KR20230173814A - 3D bound boxing method using vanishing point & Traffic flow controlling method for controlling traffic signal with sensing object - Google Patents

Abstract

The present invention includes a method for generating a 3D bounding box for an object using a vanishing point to increase the accuracy of object recognition in an image, and applying this 3D bounding box generation method as an object detection method of artificial intelligence to detect objects based on object information. This relates to an object-sensitive vehicle flow control method that controls a signaling system.
The present invention is implemented by an artificial intelligence-based image analysis unit of a computer device, and includes the steps of: (a) analyzing the camera installation angle and the directionality of the image captured by the camera to derive the longitudinal vanishing point VP1 and the transverse vanishing point VP2; ; (b) specifying a moving object in the image to be analyzed; (c) generating a left lower vanishing line Van-L1 and a right upper vanishing line Van-L2 connected from the VP1 to contact the outer edge of the object; (d) In the VP2, a front lower vanishing line Van-L3 and a rear upper vanishing line Van-L4 connected to contact the outer edge of the object are created, and the intersection point P1 of the Van-L1 and Van-L3, and the Van-L2 A step in which the intersection P2 of Van-L4 is created; (e) Create a left rear vertical line Ver-L1 in contact with the leftmost end of the object and a right front vertical line Ver-L2 in contact with the rightmost end of the object, thereby creating an intersection P3 of the Ver-L1 and Van-L1, and the Ver-L2 generating an intersection point P4 of Van-L3, an intersection P5 of Ver-L1 and Van-N4, and an intersection P6 of Ver-L2 and Van-L2; (f) generating a vanishing line Van-L5 passing from the VP2 to P3 and a vanishing line Van-L6 passing from the VP2 to P6; (g) Creating a left front vertical line Ver-L3 passing through P1 and a right rear vertical line Ver-L4 passing through P2, creating an intersection point P7 of Ver-L3 and Van-L6, and Ver-L4 and Van-L5. A step in which the intersection point P8 is created; and (h) generating a hexahedral bounding box in which the P2, P5, P7, and P6 form the upper vertices, and the P1, P4, P8, and P3 form the lower vertices; Provides a “3D bounding box generation method using vanishing points, including.”

Description

3D bounding box creation method using vanishing point and object sensitive vehicle flow control method for controlling traffic signal {3D bound boxing method using vanishing point & Traffic flow controlling method for controlling traffic signal with sensing object}

The present invention includes a method for generating a 3D bounding box for an object using a vanishing point to increase the accuracy of object recognition in an image, and applying this 3D bounding box generation method as an object detection method of artificial intelligence to detect objects based on object information. This relates to an object-sensitive vehicle flow control method that controls a signaling system.

As vehicle penetration increases and urban centers expand, traffic congestion has become routine, causing economic losses. Additionally, in emergency situations such as emergency patients or fire outbreaks, vehicle congestion in urban areas can cause irreversible losses.

Therefore, various studies are being conducted to reduce vehicle waiting times by optimizing the signal system at intersections. Representative conventional intersection signal systems include a periodic signal operation method that outputs a fixed signal period and presentation time according to TOD (Time Of Day) by utilizing accumulated data detected by vehicle detection means, and collected traffic information It is classified as an active signal operation method that controls signals according to traffic conditions.

In particular, various studies are being conducted on the active signal operation method due to its advantage in effectively reducing vehicle congestion by responding flexibly to current traffic conditions.

Meanwhile, recently, research has been actively conducted on video analysis technology to perform object detection, type, and tracking using deep learning techniques, and video analysis technology using deep learning is Its application fields are increasing exponentially due to the advantage of significantly increasing the accuracy and precision of detection by reducing the error rate through learning.

1. Publication Patent No. 10-2021-0103865 “Vanishing point extraction device and vanishing point extraction method” 2. Registered Patent 10-2311236 “Speed measurement method and system through deep learning object recognition and tracking” 3. Registered Patent 10-2391853 “Image information processing system and method” 4. Publication Patent 10-202000141834 “Video-based traffic signal control device and method”

The purpose of the present invention is to provide a 3D object detection method to increase recognition accuracy for objects in images.

In addition, the present invention analyzes images captured by cameras to generate object information and traffic information, creates an optimal presentation system according to the generated traffic information, and then controls traffic signals according to the generated optimal presentation system. Another purpose is to provide an object-sensitive vehicle flow control method that can significantly reduce vehicle congestion rate and waiting time, as well as reduce social costs by effectively reducing fuel consumption and pollutant gas emissions. there is.

In order to solve the above-mentioned problem, the present invention is implemented by an artificial intelligence-based image analysis unit of a computer device, and (a) analyzes the camera installation angle and the directionality of the image captured by the camera to determine the longitudinal vanishing point VP1 and the transverse direction. Deriving the vanishing point VP2; (b) specifying a moving object in the image to be analyzed; (c) generating a left lower vanishing line Van-L1 and a right upper vanishing line Van-L2 connected from the VP1 to contact the outer edge of the object; (d) In the VP2, a front lower vanishing line Van-L3 and a rear upper vanishing line Van-L4 connected to contact the outer edge of the object are created, and the intersection point P1 of the Van-L1 and Van-L3, and the Van-L2 A step in which the intersection P2 of Van-L4 is created; (e) Create a left rear vertical line Ver-L1 in contact with the leftmost end of the object and a right front vertical line Ver-L2 in contact with the rightmost end of the object, thereby creating an intersection P3 of the Ver-L1 and Van-L1, and the Ver-L2 generating an intersection point P4 of Van-L3, an intersection P5 of Ver-L1 and Van-N4, and an intersection P6 of Ver-L2 and Van-L2; (f) generating a vanishing line Van-L5 passing from the VP2 to P3 and a vanishing line Van-L6 passing from the VP2 to P6; (g) Creating a left front vertical line Ver-L3 passing through P1 and a right rear vertical line Ver-L4 passing through P2, creating an intersection point P7 of Ver-L3 and Van-L6, and Ver-L4 and Van-L5. A step in which the intersection point P8 is created; and (h) generating a hexahedral bounding box in which the P2, P5, P7, and P6 form the upper vertices, and the P1, P4, P8, and P3 form the lower vertices; Provides a “3D bounding box generation method using vanishing points, including.”

In addition, the present invention is a vehicle flow control system including cameras that acquire images by photographing at least one road that is a detection area, a traffic signal controller that controls the operation of traffic lights installed in the detection area, and a controller. An object-sensitive vehicle flow control method, which is an operation process, comprising: (A) acquiring images by having the cameras photograph a road assigned to them, and then transmitting the obtained images to the controller; (B) The controller analyzes the images received through step (A) using a deep learning-based object analysis algorithm and detects objects according to the 3D bounding box creation method using the vanishing point. simultaneously detecting object information including at least one of the location, type, and size of the detected object; (C) The controller tracks the detected object by referring to the object information detected in step (B), the location information of the preset shooting area of each camera, and the location information of the road, and determines the location of each object. Generating object analysis information including at least one of lane type, lane location, moving speed, and direction; (D) the controller utilizing and processing the object analysis information generated in step (C) to generate traffic information including at least one of queue, waiting time, and traffic volume for each connecting lane; (E) the controller establishing an optimal signal system based on the traffic information generated in step (D); and (F) the controller transmits the optimal signal system information established in step (E) to the traffic signal controller, and the traffic signal controller operates the traffic lights according to the optimal signal system information received from the controller. controlling traffic lights; An object-sensitive vehicle flow control method comprising an object-sensitive vehicle flow control method is also provided.

In step (B), the controller can be configured to detect object information by estimating the type and size of the detected object according to bounding box analysis reflecting the camera installation angle and distance information.

According to the method of generating a 3D bounding box using a vanishing point provided by the present invention, the accuracy and reliability of analysis can be dramatically increased by improving the object recognition rate.

In addition, according to the object-sensitive vehicle flow control method for traffic signal control provided by the present invention, images captured by cameras are analyzed to generate object information and traffic information, and then an optimal display system is created according to the generated traffic information. After generating, traffic signals are configured to be controlled according to the created optimal display system, which not only significantly reduces vehicle congestion rate and waiting time, but also effectively reduces fuel consumption and pollutant gas emissions, thereby reducing social costs. It becomes possible.

[FIG. 1A] to [FIG. 1H] are step-by-step illustrations of a method for generating a 3D bounding box using a vanishing point provided by the present invention.
[Figure 2] is a configuration diagram showing a vehicle flow control system to which the object-sensitive vehicle flow control method, which is an embodiment of the present invention, is applied.
[Figure 3] is an example diagram showing [Figure 2].
[Figure 4] is a flowchart showing the object-sensitive vehicle flow control method (S1), which is an embodiment of the present invention.
[Figure 5] is a flow chart showing the image classification steps in [Figure 4].
[Figure 6] is a flow chart showing the artificial intelligence-based image analysis steps in [Figure 4].
[Figure 7] is an example diagram to explain the number recognition step in [Figure 4].
[FIG. 8] is an example diagram showing a tail-biting violation vehicle determined by the violation vehicle determination step in [FIG. 4].
[Figure 11] is a flow chart showing the classification table optimization step in [Figure 4].

1. How to create a 3D bounding box using vanishing points

The present invention provides a “3D bounding box generation method using vanishing points” implemented according to steps (a) to (h) below. Steps (a) to (h) are all implemented by the artificial intelligence-based image analysis unit of the computer device, and in the “object-sensitive vehicle flow control method” described later, the “controller” is the “artificial intelligence of the computer device.” It performs the role of “based video analysis department.”

Below, each step of the “3D bounding box creation method using vanishing points” is explained along with the attached drawings. Therefore, the following positional concepts such as front/back, top/bottom, left/right, etc. are established based on the drawing.

In step (a), as shown in [Figure 1a], the camera installation angle and the directionality of the image captured by the camera are analyzed to derive the vertical vanishing point VP1 and the horizontal vanishing point VP2.

In step (b), the moving object in the image to be analyzed is specified as shown in [Figure 1b]. In [FIG. 1B], Object 1 and Object 2 are moving objects, and a 3D bounding box will be created for each of them. However, for convenience of explanation, the explanation will focus on Object 1 below. Hereinafter, “object” corresponds to “Obcect 1”.

In step (c), as shown in [FIG. 1C], a left lower vanishing line Van-L1 and a right upper vanishing line Van-L2 connected from the VP1 to contact the outer edge of the object are created. Van-L1 is a vanishing line that touches the lower left end of the object in VP1, and Van-L2 is a vanishing line that touches the upper right end of the object.

In step (d), as shown in [FIG. 1d], a front lower vanishing line Van-L3 and a rear upper vanishing line Van-L4 connected from the VP2 to contact the outer edge of the object are created. Van-L3 is a vanishing line that touches the lower front end of the object in VP2, and Van-L4 is a vanishing line that touches the upper rear end of the object in VP2. Accordingly, in step (d), an intersection point P1 between Van-L1 and Van-L3 and an intersection point P2 between Van-L2 and Van-L4 are created. P1 is the front lower left vertex of the 3D bounding box (hereinafter abbreviated as 'bounding box') for the object, and P2 is the rear upper right vertex of the bounding box.

In step (e), as shown in [FIG. 1e], a left rear vertical line Ver-L1 in contact with the leftmost end of the object and a right front vertical line Ver-L2 in contact with the rightmost end of the object are created. Ver-L1 is a vertical line created from the leftmost end point of the object, and Ver-L2 is a vertical line created from the rightmost end point of the object. Accordingly, in step (e), the intersection point P3 of Ver-L1 and Van-L1, the intersection point P4 of Ver-L2 and Van-L3, the intersection point P5 of Ver-L1 and Van-N4, and the intersection point P5 of Ver-L1 and Van-N4, and Ver-L2 The intersection point P6 of Van-L2 is created. P3 is the rear lower left vertex of the bounding box, P4 is the front lower right vertex of the bounding box, P5 is the rear upper left vertex of the bounding box, and P6 is the front upper right vertex of the bounding box.

In step (f), as shown in [FIG. 1f], a vanishing line Van-L5 passing from VP2 to P3 and a vanishing line Van-L6 passing from VP2 to P6 are generated. The rear bottom horizontal line of the bounding box is formed by Van-L5, and the front top horizontal line of the bounding box is formed by Van-L6.

In step (g), a left front vertical line Ver-L3 passing through P1 and a right rear vertical line Ver-L4 passing through P2 are created. Ver-L3 forms the front left vertical line of the bounding box, and Ver-L4 forms the rear right vertical line of the bounding box. Accordingly, in step (g), the intersection point P7 of Ver-L3 and Van-L6 and the intersection point P8 of Ver-L4 and Van-L5 are created. P7 is the front upper left vertex of the bounding box, and P8 is the rear lower right vertex of the bounding box.

In step (h), a six-sided bounding box is created in which P2, P5, P7, and P6 form the top vertices, and P1, P4, P8, and P3 form the bottom vertices.

The actual size and aspect ratio of the bounding box created by reflecting the camera installation location and distance information from the object can be analyzed, and the type and size of the detected object can be estimated accordingly.

2. Object-sensitive vehicle flow control method

[FIG. 2] is a configuration diagram showing a vehicle flow control system to which the object-sensitive vehicle flow control method, which is an embodiment of the present invention, is applied, and [FIG. 3] is an exemplary diagram showing [FIG. 2].

As shown in [FIG. 2] and [FIG. 3], the vehicle flow control system (1) to which the object-sensitive vehicle flow control method (S1) of the present invention is applied includes a camera (5) that photographs each lane group at an intersection. -1), , (5-N), and object analysis information is detected by analyzing the images obtained by shooting each camera 5 using a deep learning-based object analysis algorithm. , a controller 3 that generates an optimal signal system corresponding to the detected object analysis information and transmits a response signal according to the generated optimal signal system to the traffic signal controller 7, which will be described later, and an optimal signal system received from the controller 3. A traffic signal controller (7) that controls the operation of the traffic lights (8) according to the response signal according to the signal system, and an exhibition means installed on the shoulder of the road or a traffic light pole to display a preset warning message under the control of the controller (3) (11) and the control center server (9) that stores and monitors object analysis information, traffic information, violation information, and violation images transmitted from the controller (3), and between the controller (3) and the control center server (9) It consists of a communication network (10) that provides a data movement path.

Cameras 5-1, , (5-N) are installed to photograph each connecting lane connected to the intersection S, and transmit the image acquired by photographing the corresponding connecting lane to the controller 3.

[Figure 4] is a flowchart showing the object-sensitive vehicle flow control method (S1), which is an embodiment of the present invention.

The object-sensitive vehicle flow control method (S1), which is an embodiment of the present invention, optimizes the model of the deep learning algorithm for detecting objects from the input image to suit the road environment and at the same time captures the object using a camera. When acquiring images, the images are analyzed using deep learning to detect objects (vehicles and pedestrians), and then a signal system is established based on the detected object information, significantly increasing the accuracy of object detection and tracking through image analysis. As a result, not only can the reliability of signal control be increased, but vehicle waiting time at intersections is significantly reduced, and real-time image processing is possible in the field.

As shown in [Figure 4], the object-sensitive vehicle flow control method (S1) of the present invention includes a learning step (S10), a camera shooting step (S20), an image transmission step (S30), and an image classification step (S40). ), artificial intelligence-based image analysis step (S50), usage monitoring step (S60), object detailed information generation step (S70), traffic information generation step (S80), number recognition step (S90), violation vehicle determination step (S100) , It can be divided into a signal system establishment stage (S110), a signal control stage (S120), and a classification table optimization stage (S130).

Hereinafter, the “object-sensitive vehicle flow control method” will be divided into steps (A) to (E) and explained together with the attached drawings.

Step (A) is a step where the cameras acquire images by filming the road assigned to them and then transmit the acquired images to the controller. This step (A) includes the learning step (S10) and the camera shooting step. (S20) and video transmission step (S30) are included.

The learning step (S10) is performed every preset period 1 (T1), and is a step in which the controller 3 learns an object analysis algorithm based on deep learning. In addition, the learning step (S10) uses the images collected during cycle 1 (T1) and preset object types to generate learning data that can learn the correlation between images and object types, and uses the generated learning data to create objects. An extraction model, which is a set of parameter values for the correlation between images and object types, is derived.

The camera shooting step (S20) is a step in which the cameras 5 acquire images by photographing the connecting lane assigned to them.

At this time, the cameras 5 can be installed in various ways at the intersection (S) to be able to photograph the connecting lane, which is a preset shooting area. In detail, a separate auxiliary arm is installed at the upper end of the traffic light support that was previously installed. As it is combined through the camera, it is preferable to install it with a shooting angle from the top to the bottom and implement it as a regular PTZ camera capable of PTZ (Pan-Tilt-Zoom) control.

The image transmission step (S30) is a step in which images captured by the cameras 5 are transmitted to the controller 3.

In step (B), the controller uses a deep learning-based object analysis algorithm to analyze the images transmitted through step (A) and performs the above-mentioned "3D bounding box generation method using vanishing points." This is a step of detecting an object and simultaneously detecting object information including at least one of the location, type, and size of the detected object. This step (B) may include an image classification step (S40) and an artificial intelligence-based image analysis step (S50).

[Figure 5] is a flow chart showing the image classification step (S40) in [Figure 4].

As shown in FIG. 7, the image classification step (S40) includes an image input step (S41), a comparison step (S42), a classification mode determination step (S43), a one-to-one based classification step (S44), and a classification table-based classification. It consists of step S45.

The image input step (S41) is a step in which the controller 3 receives images transmitted from the cameras 5 through the video transmission step (S30).

In the comparison step (S42), the controller 3 determines the quantity (N) of images input through the image input step (S41) and the total number (M) of GPUs (Graphic Processing Units) provided in the controller 3. This is the comparison stage.

The classification mode determination step (S43) is performed in the comparison step (S42): 1) If the input image quantity (N) is less than or equal to the GPU quantity (M) (N ≤ M), each image is processed on a single GPU (one-to-one based image 2) If the input image quantity (N) exceeds the GPU quantity (M) (N > M), two or more images must be processed (classification table-based image classification) on at least one GPU. I judge that it does.

The one-to-one based classification step (S44) is performed when the input image quantity (N) is less than the GPU quantity (N) (N ≤ M) in the classification mode determination step (S43), and the images from each camera 5 are Images are classified and input into each GPU to be processed on a single GPU. At this time, when the controller 3 is on a one-to-one basis, GPU information by which images from each camera are classified is preset and stored.

That is, the one-to-one based classification step (S44) inputs the images obtained by shooting with the camera 5 to each GPU on a one-to-one basis.

The classification table-based classification step (S45) is performed when the input image quantity (N) exceeds the GPU quantity (N) (N > M) in the classification mode determination step (S33), and the classification table optimization step (described later) After extracting the classification table optimized by S120) from the memory (M), the images received from the camera 5 are classified into GPUs according to the extracted classification table, and each image is input to the classified GPU. .

At this time, the classification table refers to data matched by the GPU, which is the object of processing the images input from each camera, and is optimized and stored in the memory (M) every preset period 2 (T2) by the classification table optimization step (S130), which will be described later. do.

[Figure 6] is a flow chart showing the artificial intelligence-based image analysis step (S50) of [Figure 4].

As shown in [FIG. 6], the artificial intelligence-based image analysis step (S50) includes image input steps (S51-1, ..., 51-N), algorithm extraction step (S52), and image analysis step (S53). -1, …, S53-N).

The image input steps (S51-1, ..., S51-N) input the video received from the camera 5 to each GPU according to the controller 3's classification by the video classification step (S40) described above. It's a step. That is, each GPU receives an image according to the classification in the image classification step (S40).

The algorithm extraction step (S52) is a step in which the controller 3 extracts the object analysis algorithm learned in the learning step (S10) and stored in the memory (M).

In the image analysis steps (S53-1, ..., S53-N), each GPU uses the object analysis algorithm extracted by the algorithm extraction step (S52) to analyze the image input to itself through the image input step (S51). This is the step of analyzing and outputting object information (recognition, location, type, size, etc.). At this time, object analysis is performed using a 3D bounding box created according to the above-described “3D bounding box generation method using vanishing points.” The controller 3 can detect object information by estimating the type and size of the detected object according to bounding box analysis reflecting the camera installation angle and distance information.

In step (C), the controller tracks the detected objects by referring to the object information detected in step (B), the location information of the preset shooting area of each camera, and the location information of the road, and tracks each object. This is the step of generating object analysis information including at least one of the lane type, lane location, moving speed, and direction in which the is located. This step (C) may include a usage monitoring step (S60) and an object detailed information creation step (S70).

The usage monitoring step (S60) is performed when image processing is performed on at least one of the GPUs through the above-described image analysis steps (S53-1, ..., S53-N), and image analysis is performed on the GPU on which image processing was performed. This is the step of detecting the usage (load) for and storing the detected usage information in memory (M).

At this time, the usage monitoring step (S60) is not performed separately if all of the image analysis steps (S53-1, ..., S53-N) of each GPU are not performed. That is, the usage monitoring step (S60) detects in real time the usage of each GPU due to computational processing for image analysis and then stores it in the memory (M).

In the object detailed information generation step (S70), the controller 3 generates object information detected by the image analysis steps 35-1, ..., 35-M and the preset detection area of each camera 5. By referring to the location information, the location information of the intersection (S) and each connecting lane (S1, S2, S3, S4), and tracking the detected objects, the type of lane in which each object is located, lane location, moving speed and direction, etc. This is the step of generating object analysis information including.

Step (D) is a step in which the controller utilizes and processes the object analysis information generated in step (C) to generate traffic information including at least one of the queue, waiting time, and traffic volume of each connecting lane. . This step (D) may include a traffic information generation step (S80), a number recognition step (S90), and a violation vehicle determination step (S100).

The traffic information generation step (S80) utilizes and processes the object analysis information generated in the object analysis information generation step (S70) to generate traffic information including the queue, waiting time, and traffic volume of each connecting lane. am.

[Figure 7] is an example diagram for explaining the number recognition step (S90) of [Figure 4].

The number recognition step (S90) is performed in parallel with the traffic information generation step (S80) after the object detailed information generation step (S70), and the controller 3 controls the object generated by the object detailed information generation step (S70). By utilizing the analysis information, the current display information of the traffic signal controller 7, and the location information of the intersection (S) and each connecting lane (S1, S2, S3, and S4), the vehicle is connected to the intersection (S) and each connecting lane. This is the step to recognize the license plate number by analyzing the video of the vehicle passing through the connected section.

[FIG. 8] is an example diagram showing a tail biting violation vehicle determined by the violation vehicle determination step 100 of FIG. 4.

The violation vehicle determination step (S100) is performed by the controller (3) using the object analysis information generated by the object detailed information generation step (S70), the current display information of the traffic signal controller (7), the intersection (S), and each connecting lane. This is the step of detecting violation vehicles such as tailgating, as shown in [Figure 8], by utilizing the location information of (S1, S2, S3, S4).

Step (E) is a step in which the controller establishes an optimal signal system based on the traffic information generated in step (D).

This step (E) can be referred to as the signal system establishment step (S110), in which the controller 3 uses a preset optimal signal detection algorithm to generate traffic information (S80). This is the step of analyzing the traffic information generated and establishing an optimal signal system according to the current traffic information.

In step (F), the controller transmits the optimal signal system information established in step (E) to the traffic signal controller, and the traffic signal controller operates the traffic lights according to the optimal signal system information received from the controller. This is the step of controlling the traffic lights. This step (F) may include a signal control step (S120) and a classification table optimization step (S130).

In the signal control step (S120), the controller (3) transmits the optimal signal system information established by the signal system establishment step (S110) to the traffic signal controller (7), and the traffic signal controller (7) transmits the optimal signal system information established by the signal system establishment step (S110) to the traffic signal controller (7). This is the step of managing and controlling the operation of the traffic lights 8 according to the optimal signaling system information received from.

At this time, since the technology and method for the optimal signal detection algorithm that establishes the optimal signal system based on traffic information is a technology and method commonly used in signal control systems, a detailed description will be omitted, and the known algorithm or future development will be omitted. Algorithms can be applied.

[FIG. 9] is a flow chart showing the classification table optimization step (S130) in [FIG. 4].

The classification table optimization step (S130) in [FIG. 9] is executed every preset period 2 (T2).

In addition, the classification table optimization step (S130) includes a data collection step (S131), a total usage calculation step for each GPU (S132), an average usage calculation step for each GPU (S133), a GPU identification number initialization step (S134), and GPU selection. Step (S135), comparison and determination step (S136), inspection candidate determination step (S137), first remaining GPU presence determination step (S138), GPU identification number addition step (S139), second remaining GPU presence determination step It consists of (S140), a step to determine whether an inspection candidate exists (S141), a GPU sorting step (S142), a change target decision step (S143), and a classification table optimization step (S144).

The data collection step (S131) is a step in which the controller 3 collects usage data for each GPU stored in the memory (M) during a preset period 2 (T2).

The step of calculating the total usage of each GPU (S132) is a step of calculating the total usage of each GPU by utilizing the usage data of each GPU collected in the data collection step (S131).

In the step of calculating the average usage of each GPU (S133), assuming that the quantity of GPUs provided in the controller 3 is 'm' and that each GPU is assigned an identification number (i) from 1 to m, the total amount of each GPU is Each calculated by the usage calculation step (S132) Using the total usage and cycle (T2) for each, the average usage per hour for each ( ) is calculated.

The GPU identification number initialization step (S134) is a step of initializing the identification number (i) of the GPU to be inspected to ‘1’.

The GPU selection step (S135) is a step of selecting a GPU with the identification number (i) calculated by the GPU identification number initialization step (S134) or the GPU identification number addition step (S138).

The comparison and determination step (S136) is the average usage of GPUi selected by the GPU selection step (S135) ( ) is compared with the preset upper limit value (TH), and in detail, the average usage of the corresponding GPUi ( ) is more than the upper limit setting value (TH) ( ≥ TH) is the step to compare.

At this time, the upper limit value (TH) is defined as the minimum usage value at which it can be determined that an overload has occurred in the relevant GPU.

In addition, the comparison and determination step (S136) is 1) the average usage of GPUi ( ) is more than the upper limit setting value (TH) ( ≥ TH), it is determined that the GPUi is overloaded and the next step is the inspection candidate decision step (S137), but 2) the average usage of GPUi ( ) is less than the upper limit setting value (TH) ( < TH), it is determined that the GPUi is not overloaded, and the next step is to determine whether the first remaining GPU exists (S138).

In the inspection candidate decision step (S137), the average usage of GPUi ( ) is more than the upper limit setting value (TH) ( ≥ TH), it proceeds when it is determined that the GPUi is overloaded, and this is the step to determine the GPUi as the inspection target.

In the first remaining GPU presence determination step (S138), in the comparison and determination step (S136), the average usage of GPUi ( ) is less than the upper limit setting value (TH) ( < TH), it proceeds when it is determined that the GPUi is not overloaded, and is a step to compare whether the current identification number (i) is greater than or equal to the final identification number 'm'. At this time, the remaining GPU refers to a GPU for which the average usage and upper limit value (TH) have not yet been compared in step 136 (S136).

In addition, the first remaining GPU presence determination step (S138) is 1) if the current identification number (i) is greater than or equal to the final identification number 'm', it is determined that no remaining GPU exists and the next step is the inspection candidate presence determination step ( Proceed with S141), and 2) if the current identification number (i) is less than the final identification number 'm', it is determined that there is a remaining GPU and the next step is to add the GPU identification number (S139).

The GPU identification number adding step (S139) is performed when it is determined that a remaining GPU exists because the current identification number (i) is less than the final identification number 'm' in the first remaining GPU presence determination step (S138). , This is the step of adding '1' to the current GPU identification number (i).

In addition, the GPU identification number addition step (S139) proceeds to the GPU selection step (S135) as the next step and repeats the process of the GPU selection step (S135).

In addition, the second remaining GPU presence determination step (S140) is performed after the inspection candidate determination step (S137) and is a step of comparing whether the current identification number (i) is greater than or equal to the final identification number ‘m’.

In addition, the second remaining GPU presence determination step (S140) is 1) If the current identification number (i) is greater than or equal to the final identification number 'm', it is determined that no remaining GPU exists and the next step is the inspection candidate presence determination step ( Proceed with S141), and 2) if the current identification number (i) is less than the final identification number 'm', it is determined that there is a remaining GPU and the next step is to add the GPU identification number (S139).

The inspection candidate presence determination step (S141) is a step of determining whether at least one GPU has been determined as an inspection candidate by the inspection candidate determination step (S137).

In addition, in the above inspection candidate presence determination step (S141), 1) if the GPU determined as the inspection candidate exists, the GPU sorting step (S142) proceeds to the next step, and 2) if the GPU determined as the inspection candidate does not exist, Quit without proceeding to the next step.

The GPU sorting step (S142) is performed when the GPU determined as an inspection candidate exists in the inspection candidate presence determination step (S141), and the average usage calculated in the average usage calculation step (S133) of each GPU ( ), using the average usage ( ) Sorts the GPUs in order from lowest to highest order.

In the change target determination step (S143), when there are L GPUs determined as inspection candidates, the average usage ( ) L GPUs are determined to be changed, starting from the lowest order.

In the classification table optimization step (S144), for each GPU determined as an inspection candidate in the change target determination step (S143), one of the images input to the GPU of the inspection candidate is selected as the GPU determined to be a change target of the inspection candidate. After deciding to input, the classification table is optimized to reflect the determined information and then stored in memory (M).

In this way, the object-sensitive vehicle flow control method (S1), which is an embodiment of the present invention, analyzes images captured by cameras to generate object information and traffic information, and then generates an optimal display system according to the generated traffic information. Afterwards, traffic signals are configured to be controlled according to the created optimal display system, which not only significantly reduces vehicle congestion rate and waiting time, but also effectively reduces fuel consumption and pollutant gas emissions, thereby reducing social costs. do.

In addition, the object-sensitive vehicle flow control method (S1) of the present invention can dramatically increase the accuracy and reliability of analysis by improving the object recognition rate by analyzing input images using a deep-learning algorithm.

In addition, the object-sensitive vehicle flow control method (S1) of the present invention can further increase the object recognition rate by applying the YOLO model based on a convolution neural network (CNN) as a deep-learning algorithm.

In addition, the object-sensitive vehicle flow control method (S1) of the present invention is configured to limit the recognition target to 5 types frequently seen on the road from the conventional 80 types when learning the deep-learning algorithm, so that the object The recognition rate can be further improved.

In addition, the object-sensitive vehicle flow control method (S1) of the present invention analyzes images using a plurality of GPUs (Graphic Processing Units), enabling real-time processing and analysis of high-capacity images obtained by shooting with cameras.

S1: Object-sensitive vehicle flow control method
S10: Learning stage S20: Camera shooting stage
S30: Video transmission step S40: Video classification step
S41: Video input step S42: Comparison step
S43: Classification mode determination step S44: One-to-one based classification step
S45: Classification table-based classification step S50: Artificial intelligence-based image analysis step
S51-1, , 51-N: Video input step S52: Algorithm extraction step
S53-1, , S53-N: Video analysis steps S60: Usage monitoring steps
S70: Object detailed information creation step S80: Traffic information creation step
S90: Number recognition step S100: Violating vehicle determination step
S110: Signal system establishment step S120: Signal control step
S130: Classification table optimization step S131: Data collection step
S132: Total usage calculation step for each GPU S133: Average usage calculation step for each GPU
S134: GPU identification number initialization step S135: GPU selection step
S136: Comparison and determination step S137: Inspection candidate decision step
S138: First remaining GPU presence determination step
S139: Step of adding GPU identification number S140: Step of determining whether a second remaining GPU exists
S141: Determination step for existence of inspection candidate S142: GPU sorting step
S143: Change target decision step S144: Classification table optimization step

Claims (3)

It is implemented by the artificial intelligence-based image analysis unit of the computer device,
(a) analyzing the camera installation angle and the directionality of the image captured by the camera to derive a longitudinal vanishing point VP1 and a transverse vanishing point VP2;
(b) specifying a moving object in the image to be analyzed;
(c) generating a left lower vanishing line Van-L1 and a right upper vanishing line Van-L2 connected from the VP1 to contact the outer edge of the object;
(d) In the VP2, a front lower vanishing line Van-L3 and a rear upper vanishing line Van-L4 connected to contact the outer edge of the object are created, and the intersection point P1 of the Van-L1 and Van-L3, and the Van-L2 A step in which the intersection P2 of Van-L4 is created;
(e) Create a left rear vertical line Ver-L1 in contact with the leftmost end of the object and a right front vertical line Ver-L2 in contact with the rightmost end of the object, thereby creating an intersection P3 of the Ver-L1 and Van-L1, and the Ver-L2 generating an intersection point P4 of Van-L3, an intersection P5 of Ver-L1 and Van-N4, and an intersection P6 of Ver-L2 and Van-L2;
(f) generating a vanishing line Van-L5 passing from the VP2 to P3 and a vanishing line Van-L6 passing from the VP2 to P6;
(g) Creating a left front vertical line Ver-L3 passing through P1 and a right rear vertical line Ver-L4 passing through P2, creating an intersection point P7 of Ver-L3 and Van-L6, and Ver-L4 and Van-L5. A step in which the intersection point P8 is created; and
(h) generating a hexahedral bounding box in which the P2, P5, P7, and P6 form the upper vertices, and the P1, P4, P8, and P3 form the lower vertices; Including,
How to create a 3D bounding box using vanishing points.
Object sensitive type, which is the operating process of a vehicle flow control system that includes cameras that acquire images by photographing at least one road in the detection area, a traffic signal controller that controls the operation of traffic lights installed in the detection area, and a controller. As a vehicle flow control method,
(A) the cameras acquire images by photographing roads assigned to them, and then transmit the acquired images to the controller;
(B) The controller uses a deep learning-based object analysis algorithm to analyze the images transmitted through step (A) and creates an object according to the 3D bounding box creation method using the vanishing point of paragraph 1. simultaneously detecting object information including at least one of the location, type, and size of the detected object;
(C) The controller tracks the detected object by referring to the object information detected in step (B), the location information of the preset shooting area of each camera, and the location information of the road, and determines the location of each object. Generating object analysis information including at least one of lane type, lane location, moving speed, and direction;
(D) the controller utilizing and processing the object analysis information generated in step (C) to generate traffic information including at least one of queue, waiting time, and traffic volume for each connecting lane;
(E) the controller establishing an optimal signal system based on the traffic information generated in step (D); and
(F) The traffic light is such that the controller transmits the optimal signal system information established in step (E) to the traffic signal controller, and the traffic light controller operates the traffic lights according to the optimal signal system information transmitted from the controller. controlling them; An object-sensitive vehicle flow control method comprising:
In paragraph 2,
In step (B), the object-sensitive vehicle flow control method is characterized in that the controller detects object information by estimating the type and size of the detected object according to bounding box analysis reflecting camera installation angle and distance information.