JP2022074223A - Traffic control device and learning model production method - Google Patents

Abstract

To preliminarily produce a learning model adaptable to a new section of a route.SOLUTION: A traffic control device of an embodiment includes a classification unit that classifies plural sections of plural routes on a road on which a vehicle travels into plural groups on the basis of the property of each section, and a learning processing unit that performs mechanical learning on the basis of previous traffic data for each of the groups so as to produce a learning model of accident forecast which treats an accident occurrence risk.SELECTED DRAWING: Figure 1

Description

Embodiments of the present invention relate to a traffic control device and a learning model generation method.

Conventionally, for example, there is a technique for generating an accident prediction learning model by performing machine learning based on past traffic data for each section of a road on which a vehicle travels. In that case, it is possible to predict an accident (prediction of the likelihood of an accident) using a learning model and current traffic data for each section.

JP-A-2017-151545 Japanese Unexamined Patent Publication No. 2018-18214 Japanese Unexamined Patent Publication No. 2020-135243 Japanese Unexamined Patent Publication No. 2020-61088 Japanese Unexamined Patent Publication No. 2012-59058

However, in the prior art, for example, a section of a new route, a new section when the route is extended, or a section where the traffic condition changes due to an increase or decrease in the number of lanes even in an existing section (hereinafter, they). The section is referred to as the "new section"), which has a problem that it takes a long period of time such as several months to collect the past traffic data necessary for generating the learning model.

Therefore, an object of the embodiment of the present invention is to provide a traffic control device capable of generating a learning model applicable to a new section in advance, and a learning model generation method.

The traffic control device of the embodiment has a classification unit that classifies a plurality of sections of a plurality of routes of a road on which a vehicle travels into a plurality of groups based on the characteristics of the sections, and past traffic data for each of the above groups. It is provided with a learning processing unit that generates a learning model of an accident forecast regarding an accident occurrence risk by performing machine learning based on the machine learning.

FIG. 1 is a block diagram showing a functional configuration and the like of the traffic control device of the first embodiment. FIG. 2 is a schematic diagram showing the relationship between a section that is a unit of accident prediction for a road and a vehicle detector in the first embodiment. FIG. 3 is an explanatory diagram of section classification in the first embodiment. FIG. 4 is a diagram schematically showing QV distribution information in the first embodiment. FIG. 5 is a flowchart showing a learning process of SOM by the traffic control device of the first embodiment. FIG. 6 is a flowchart showing an application process of SOM by the traffic control device of the second embodiment.

Hereinafter, the traffic control device according to the embodiment of the present invention and the learning model generation method will be described with reference to the drawings. In the following, the term "route" includes, but is not limited to, the meaning of a route in the Road Act, and also includes the meaning of a road as a management unit.

In order to facilitate the understanding of the present embodiment, the background technique will be described again. When an accident or traffic jam occurs on a motorway such as a highway, road users become inconvenient because the travel time increases and it becomes necessary to consider detours.

Therefore, techniques for forecasting and forecasting using machine learning (including AI (Artificial Intelligence)) such as accident forecasting (accident prediction) and traffic congestion forecasting are being researched. On the other hand, machine learning requires a certain amount or more of sample data at the time of learning, such as past accident data.

For example, if the expressway has been used for many years, many sample data such as accidents and traffic jams have been accumulated. Therefore, for example, it is possible to generate a learning model for accident prediction by performing machine learning based on past traffic data for each section, and to perform accident prediction using the learning model and current traffic data.

However, depending on the section, past accident data is scarce, so there are sections where learning cannot be sufficiently performed in order to make an accident forecast. Furthermore, for sections where traffic conditions have changed due to an increase or decrease in the number of lanes or the opening of a large shopping center in the neighborhood, even in sections where there is sufficient accident data in existing sections, past traffic data necessary for generating learning models will be collected. However, there was a problem that it took a long time such as several months.

Therefore, in the following, a technique for generating a learning model that can be immediately applied to a new section will be described.

[First Embodiment]
Hereinafter, the first embodiment will be described.
FIG. 1 is a block diagram showing a functional configuration and the like of the traffic control device 3 of the first embodiment.
FIG. 2 is a schematic diagram showing the relationship between a section that is a unit of accident prediction for a road and a vehicle detector in the first embodiment.

The traffic control device 3 shown in FIG. 1 uses traffic data measured by a vehicle detector 1 corresponding to sections 1, 2, 3, ... On the road R shown in FIG. 2, and causes an accident in each section. It is a computer device that makes predictions (prediction of the likelihood of accidents). The white arrow Y in FIG. 2 indicates the flow direction of the vehicle on the road R (hereinafter, also simply referred to as “road”).

The vehicle detector 1 includes a sensor capable of detecting a vehicle traveling on the road R. This sensor is composed of, for example, a loop coil installed under the road surface, a camera or an ultrasonic sensor that monitors the road surface from above, and the like.

Further, the vehicle detector 1 includes a traffic data processing unit. Specifically, the traffic data processing unit is based on the traffic data measured by the sensor, and the traffic volume [vehicles / h] and the average speed of the vehicle traveling on the road R (hereinafter, also simply referred to as "speed"). [Km / h], vehicle density [vehicles / km], occupancy rate (occupancy) [%], etc. are calculated, and the calculation results are transmitted to the traffic control device 3. This calculation and transmission are performed in time units such as 1 minute and 5 minutes. The vehicle detector 1 may transmit only the measurement result by the sensor to the traffic control device 3, and the traffic control device 3 may calculate traffic data such as traffic volume.

Although the traffic control device 3 is shown as one computer device in FIG. 1 for the sake of brevity, it may be realized by a plurality of computer devices. The traffic control device 3 includes a processing unit 31, a storage unit 32, an input unit 33, a display unit 34, and a communication unit 35.

The processing unit 31 includes, for example, a CPU (Central Processing Unit), a ROM (Read Only Memory), and a RAM (Random Access Memory), and executes various processes.

The CPU comprehensively controls the operation of the traffic control device 3. ROM is a storage medium for storing various programs and data. RAM is a storage medium for temporarily storing various programs and rewriting various data.

Then, the CPU executes the program stored in the ROM, the storage unit 32, etc., using the RAM as a work area (work area). The details of the processing unit 31 will be described later.

The storage unit 32 is a storage device such as an HDD (Hard Disk Drive) or SSD (Solid State Drive), and stores various information. The storage unit 32 stores, for example, the accident form data 321 and the classification data 322, and the learning model 323.

The accident form data 321 is form data recording the details of the accident on the road. The accident form data 321 is, for example, the location of the accident, the cause of the accident, the accident situation, the dispatch information of emergency vehicles (emergency vehicle, accident handling vehicle, patrol car, etc.), the time until the accident processing is completed, the lane blockage status, the injured person. Includes various information such as information, the presence or absence of falling objects, and information on whether or not the accident vehicle can run on its own.

The classification data 322 is data of the classification result of a plurality of sections by the classification unit 312.

The learning model 323 is, for example, a learning model for accident prediction using a self-organizing map (SOM) created for each group by the learning processing unit 314 (details will be described later). In general, a self-organizing map is a type of neural network applied to important fields in the real world such as process analysis, control, search system, and information analysis for management, and it is a kind of neural network that inputs high-dimensional input data. , An algorithm for clustering without prior knowledge such as teacher signals (outputs considered ideal for input data). Since the details of the self-organizing map are disclosed in, for example, Japanese Patent Application Laid-Open No. 2014-35639, further description thereof will be omitted here.

Further, the method for creating the learning model 323 is not limited to the method using the self-organizing map described above, and may be a method using a neural network, deep learning, random forest, SVM (support vector machine), or the like. good.

The storage unit 32 also acquired road information such as road sections, number of lanes, interchanges, parking area locations, sensor information acquired from the vehicle detector 1, and weather data management device 2. It stores various information such as weather data and various calculation processing results by the processing unit 31.

The input unit 33 is an input device that receives a user's operation on the traffic control device 3, and is, for example, a keyboard, a mouse, or the like.

The display unit 34 is realized by a liquid crystal display device (LCD (Liquid Crystal Display)), an organic EL (Electro-Luminescence) display device, or the like.

The communication unit 35 is a communication interface for communicating with an external device (vehicle detector 1, weather data management device 2, etc.).

The processing unit 31 includes an acquisition unit 311, a classification unit 312, a determination unit 313, a learning processing unit 314, a prediction processing unit 315, and a control unit 316 as functional configurations.

The acquisition unit 311 acquires various information from the external device and the storage unit 32. The acquisition unit 311 acquires sensor information (traffic data) from the vehicle detector 1, for example. Further, the acquisition unit 311 acquires meteorological data from the meteorological data management device 2.

In addition, the acquisition unit 311 acquires traffic data from the vehicle detector 1 for a new section that is newly targeted for accident prediction.

The classification unit 312 classifies a plurality of sections of a plurality of routes on a road on which a vehicle travels into a plurality of groups based on the characteristics of the sections (hereinafter, also referred to as "grouping"). More specifically, for example, when the classification unit 312 classifies a plurality of sections into a plurality of groups based on the characteristics of the sections, the classification unit 312 is based on the type of route, the type of lane, and the type of section. And classify.

FIG. 3 is an explanatory diagram of section classification in the first embodiment.
First, the lines are classified into radiation, which is a line extending radially outward from the center of a big city, and a circular line, which is a line arranged around the big city and connecting a plurality of radiations. Next, each of the radiation and the ring road is classified into a three-lane section, a two-lane section, and a one-lane section according to the number of lanes.

The three-lane section is further classified into a section without additional lane (additional lane), a section with additional lane, and a section with a shared lane. In addition, each of the sections without additional lanes, sections with additional lanes, and sections with shared lanes is classified into the main line (merging / no divergence), diverging section, merging section, and main line tollhouse. Further, although not shown in FIG. 3, a two-lane section and a one-lane section are also classified in the same manner.

In this way, a plurality of sections of a plurality of routes can be classified into a plurality of groups (groups 1 to 12 in FIG. 3 and the like). Further, the classification unit 312 classifies the accident form data 321 for machine learning into groups. That is, classifying the sections into groups and classifying the accident form data 321 into each group are substantially synonymous.

Returning to FIG. 2, the determination unit 313 for each group, based on the information indicating the relationship between the traffic volume and the speed for each section belonging to the group (for example, QV distribution information), for each section. Determine if it is appropriate to belong to the group.

Here, FIG. 4 is a diagram schematically showing the QV distribution information in the first embodiment. Here, in the QV distribution information, the vertical axis indicates the speed (V), and the horizontal axis indicates the traffic volume (Q: the number of vehicles passing per unit time).

FIG. 4A is QV distribution information in the case of a three-lane section, and FIG. 4B is QV distribution information in the case of a two-lane section. As shown in FIG. 4, since the number of vehicles that can pass through varies greatly between the three-lane section and the two-lane section, there is also a significant difference in the QV distribution information. On the contrary, when comparing a plurality of three-lane sections, their QV distribution information is considered to be similar. Similarly, when comparing a plurality of two-lane sections, their QV distribution information is considered to be similar.

Similarly, for example, for groups 1 to 12 in FIG. 3, the QV distribution information is considered to be similar among a plurality of sections belonging to each group. In addition, it is considered that there is a significant difference in the QV distribution information between the two sections belonging to different groups.

Therefore, the determination unit 313 can determine whether or not it is appropriate for each section to belong to the group for each group by, for example, pattern matching processing of the image corresponding to the QV distribution information. If there is an inappropriate section, for example, the classification unit 312 may perform reclassification.

Returning to FIG. 2, the learning processing unit 314 performs machine learning for each group (groups 1 to 12 in FIG. 3) based on past traffic data, thereby performing machine learning to learn an accident forecast learning model 323 regarding the risk of accident occurrence. (SOM1 to 12 in FIG. 3) are generated. Further, the learning processing unit 314 may further use meteorological data (temperature, road surface temperature, presence / absence of precipitation, amount of precipitation, solar radiation angle, sunshine intensity, etc.) when performing machine learning.

For the new section, the prediction processing unit 315 has a learning model 323 corresponding to the current (most recent) traffic data (for example, speed, traffic volume, occupancy rate) and a group of a plurality of learning models 323 in which the characteristics of the section match. Predict accidents based on. Further, the prediction processing unit 315 may further use meteorological data (temperature, road surface temperature, presence / absence of precipitation, amount of precipitation, solar radiation angle, sunshine intensity, etc.) when predicting an accident.

The control unit 316 executes a process other than the process executed by each unit 311 to 315. The control unit 316 causes the display unit 34 to display, for example, the prediction processing result by the prediction processing unit 315.

FIG. 5 is a flowchart showing a learning process of the learning model 323 by the traffic control device 3 of the first embodiment.
In step S1, the acquisition unit 311 acquires accident data (accident form data 321).

Next, in step S2, the classification unit 312 classifies the accident data by the type of route (see FIG. 3).

Next, in step S3, the classification unit 312 classifies the accident data by the type of lane (see FIG. 3).

Next, in step S4, the classification unit 312 classifies the accident data by the type of section (see FIG. 3).

Next, in step S5, the learning processing unit 314 machine-learns the learning model 323 using the past traffic data, the meteorological data, and the accident data for each group.

As described above, according to the traffic control device 3 of the first embodiment, a plurality of sections of a plurality of routes are classified into a plurality of groups based on the characteristics of the sections, and machine learning is performed for each group based on past traffic data. By performing the above, the learning model 323 of the accident forecast can be generated in advance. That is, it is possible to efficiently learn by learning the learning model 323 with the data of a plurality of sections having similar section characteristics, instead of learning the learning model with the data of each route unit or each point as in the conventional technique. ..

Further, it can be determined whether or not the grouping is appropriate based on the QV distribution information (FIG. 4).

Further, at the time of grouping, highly effective grouping can be performed based on the type of route, the type of lane, and the type of section.

[Second Embodiment]
Hereinafter, the second embodiment will be described. The same matters as in the first embodiment will be omitted as appropriate.

The second embodiment describes a case where the accident forecast function is applied to a newly opened section or the like that is a target of new accident prediction and the accumulation of accident data used for learning is not sufficient.
In the section of the new route or the new section when the route is extended, there is a problem that it takes a long time such as several months to collect the past traffic data necessary for generating the learning model.
Since the configuration of the traffic control device 3 of the second embodiment is the same as that of the first embodiment, the description thereof will be omitted.
The traffic control device 3 in the second embodiment performs the process shown in FIG. 6 below.

FIG. 6 is a flowchart showing an application process of the learning model 323 by the traffic control device 3 of the second embodiment.
In step S11, the acquisition unit 311 acquires the information of the new section. The information of the new section may be input by the user, for example, by the input unit 33, but the information is not limited to this.

Next, in step S12, the prediction processing unit 315 determines the type of the route in the new section.

Next, in step S13, the prediction processing unit 315 determines the type of lane in the new section.

Next, in step S14, the prediction processing unit 315 determines the type of the section of the new section. By the processing up to this point, the group to which the new section belongs is determined.

Next, in step S15, the prediction processing unit 315 applies the learning model 323 corresponding to the group to which the new section belongs. That is, the prediction processing unit 315 makes an accident prediction for the new section based on the traffic data, the corresponding learning model 323, and the meteorological data.

As described above, according to the traffic control device 3 of the present embodiment, a plurality of sections of a plurality of routes are classified into a plurality of groups based on the characteristics of the sections, and machine learning is performed for each group based on past traffic data. Thereby, the learning model 323 applicable to the new section can be generated in advance. In other words, instead of learning the learning model with data for each route or point as in the conventional technique, it is efficient by learning the learning model (learning model 323) with data for multiple sections with similar section characteristics. Can be learned.

Further, it can be determined whether or not the grouping is appropriate based on the QV distribution information (FIG. 4).

Further, at the time of grouping, highly effective grouping can be performed based on the type of route, the type of lane, and the type of section.

Further, when a new section to be the target of accident prediction occurs, the accident prediction can be performed with high accuracy by using the learning model 323 corresponding to the group to which the new section belongs.

The program executed by the CPU of the traffic control device 3 of the present embodiment is a file in an installable format or an executable format, such as a CD-ROM, a flexible disk (FD), a CD-R, or a DVD (Digital Versatile Disk). It may be configured to be recorded and provided on a computer-readable recording medium.

Further, the program may be stored on a computer connected to a network such as the Internet and provided by downloading via the network. Further, the program executed in the present embodiment may be configured to be provided or distributed via a network such as the Internet.

Although the embodiments of the present invention have been described above, the above embodiments are merely examples and are not intended to limit the scope of the invention. The above-described embodiment can be implemented in various forms, and various omissions, replacements, and changes can be made without departing from the gist of the invention. The above-described embodiment and its modifications are included in the scope and gist of the invention, and are also included in the scope of the invention described in the claims and the equivalent scope thereof.

For example, traffic data (speed, traffic volume, occupancy rate, etc.) and meteorological data may be further used for grouping.

1 ... Vehicle detector, 2 ... Meteorological data management device, 3 ... Traffic control device, 31 ... Processing unit, 32 ... Storage unit, 33 ... Input unit, 34 ... Display unit, 35 ... Communication unit, 311 ... Acquisition unit, 312 ... Classification unit, 313 ... Judgment unit, 314 ... Learning processing unit, 315 ... Prediction processing unit, 316 ... Control unit, 321 ... Accident form data, 322 ... Classification data, 323 ... Learning model, R ... Road

Claims (5)

A classification unit that classifies multiple sections of multiple routes on the road on which vehicles travel into multiple groups based on the characteristics of the sections.
A traffic control device including a learning processing unit that generates a learning model of an accident forecast regarding an accident occurrence risk by performing machine learning based on past traffic data for each of the above groups. For each of the above groups, a determination unit for determining whether or not it is appropriate to belong to the relevant group for each of the relevant sections is further provided based on the information indicating the relationship between the traffic volume and the speed of each of the sections belonging to the relevant group. The traffic control device according to claim 1. When classifying the plurality of sections into the plurality of groups based on the characteristics of the sections, the classification unit classifies the plurality of sections based on the type of the route, the type of the lane, and the type of the section. Item 1. The traffic control device according to item 1. An acquisition unit that acquires traffic data from a road information collection terminal for a new section that is a new target of accident prediction or a section where traffic conditions have changed among the plurality of sections.
A claim further comprising the traffic data, the learning model corresponding to the group having the same section characteristics among the plurality of learning models, and a prediction processing unit for predicting an accident based on the new section. The traffic control device according to 1. A classification step that classifies multiple sections of multiple lines of the road on which the vehicle travels into multiple groups based on the characteristics of the sections.
For each of the above groups, a learning process step that generates a learning model of an accident forecast regarding the accident occurrence risk by performing machine learning based on past traffic data, and
Learning model generation method including.

Images