KR20220098677A - Traffic accident prediction method and system - Google Patents

Abstract

The present invention relates to a traffic accident prediction system. The system may include: an image filming unit acquiring a real-time image; an object information extraction unit detecting an object from the acquired image and extracting object information from the detected object; and a traffic accident prediction unit predicting a future location and size of the object from the extracted object information and predicting a possibility of an accident based on the predicted future location and size of the object.

Description

Traffic accident prediction method and system

The present invention relates to a method and system for predicting a traffic accident, and more particularly, predicts objects (vehicles and pedestrians, etc.) that are likely to cause an accident based on artificial intelligence using only information from a camera (black box, etc.) attached to a vehicle It relates to a system and method for

Representative ADAS (Advanced Driver Assistance Systems) technologies applied for predicting a traffic accident while driving operate only in a predefined situation, so there is a problem that avoidance control is impossible for an accident that is not predefined.

Accordingly, around 2006, a method of recognizing and avoiding the possibility of collision by transmitting and receiving information between vehicles was proposed (KR 10-2006-0050427). However, in this case, an infrastructure for communication between vehicles must be built, and in the case of a vehicle in which some communication is not made, there are problems such as unpredictable collision and security problems.

Recently, research on predicting vehicle accidents using artificial intelligence technology has been actively conducted. Most of the inventions were made in a way that predicts accidents by analyzing information from the cameras installed in vehicles with artificial intelligence.

In addition, a method for predicting accident probability using CNN has been proposed (KR 10-2017-0157616). However, according to 2016 ACCV “Anticipating Accidents in Dashcam Videos”, it was shown that using RNN outperformed single frame classify. That is, in predicting an accident, it is more effective when analyzed according to the passage of time rather than analyzing only one scene.

Also, around 2017, a method for predicting an accident based on CNN-RNN was proposed, but there is a problem in that it does not determine which vehicle or pedestrian is a risk factor (KR 10-2017-0173141).

In the case of the most recent method around 2019, there is a problem that a large number of additional sensors are required (KR 10-2019-0166389).

[Patent Document 1] Korean Patent Registration No. 10-0754135. 2007.08.24. registration.

The present invention was created to solve the above problems, and an object of the present invention is to provide a method and system for predicting a traffic accident.

Objects of the present invention are not limited to the objects mentioned above, and other objects not mentioned will be clearly understood from the description below.

In order to achieve the above objects, a traffic accident prediction system according to an embodiment of the present invention is disclosed. The system includes an image capture unit that acquires a real-time image, an object information extractor that detects an object from the acquired image, and an object information extractor that extracts object information from the detected object, and predicts and predicts the future location and size of the object from the extracted object information It may include a traffic accident prediction unit that predicts the possibility of an accident based on the future location and size of the object.

Further, according to an embodiment of the present invention, the object information extraction unit may include an object detection unit that detects a vehicle or a pedestrian from an image acquired in real time, and an object motion information extraction unit that extracts motion information of the detected object. have.

In addition, according to an embodiment of the present invention, the object detection unit inputs an image acquired in real time to an object detection deep learning model, and from at least one object, an object identifier, and an object display time from an object-related image acquired in real time. At least one may be detected, and the at least one detected object may be tracked in real time through a tracking algorithm.

In addition, according to an embodiment of the present invention, the object detection unit inputs an image acquired in real time to an object detection deep learning model, and from at least one object, an object identifier, and an object display time from an object-related image acquired in real time. Detect at least one, generate at least one bounding box corresponding to at least one detected object, respectively, and track the detected at least one bounding box in real time through a tracking algorithm (tracking) is possible.

Also, according to an embodiment of the present invention, the object motion information extracting unit may extract motion information using information on the amount of time-series change in coordinates of the generated bounding box.

In addition, according to an embodiment of the present invention, the traffic accident prediction unit predicts the future location and size of the object using the result obtained by inputting the extracted object information into a recurrent neural network (RNN) model. It may include a location and size prediction unit and an accident probability prediction unit for predicting the possibility of a future accident through a result obtained by inputting the future location and size of the predicted object into the accident probability prediction model.

In addition, according to an embodiment of the present invention, the accident probability prediction unit extracts the score of each object from the real-time image by using the previously labeled data, and determines whether the extracted score exceeds a preset threshold in the future. predict the likelihood of an accident.

In addition, in order to achieve the above objects, a traffic accident prediction method according to an embodiment of the present invention is disclosed. The method includes an image capturing step of acquiring a real-time image, an object information extraction step of detecting an object from the acquired image, and extracting object information from the detected object, and predicting a future location of an object from the extracted object information, and predicting the predicted object It may include a traffic accident prediction step of predicting the probability of an accident based on the future location of

Further, according to an embodiment of the present invention, the step of extracting the object information may include the step of detecting an object of detecting a vehicle or a pedestrian from the image acquired in real time and the step of extracting motion information of the detected object.

In addition, according to an embodiment of the present invention, in the object detection step, at least one of an object, an object identifier, and an object display time from an image acquired in real time by inputting an image acquired in real time to an object detection deep learning model It may be a step of detecting one and tracking at least one detected object in real time through a tracking algorithm.

In addition, according to an embodiment of the present invention, in the object detection step, at least one object, an object identifier, and an object display time from an object-related image acquired in real time by inputting an image acquired in real time to an object detection deep learning model detects at least one of, generates at least one bounding box corresponding to at least one detected object, respectively, and tracks the generated at least one bounding box in real time through a tracking algorithm It may be a step of tracking.

In addition, according to an embodiment of the present invention, the step of extracting the motion information of the object may be a step of extracting the motion information using information on the amount of change in which the coordinates of the generated bounding box change temporally. have.

In addition, according to an embodiment of the present invention, the traffic accident prediction step is to predict the future location and size of the object using the result value obtained by inputting object information into a recurrent neural network (RNN) model. It may include an accident probability prediction step of predicting the possibility of a future accident through the result value obtained by inputting the location and size prediction step and the future location and size of the predicted object into the accident probability prediction model.

In addition, according to an embodiment of the present invention, the accident probability prediction step extracts the score of each object from the real-time image using previously labeled data, and determines whether the extracted score exceeds a preset threshold. It may be a step in predicting the possibility of future accidents.

Specific details for achieving the above objects will become clear with reference to the embodiments to be described in detail below in conjunction with the accompanying drawings.

However, the present invention is not limited to the embodiments disclosed below, and may be configured in various different forms, and those of ordinary skill in the art to which the present invention pertains ( Hereinafter, "a person skilled in the art") is provided to fully inform the scope of the invention.

According to an embodiment of the present invention, because an accident is predicted by using only the information of the hidden states of the RNN during the prediction process through the prediction of the future location of the object, the accident is predicted by the accident prediction unit. As it can warn the driver of an accident, it can be used for active accident avoidance operation technology in the autonomous driving field.

The effects of the present invention are not limited to the above-described effects, and potential effects expected by the technical features of the present invention will be clearly understood from the following description.

BRIEF DESCRIPTION OF THE DRAWINGS In order that the above-mentioned features of the present invention may be understood in detail, with a more specific description, with reference to the following examples, some of the embodiments are shown in the accompanying drawings. Also, like reference numerals in the drawings are intended to refer to the same or similar functions throughout the various aspects. It should be noted, however, that the accompanying drawings show only certain typical embodiments of the present invention and are not to be considered as limiting the scope of the present invention, and other embodiments having the same effect may be sufficiently recognized. do it
1 is a block diagram of a traffic accident prediction system according to an embodiment of the present invention.
2 is a diagram illustrating a structure diagram of an artificial intelligence model for predicting traffic accidents according to an embodiment of the present invention.
3 is an exemplary diagram of an object information extraction result according to an embodiment of the present invention.
4 is a diagram illustrating an example of calculating a change amount of a bounding box through a coordinate value of a bounding box according to an embodiment of the present invention.
5 is a diagram illustrating prediction of a traffic accident according to an embodiment of the present invention.
6 is another exemplary diagram of a traffic accident prediction according to an embodiment of the present invention.
7 is a flowchart of a method for predicting a traffic accident according to an embodiment of the present invention.
8 is a flowchart of a traffic accident prediction step according to an embodiment of the present invention.

Since the present invention can have various changes and can have various embodiments, specific embodiments are illustrated in the drawings and described in detail.

Various features of the invention disclosed in the claims may be better understood upon consideration of the drawings and detailed description. The apparatus, methods, preparations, and various embodiments disclosed herein are provided for purposes of illustration. The disclosed structural and functional features are intended to enable those skilled in the art to specifically practice the various embodiments, and are not intended to limit the scope of the invention. The terms and sentences disclosed are for the purpose of easy-to-understand descriptions of various features of the disclosed invention, and are not intended to limit the scope of the invention.

In describing the present invention, if it is determined that a detailed description of a related known technology may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted.

Hereinafter, a method and system for predicting a traffic accident according to an embodiment of the present invention will be described.

1 is a block diagram of a traffic accident prediction system according to an embodiment of the present invention, and FIG. 2 is a diagram illustrating a traffic accident prediction artificial intelligence model structural diagram according to an embodiment of the present invention.

In addition, FIG. 3 is an exemplary diagram of an object information extraction result according to an embodiment of the invention, and FIG. 4 is an exemplary diagram of calculating a change amount of a bounding box through a coordinate value of a bounding box according to an embodiment of the invention is a diagram showing

In addition, FIG. 5 is a diagram illustrating prediction of a traffic accident according to an embodiment of the invention, and FIG. 6 is another diagram illustrating prediction of a traffic accident according to an embodiment of the invention.

1 to 6 , the traffic accident prediction system 10 may include an image capturing unit 100 , an object information extracting unit 300 , and a traffic accident prediction unit 500 .

In an embodiment, the image capturing unit 100 may acquire a real-time image through a camera mounted on a vehicle. For example, when driving or stopping a vehicle, an image may be acquired in real time through a black box camera mounted in the vehicle or a separately attached camera.

In an embodiment, the object information extracting unit 300 may detect an object from an image obtained from the image capturing unit 100 and extract object information from the detected object.

More specifically, the object information extraction unit 300 includes an object detection unit 310 that detects a vehicle or a pedestrian from an image acquired in real time by the image capturing unit 100 and a motion of an object that extracts motion information of the detected object. It may include an information extraction unit.

Also, the object detection unit 310 may detect a vehicle or a pedestrian from the image acquired in real time by the image capturing unit 100 .

More specifically, the object detection unit 310 inputs the image acquired in real time by the image capturing unit 100 into the object detection deep learning model, and at least one object, an object identifier and an object from the object-related image acquired in real time. At least one of the display times may be detected, and the detected at least one object may be tracked in real time through a tracking algorithm.

In addition, the object detection unit 310 inputs the image acquired in real time by the image capture unit 100 to the object detection deep learning model, and at least one object, object identifier, and object display time from the object-related image acquired in real time detects at least one of, generates and displays at least one bounding box corresponding to at least one detected object, and tracks the generated bounding box in real time through a tracking algorithm (tracking) is possible.

In addition, the object detection unit 310 sets a bounding box and tracks it, so that the tracking target becomes clear, and accordingly, the change in the movement of the object can be easily tracked. . In addition, since the size of the bounding box can be adjusted, it is possible to easily express whether the object is moving away or closer by adjusting the size of the bounding box.

Also, the object detector 310 may detect objects in the image using an object detection deep learning model (eg, Mask-RCNN). Thereafter, it may be tracked using the SORT algorithm, which is a tracking technique.

Also, the object motion information extractor may extract motion information by using information on the amount of time-series change in coordinates of the generated bounding box.

In addition, the bounding box change amount may be obtained through a difference in coordinate values of time-series continuous bounding box information. By acquiring motion information of an object through a bounding box change amount, it is possible to obtain motion information of an object without using additional sensors or models, thereby reducing the use of additional models.

More specifically, the object motion information extraction unit may extract motion information using information on the amount of time-series change in the coordinates of a bounding box generated from an image acquired in real time by the image capturing unit 100 . . That is, since the image acquired in real time through the image capturing unit 100 has a fairly large file size, it takes a lot of time to process the image, and since it is difficult to store all the images, the time series of the coordinates of the bounding box Since the size of an acquired image can be reduced through a method of obtaining motion information by extracting change amount information, processor processing speed can be increased and storage capacity can be reduced.

In addition, referring to FIG. 4 , the object motion information extracting unit may obtain the amount of change of the bounding box through the difference in the coordinate values of the bounding box information, and specifically, it may be obtained through Equation 1 below.

Accordingly, the object information extracted through the object information extraction unit 300 may include time-series object position, object size, and object motion information.

In addition, the object information extraction unit 300 is formed of time series (time series) data through the tracking (tracking) algorithm for the location of the object in various time periods extracted with respect to the time series (time series) images. The motion information of each object can be obtained through the amount of coordinate change of each object in time-series continuous images. In addition, the extracted time series object position and object motion information is transmitted to the traffic accident prediction unit 500 .

In one embodiment, the traffic accident prediction unit 500 predicts the future location and size of the object from the object information extracted through the object information extraction unit 300, and predicts the possibility of an accident based on the predicted future location and size of the object. predictable.

In addition, the traffic accident prediction unit 500 consists of an artificial intelligence model, an object future location prediction model, and an accident probability prediction model. The object future location prediction model predicts the future location of each object for a certain period based on information obtained through the object information extraction unit 300 . The object future position prediction model is learned through normal driving images.

In addition, the accident possibility prediction model uses only the information of the hidden states of the RNN in the traffic accident prediction unit 500 during the prediction process based on the future location information obtained through the object future location prediction model to predict the possibility of future accidents. predict The accident probability prediction model is learned based on the information of the accident and non-accident vehicles in the accident video.

More specifically, the traffic accident prediction unit 500 predicts the future position and size of an object that predicts the future position and size of the object using the result obtained by inputting the extracted object information into a recurrent neural network (RNN) model. It may include a unit 510 and an accident probability prediction unit 530 for predicting the possibility of a future accident through the predicted future location and size of the object.

In addition, the traffic accident prediction unit 500 predicts the future location and size of an object according to a time series flow by using an artificial neural network model for object information including the object position, object size, and motion information configured in time series. And, it is possible to predict the possibility of a future accident with the own vehicle based on the predicted future location and size of the object.

In addition, the object future position and size prediction unit 510 may predict the future position and size of the object using the result obtained by inputting the extracted object information into a recurrent neural network (RNN) model.

In addition, the object future location and size prediction unit 510 may predict the future bounding box coordinates of the object through a recurrent neural network (RNN Encoder-Decoder structure) model using the information obtained from the object information extraction unit 300 . . A recurrent neural network (RNN) model can be trained through normal driving images to become future bounding box coordinates.

An artificial neural network according to an embodiment of the present invention is a convolutional neural network (CNN), an auto encoder, a feedforward neural network, a radial basis function network, and Cohen self-organization. It may be composed of a kohonen self-organizing network, a recurrent neural network (RNN), a multilayer perceptron, and the like, and the present invention is not limited thereto and may include any neural network. The above example is merely an example for describing the present disclosure, and the present disclosure is not limited thereto.

In addition, the accident possibility prediction unit 530 may predict the possibility of a future accident through the predicted future location and size of the object.

More specifically, the accident probability prediction unit 530 extracts the score of each object from the real-time image by using the previously labeled data, and determines whether the extracted score exceeds a preset threshold, the likelihood of a future accident can be predicted. The accident probability prediction unit 530 may use a cyclic neural network model learned through the own vehicle accident image.

In addition, the recurrent neural network (RNN) model can be trained so that the output value of the model matches the labeled value using the data labeled 0 for non-accident objects and 1 for future accidents.

In addition, the accident probability prediction unit 530 extracts the score of each object from each frame using the recurrent neural network (RNN) model learned by the above method and predicts future accidents based on a preset threshold (threshold value). can be predicted The threshold value (threshold value) may be selected as an optimal value through sample data.

The present invention is a traffic accident prediction method and system predicting the future location and size of objects, and based on the predicted future information, only the information of the hidden states of the RNN during this prediction process is used in the traffic accident prediction unit to the object We want to predict each future traffic accident.

In addition, the traffic accident prediction method and system can predict the future location with a few frames of each object and use the predicted information to calculate the probability that each object will cause a traffic accident with its own vehicle in the future.

In addition, the traffic accident prediction method and system improve performance in predicting objects with high risk of accidents in the future. Since the use of information does not have a significant effect on learning and judgment, but rather can become a hindrance, performance improvement is attempted by using only object information.

In addition, the traffic accident prediction method and system have a model configured to directly predict the probability of an accident for each object, and learning is made for that purpose.

In addition, since the traffic accident prediction method and system predict an accident by predicting the future location of an object, it is possible to warn the driver about an object with a high probability of an accident, so it can be used for active accident avoidance operation technology in the autonomous driving field. .

7 is a flowchart of a traffic accident prediction method according to an embodiment of the present invention, and FIG. 8 is a flowchart of a traffic accident prediction step according to an embodiment of the present invention.

7 and 8 , the traffic accident prediction method ( S1000 ) includes an image capturing step ( S1100 ) of acquiring a real-time image through a camera mounted on a vehicle, detecting an object from the acquired image, and from the detected object It may include an object information extraction step (S1200) of extracting object information and a traffic accident prediction step (S1300) of predicting the future location of the object from the extracted object information, and predicting the possibility of an accident based on the predicted future location of the object have.

In one embodiment, the image taking step (S1100) may be a step of acquiring a real-time image through a camera mounted on the vehicle. For example, when driving or stopping a vehicle, an image may be acquired in real time through a black box camera mounted in the vehicle or a separately attached camera.

In one embodiment, the step of extracting object information ( S1200 ) may be a step of detecting an object from the image obtained in the image taking step ( S1100 ) and extracting object information from the detected object.

More specifically, the object information extraction step (S1200) may include an object detection step of detecting a vehicle or a pedestrian from the image acquired in real time in the image capturing step (S1100) and a step of extracting motion information of the detected object. .

In addition, the object detection step may detect a vehicle or a pedestrian from the image acquired in real time in the image capturing step (S1100).

More specifically, in the object detection step, at least one object, an object identifier, and an object display time from the object-related image acquired in real time by inputting the image acquired in real time in the image capturing step ( S1100 ) to the object detection deep learning model. It may be a step of detecting at least one and tracking the at least one detected object in real time through a tracking algorithm.

In addition, the object detection step inputs the image acquired in real time to the object detection deep learning model, detects at least one of at least one object, an object identifier, and an object display time from the object-related image acquired in real time, and at least the detected at least one It may be a step of generating and displaying at least one bounding box corresponding to one object, respectively, and real-time tracking of the created bounding box through a tracking algorithm.

In addition, the tracking target becomes clear by setting and tracking a bounding box in the object detection step, and accordingly, it is possible to easily track the change in the movement of the object. In addition, since the size of the bounding box can be adjusted, it is possible to easily express whether the object is moving away or closer by adjusting the size of the bounding box.

In addition, the step of extracting the motion information of the object may be a step of extracting the motion information using information on the amount of change in which the coordinates of the generated bounding box change temporally.

In addition, the bounding box change amount may be obtained through a difference in coordinate values of time-series continuous bounding box information. By acquiring motion information of an object through a bounding box change amount, it is possible to obtain motion information of an object without using additional sensors or models, thereby reducing the use of additional models.

More specifically, the step of extracting object motion information is a step of extracting motion information using information on the amount of time-series change in the coordinates of a bounding box generated from the image acquired in real time in the image capturing step (S1100). can be That is, since the image acquired in real time through the image capturing step (S1100) has a fairly large file size, it takes a lot of time to process the image, and it is difficult to store all the images, so the coordinates of the bounding box are time-series Since the size of an acquired image can be reduced through a method of obtaining motion information by extracting change amount information, processor processing speed can be increased and storage capacity can be reduced.

In addition, in the step of extracting the object motion information, the amount of change of the bounding box may be obtained through the difference in the coordinate values of the information of the bounding box.

Accordingly, the object information extracted through the object information extraction step S1200 may include time-series object position, object size, and movement information of the object.

In one embodiment, the traffic accident prediction step (S1300) may predict the future location of the object from the object information extracted through the object information extraction step (S1200), and predict the possibility of an accident based on the predicted future location of the object.

More specifically, the traffic accident prediction step (S1300) is performed by inputting the object information extracted through the object information extraction step (S1200) into a recurrent neural network (RNN) model and using the result value, the future location of the object and It may include a prediction step (S1310) of the future location and size of the object for predicting the size and an accident possibility prediction step (S1330) of predicting the possibility of a future accident through the predicted future location and size of the object.

In addition, the traffic accident prediction step (S1300) predicts the future location and size of the object according to the time series flow by using an artificial neural network model for object information including the object position, object size, and motion information configured in time series. And, it is possible to predict the possibility of a future accident with the own vehicle based on the predicted future location and size of the object.

In addition, the step of predicting the future location and size of the object ( S1310 ) may be a step of predicting the future location and size of the object using the result obtained by inputting the extracted object information into a recurrent neural network (RNN) model.

In addition, in the object future location and size prediction step ( S1310 ), the future bounding box coordinates of the object may be predicted using the information obtained from the object information extraction unit through a recurrent neural network (RNN encoder-decoder structure) model. The cyclic neural network model can be trained through normal driving images to become the coordinates of the future bounding box.

An artificial neural network according to an embodiment of the present invention is a convolutional neural network (CNN), an auto encoder, a feedforward neural network, a radial basis function network, and Cohen self-organization. It may be composed of a kohonen self-organizing network, a recurrent neural network (RNN), a multilayer perceptron, and the like, and the present invention is not limited thereto and may include any neural network. The above example is merely an example for describing the present disclosure, and the present disclosure is not limited thereto.

In addition, the accident probability prediction step ( S1330 ) may be a step of predicting the future accident possibility through the predicted future location and size of the object.

More specifically, the accident possibility prediction step (S1330) extracts the score of each object from the real-time image using the previously labeled data, and determines the possibility of a future accident by whether the extracted score exceeds a preset threshold value. This may be a predictive step. The accident possibility prediction step ( S1330 ) may use a cyclic neural network model learned through the own vehicle accident image.

In addition, the recurrent neural network (RNN) model can be trained so that the output value of the model matches the labeled value using the data labeled 0 for non-accident objects and 1 for future accidents.

In addition, the accident probability prediction step (S1330) extracts the score of each object from every frame using the recurrent neural network (RNN) model learned by the method, and future accidents based on a preset threshold (threshold value) can be predicted The threshold value (threshold value) may be selected as an optimal value through sample data.

The above description is merely illustrative of the technical idea of the present invention, and various changes and modifications may be made by those skilled in the art without departing from the essential characteristics of the present invention.

Accordingly, the embodiments disclosed in the present specification are not intended to limit the technical spirit of the present invention, but to illustrate, and the scope of the present invention is not limited by these embodiments.

The protection scope of the present invention should be interpreted by the claims, and all technical ideas within the scope equivalent thereto should be understood to be included in the scope of the present invention.

10: Traffic Accident Prediction System
100: video recording unit
300: object information extraction unit
310: object detection unit
500: traffic accident prediction unit
510: object future position and size prediction unit
530: accident probability prediction unit

Claims (14)

In the traffic accident prediction system,
an image capturing unit for acquiring a real-time image;
an object information extraction unit for detecting an object from the obtained image and extracting object information from the detected object; and
Predicting the future location and size of the object from the extracted object information, and a traffic accident prediction unit for predicting the possibility of an accident based on the predicted future location and size of the object,
Traffic accident prediction system.
According to claim 1,
The object information extraction unit,
an object detector for detecting a vehicle or a pedestrian in the image acquired in real time; and
Comprising an object motion information extraction unit for extracting the motion information of the detected object,
Traffic accident prediction system.
3. The method of claim 2,
The object detection unit,
inputting the real-time acquired image into an object detection deep learning model, and detecting at least one of at least one object, an object identifier, and an object display time from the real-time acquired object-related image,
Real-time tracking of the detected at least one object through a tracking algorithm,
Traffic accident prediction system.
4. The method of claim 3,
The object detection unit,
inputting the real-time acquired image into an object detection deep learning model, and detecting at least one of at least one object, an object identifier, and an object display time from the real-time acquired object-related image,
generating at least one bounding box corresponding to each of the at least one detected object,
Real-time tracking of the detected at least one bounding box through a tracking algorithm,
Traffic accident prediction system.
3. The method of claim 2,
The object motion information extraction unit,
Extracting motion information using information on the amount of time-series change in the coordinates of the generated bounding box,
Traffic accident prediction system.
3. The method of claim 2,
The traffic accident prediction unit,
an object future location and size prediction unit for predicting a future location and size of an object by inputting the extracted object information into a recurrent neural network (RNN) model; and
Including an accident probability prediction unit for predicting the possibility of a future accident through the result value obtained by inputting the future location and size of the predicted object into an accident probability prediction model,
Traffic accident prediction system.
7. The method of claim 6,
The accident probability prediction unit,
Extracting the score of each object from a real-time image using previously labeled data, and predicting the possibility of a future accident by whether the extracted score exceeds a preset threshold,
Traffic accident prediction system.
In the method of predicting a traffic accident,
an image capturing step of acquiring a real-time image;
an object information extraction step of detecting an object from the obtained image and extracting object information from the detected object; and
Predicting the future location of the object from the extracted object information, and predicting the possibility of an accident based on the predicted future location of the object, including a traffic accident prediction step,
How to predict a traffic accident.
9. The method of claim 8,
The object information extraction step is
an object detection step of detecting a vehicle or a pedestrian in the image acquired in real time; and
Including the step of extracting the motion information of the detected object,
How to predict a traffic accident.
10. The method of claim 9,
The object detection step is
inputting the real-time acquired image into an object detection deep learning model to detect at least one of an object, an object identifier, and an object display time from the real-time acquired image,
A step of real-time tracking of the detected at least one object through a tracking algorithm,
How to predict a traffic accident.
11. The method of claim 10,
The object detection step is
inputting the real-time acquired image into an object detection deep learning model, and detecting at least one of at least one object, an object identifier, and an object display time from the real-time acquired object-related image,
generating at least one bounding box corresponding to each of the at least one detected object,
A step of real-time tracking of the generated at least one bounding box through a tracking algorithm,
How to predict a traffic accident.
12. The method of claim 11,
The step of extracting the motion information of the object,
A step of extracting motion information by using the change amount information in which the coordinates of the generated bounding box change temporally,
How to predict a traffic accident.
10. The method of claim 9,
The traffic accident prediction step is,
an object future location and size prediction step of predicting a future location and size of an object by inputting the object information into a recurrent neural network (RNN) model; and
Including an accident probability prediction step of predicting the possibility of a future accident through the result value obtained by inputting the future location and size of the predicted object into an accident probability prediction model,
How to predict a traffic accident.
14. The method of claim 13,
The accident probability prediction step is,
Extracting the score of each object from a real-time image using previously labeled data, and predicting the possibility of a future accident by whether the extracted score exceeds a preset threshold value,
How to predict a traffic accident.

Images