are collected, the degree (numerical value) of each of the plurality of image parameters is calculated from the collected images, and grades (levels) are determined according to a reference table for each parameter to generate environmental characteristic information and label the corresponding images, and the learning data is classified according to the environmental characteristic information, and each learning model (MN) that matches the environmental characteristic information is trained using the classified learning data. Claim 5 delete Claim 6 delete Claim 7 delete Claim 8 delete

Images