JP5083345B2 - Traffic information prediction apparatus, computer program for traffic information prediction, and traffic information prediction method - Google Patents

Description

  The present invention relates to a traffic information prediction apparatus, a computer program for traffic information prediction, and a traffic information prediction method.

As a technology for providing road traffic information to drivers, VICS (Vehicle Information and Communication System: “VICS” is a registered trademark of the foundation) is widely known.
This VICS tabulates traffic information including traffic congestion and link travel time on each route based on fixed point observation information consisting of the number of vehicles, vehicle speed, etc. collected from various roadside sensors (vehicle detectors, loop coils, etc.). The traffic information is provided to the driver by wide-area communication such as narrow-area communication using beacons or FM broadcasting.

As another technique for providing road traffic information to a driver, a traffic information prediction system using a probe car (hereinafter referred to as a probe system) is also known.
For example, as shown in Patent Documents 1 and 2, this probe system uses a vehicle (probe vehicle) that actually travels on a road as a moving body sensor, and wirelessly transmits probe information such as the current vehicle position and time. It is collected from each probe vehicle by communication and generates road traffic information.

Japanese Patent Laid-Open No. 5-151696 Japanese Patent Laying-Open No. 2005-4467

The VICS information can be obtained only with respect to some roads such as main trunk roads where roadside sensors are installed.
On the other hand, in the probe system, since a vehicle (probe vehicle) that actually travels on the road is used as a moving body sensor, traffic information can be acquired even on a road where no roadside sensor is installed.

  However, at present, the number of probe cars is very small, and traffic information cannot be obtained for road links for which information is not obtained from either the VICS or the probe system. Moreover, traffic information at a later time cannot be predicted.

  Accordingly, an object of the present invention is to enable appropriate prediction of road link traffic information.

(1) The present invention is based on the traffic information relevant road links associated with the prediction target road link, a traffic information prediction apparatus for predicting a traffic information of the prediction target road link, it can be a prediction target road link A road link in which traffic information correlates with the road link is set as a related road link, traffic information based on actual measurement values obtained for road links that can be predicted road links, and Learning to learn prediction parameters for predicting traffic information of a prediction target road link, using as a learning data a combination with the traffic information of the related road link acquired before the actual measurement value is acquired And the prediction parameter obtained by learning, the traffic information of the prediction target road link at the predetermined time that comes later The operation of the traffic information of the associated road links that are acquired before the time the data for the calculation, characterized in that a prediction unit that predicts.

Here, for example, when the correlation between the traffic information of the prediction target road link and the traffic information of the related road link is recognized, the traffic information of the prediction target road link at time “t” and the time before time “t” The relationship between the traffic information of the related road link at “t−1” is the traffic information of the road link to be predicted at the time “t + 1” after the time “t” and the traffic of the related road link at the time “t”. It seems to be common with the relationship with information.
Therefore, the learning unit of the present invention provides traffic information based on actual measurement values acquired for road links that can be predicted road links, and traffic of related road links acquired before the actual measurement values are acquired. A prediction parameter is learned using a combination with information as learning data. As a result, the prediction parameter obtained by learning reflects the relationship at the time when the measured value is acquired and before that time.
Then, the prediction unit predicts the traffic information of the prediction target road link at a predetermined time that comes later based on the traffic information of the road link acquired before the predetermined time. By using the parameter for prediction reflecting the relationship in, it is possible to appropriately predict the traffic information of the prediction target road link.

(2) Moreover, it is preferable that the said related road link contains the road link connected with the prediction object road link.

( 3 ) In the learning unit, the combination of the learning data includes traffic information acquired before the actual measurement value is acquired regarding the road link that can be the prediction target road link. It is preferable.
In the same road link, it is considered that the traffic information is correlated even if the time is slightly different. For this reason, the combination of learning data includes traffic information acquired before the actual measurement value is acquired regarding the road link that can be a prediction target road link, and such learning data is used. Thus, if a prediction parameter for predicting traffic information of the prediction target road link is learned, a more appropriate prediction parameter can be obtained.

( 4 ) In the learning unit, the traffic information of the related road link acquired before the actual measurement value included in the combination of the learning data is included in the traffic information before the acquisition. It is preferable that the traffic information at the previous time and the traffic information at a time earlier than the previous time are included.
In this case, learning data, rather than the time of acquisition of the measured values as well traffic information relevant road links in the previous time, the traffic information may include the ones as will the further previous time than, for the associated road link Learning is performed using a prediction parameter that reflects a temporal change in traffic information.

( 5 ) Further, the learning unit is configured to provide traffic information based on an actual value acquired with respect to a road link that can be a prediction target road link, and traffic information on the related road link before ΔT time from when the actual value is acquired. Is used as learning data to learn the prediction parameter, and the prediction unit uses the prediction parameter obtained by the learning, and after ΔT time from the time of obtaining the actual measurement value It can be set as the structure which estimates the traffic information of the prediction object road link in the predetermined time which arrives based on the traffic information of the road link acquired at the time of the acquisition of the said actual value.
In this case, using as a learning data a combination of traffic information based on actual measurement values acquired for road links that can be predicted road links and traffic information of related road links before ΔT time from the acquisition time, It is possible to learn the prediction parameter, and to predict the traffic information of the prediction target road link at a predetermined time that arrives after ΔT time from the time of acquisition of the actual measurement value by using the prediction parameter.

( 6 ) Moreover, it is preferable that the said traffic information prediction apparatus is provided with the random number generation part which calculates | requires the initial value of the said parameter for prediction updated by learning by the said learning part with a random number.
In this case, by using a random number as the initial value of the prediction parameter, each value of the prediction parameter does not have a strong correlation, and the occurrence of a problem due to multicollinearity can be prevented.

( 7 ) When the traffic information based on the actual measurement value is acquired at the predetermined time that arrives later, the traffic information at the predetermined time that arrives after the prediction predicted by the prediction unit is replaced with the traffic information based on the actual measurement value. It can be set as the structure provided with the process part.
In this case, since the measured value is adopted as time passes, more accurate processing is possible.

( 8 ) Moreover, this invention is a computer program for functioning a computer as the traffic information prediction apparatus as described in any one of said (1) to ( 7 ).

(9) Further, the present invention is based on the traffic information relevant road links associated with the prediction target road link, the traffic information of the prediction target road link, a traffic information prediction method a computer to predict, prediction target A road link whose traffic information correlates with a road link that can be a road link is set in the database as a related road link, and is based on an actual measurement value obtained for a road link that can be a prediction target road link. For prediction for predicting traffic information of a prediction target road link using a combination of traffic information and traffic information of the related road link acquired before the actual measurement value is acquired as learning data parameters, the computer learns, by using the prediction parameters obtained by learning, the prediction at a predetermined time comes after Traffic information elephant road link, by calculation of the traffic information of the associated road links that are acquired before the predetermined time has data for calculation, the computer is characterized in that to predict.

  According to the present invention, it is possible to appropriately predict the traffic information of the prediction target road link.

1 is an overall configuration diagram of a traffic information system. It is a block diagram of a traffic information prediction apparatus. It is a figure which shows conversion information. It is a flowchart which shows the traffic information prediction method. It is a block diagram of a neural network. It is explanatory drawing which shows the example of a road link. It is a figure which shows the acquired VICS information and probe information. It is a figure which shows a traffic information database. It is a figure which shows a learning database. It is a figure which shows a weight database. It is a flowchart which shows the procedure of the learning process by a learning system. It is a time chart figure explaining a traffic information prediction method (1st embodiment). It is a time chart figure explaining a traffic information prediction method (1st embodiment). It is a time chart figure explaining a traffic information prediction method (2nd embodiment). It is a time chart figure explaining a traffic information prediction method (2nd embodiment). It is a time chart figure explaining a traffic information prediction method (3rd embodiment). It is a time chart figure explaining a traffic information prediction method (3rd embodiment).

Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.
[1. overall structure]
FIG. 1 is an overall configuration diagram of a traffic information system including a traffic information prediction device (central device) 1 according to the present invention. In addition to the traffic information prediction device 1, this traffic information system includes a probe vehicle 3 equipped with an in-vehicle device 2, a roadside communication device 4 that communicates wirelessly with the in-vehicle device 2, a roadside sensor 5, and the like.

  The traffic information prediction device 1 indicates, for example, one of various functions in the central device, and the traffic information prediction device 1 acquires traffic information (observation information) such as VICS information and probe information, It has a function of generating traffic information for provision for provision. The central device may perform various traffic processes such as traffic signal control and traffic control based on the traffic information.

  The traffic information prediction device (central device) 1 is constituted by a computer having a processing device (CPU) and a storage device, and a computer program for causing the computer to function as the traffic information prediction device 1 is stored in the storage device. It is remembered. This computer program is executed by the processing device, and the processing device performs functions as the traffic information prediction device 1 by inputting and outputting to the storage device and the like. In addition, the function of the traffic information prediction apparatus (central apparatus) 1 demonstrated below is implement | achieved by the said computer program unless there is particular notice.

  The in-vehicle device 2 generates probe information as observation information of the probe vehicle 3 and wirelessly transmits it to the roadside communication device 4. The probe information is traffic information including the position of the probe vehicle, the passage time of the position, the vehicle ID of the probe vehicle, and the like. The probe information may include other information such as the passing speed of the probe vehicle. A link travel time for each road link is obtained based on the position and time information of the probe information. The position of the probe vehicle is calculated based on the GPS signal received by the GPS receiver included in the in-vehicle device 2.

  The roadside communication device 4 transmits / receives information to / from the in-vehicle device 2 by wireless communication. Specifically, the roadside communication device 4 receives probe information as observation information transmitted by the in-vehicle device 2 and transfers the probe information to the traffic information prediction device (central device) 1. Further, the roadside communication device 4 can acquire traffic information for provision to the vehicle from the traffic information prediction device (central device) 1 and transmit the traffic information to the in-vehicle device 2. The roadside communication device 4 and the traffic information prediction apparatus 1 are connected by a communication line.

  The roadside sensor 5 is for detecting traffic information as observation information, for example, a vehicle detector for ultrasonically detecting a vehicle passing directly below, a loop coil for detecting the vehicle by inductance change, or It consists of an image sensor that measures the traffic volume and vehicle speed by image processing of the camera video, and is installed on expressways and major arterial roads for the purpose of measuring the number of vehicles flowing into intersections and vehicle speed. Yes.

  The observation information detected by the roadside sensor 5 is transmitted to the VICS center server 6 via the communication line, and the VICS center server 6 generates VICS information based on the observation information of the roadside sensor 5. This VICS information is transmitted to the traffic information prediction apparatus 1 via a communication line.

The VICS information is traffic information including traffic jams and link travel times on each road link. Since the VICS information is updated by acquiring observation information from the roadside sensor 5 every 5 minutes, high-density information can be obtained in time. However, the roadside sensor 5 is not installed on all roads (20% or less on main roads), and the area coverage rate is low.
On the other hand, since the probe information is acquired from the probe vehicle 3 traveling on the road, the area coverage rate can be increased. However, since the penetration rate of the in-vehicle device 2 for becoming the probe vehicle 3 is still low, only low-density data can be obtained in terms of time.

  That is, when a road link is set on a road in a predetermined area, traffic information (VICS information) is always obtained for each VICS information update unit time for a road link from which VICS information is obtained. On the other hand, as for road links for which VICS information cannot be obtained, there are road links for which probe information can be obtained as traffic information, and road links for which probe information cannot be obtained for each VICS information update unit time. become.

[2. Details of Traffic Information Prediction Device]
[2.1 Overall configuration of prediction device]
The traffic information prediction apparatus 1 according to the present embodiment is a road link from which probe information has been acquired, and road link traffic information at a predetermined time in the future that arrives after the time at which the probe information was acquired (the following information) The embodiment has a function of predicting speed information). Note that the time when the probe information is acquired can be determined based on the passage time included in the probe information.

In this way, by predicting traffic information of road links, even if VICS information and probe information cannot be obtained at a predetermined time that arrives later, highly accurate traffic information is provided to the vehicle (navigation system), It is possible to perform traffic control with high accuracy, and to provide and control traffic in advance.
In the following description, a road link for which the probe information is acquired and that predicts traffic information at a predetermined time that comes later than the time when the probe information is acquired is referred to as a “prediction target road link”.

  As shown in FIG. 2, the traffic information prediction device 1 uses a learning system 12 for learning parameters (prediction parameters) used for predicting traffic information of a prediction target road link, and the parameters. And a prediction system (prediction unit) 11 that predicts traffic information of the prediction target road link.

  The traffic information prediction apparatus 1 includes a traffic information database 13 for storing traffic information and the like, a learning database 14 for storing data used for learning in the learning system 12 among data in the traffic information database 13, A weight database (prediction parameter database) 15 for accumulating prediction parameters (“weight” in the following embodiment) is provided on the storage device.

[2.2 Input information processing section]
The traffic information prediction device 1 according to the present embodiment is an input information processing unit that performs processing on VICS information and probe information acquired by the traffic information prediction device 1 (hereinafter referred to as “input information” when both information are collectively referred to). 16 is provided.
The input information processing unit 16 performs processing for generating “speed information” of each link from the input information. The generated “speed information” is given to the traffic information database 13 to update the traffic information database 13. Used for.

The input information processing unit 16 includes a VICS information processing function that generates “speed information” from the acquired VICS information, and a probe information processing function that generates “speed information” from the acquired probe information.
The VICS information processing function is a function of calculating the link travel speed of the road link from the link travel time obtained based on the VICS information, and converting the link travel speed into “speed information” of the road link.

The VICS information processing function calculates the link travel speed [km / h] by dividing the link length of the road link for which the link travel speed is to be calculated by the link travel time of the road link. The link lengths of all road links in the target area are stored in advance in the storage device of the device 1.
Then, the VICS information processing function converts the obtained link travel speed into an index value (speed index information) called “speed information” (standardized speed). This conversion is performed by referring to “link travel speed-speed information” conversion information set in advance in the device 1. The “link travel speed-speed information” conversion information can be set as shown in FIG. 3, for example.

  The probe information processing function calculates the link travel speed of the road link based on the position and time included in the probe information, and refers to the “link travel speed-speed information” conversion information, thereby calculating the calculated link travel speed. Is converted into an index value (speed index information) called “speed information” (standardized speed). The processing content is the same as that of the VICS information processing function.

[2.3 Processing by the traffic information prediction device]
[2.3.1 Outline of processing]
FIG. 4 shows a traffic information prediction method executed by the traffic information prediction apparatus 1.
Predetermined values (for example, initial values) are set in the traffic information database 13 and the weight database 15 (step S1).
And if the traffic information prediction apparatus 1 acquires VICS information and probe information, the input information processing part 16 will generate speed information from VICS information and probe information (step S2). However, the road links for which speed information can be acquired in step S2 are a part of all road links in the target area, and there are road links for which speed information cannot be obtained in step S2.

Subsequently, the speed information obtained in step S2 is set in the traffic information database 13, and the traffic information database 13 is updated (step S3). When the probe information is acquired, a part of the contents of the traffic information database 13 updated in step S3 is accumulated as a snapshot in the learning database 14 and used as learning data by the learning system 12 (step S4). . The snapshot will be described later.
The learning system 12 learns a prediction parameter for predicting speed information of the prediction target road link using the learning data (step S5). The prediction parameters obtained by learning in step S5 are set in the weight database 15, and the weight database 15 is updated (step S6).

  Then, the prediction system 11 predicts at the time when the probe information is acquired based on the contents of the learning database 14 updated at step S6 and the contents of the traffic information database 13 updated at step S3. The speed information of road links other than the target road link is estimated, the speed information obtained by the estimation is set in the traffic information database 13, and the traffic information database 13 is updated (step S7).

  Then, the prediction system 11 arrives later than the time when the probe information is acquired using the prediction parameters obtained by learning by the learning system 12 (the contents of the learning database 14 updated in step S6). The speed information of the prediction target road link at a predetermined time is predicted based on the contents of the traffic information database 13 updated in step 7 (step S8). The speed information obtained by prediction in step S8 is set in the traffic information database 13, and the traffic information database 13 is updated (step S9).

Then, the speed information of each road link accumulated in the traffic information database 13 updated in step S9 is output to the outside of the apparatus. Specifically, it is displayed on the display of the device 1 or output as information to be provided to the in-vehicle device 2.
[2.3.2 Model for learning processing and prediction of speed information]
FIG. 5 shows a neural network for the prediction system 11 to predict speed information of a prediction target road link, and the learning system 12 executes a process of optimizing (learning) the neural network.
This neural network is configured as a simple perceptron that generates an output value y by multiplying each of N input signals x _i (i: 1 to N) having values of 0 to 1 by weights w _i .

The input signal x _i may be speed information (standardized speed) of a road link other than the prediction target road link, and may be speed information (standardized speed) of the prediction target road link and other road links other than that. Speed information (standardized speed). The output value y is speed information (standardized speed) of the prediction target road link.
The weight x _i is a value indicating how much the speed information of each of the plurality of road links is reflected, and a road link having a high correlation with the prediction target road link (for example, a road adjacent to the prediction target road link). (Link) should be set to a larger value, and a road link having a lower correlation should be set to a smaller value.
Also, in FIG. 5, unlike a general simple perceptron, an independent parameter w ₀ that is added without being multiplied by the input signal x _i is provided.

  As described above, in order to obtain the speed information of the road link (target road link) with the prediction system 11, the speed information of the road link whose correlation with the target road link is recognized to some extent, It is only necessary to obtain a weight indicating how much the speed information is reflected in the speed information of the target road link. The road link speed information is stored in the traffic information database 13, the weights are stored in the weight database 15, and the prediction system 11 acquires necessary information from both the databases 13 and 15 and performs arithmetic processing. Run.

  The neural network is provided for each road link, and the prediction system 11 displays speed information about each of a plurality of road links (target road links), speed information on related road links, and a parameter for prediction indicating the degree of contribution thereof ( It is configured to be predictable using (weight). And the learning system 12 performs the process which optimizes the parameter for prediction of the neural network provided about each road link.

[First embodiment]
[2.3.3 Details of learning process]
Assume a road link connected as shown in FIG. Vx indicates road link speed information (value of 0 to 1). The subscript x in Vx indicates the link number of the road link and takes a value of 1 to 24. That is, in FIG. 6, there are 24 road links with link numbers 1 to 24. Moreover, the arrow direction of each link shows the traveling direction of the vehicle.

  Road links (link numbers x = 1 to 10) indicated by solid arrows are links from which VICS information can be acquired, and are road links corresponding to highways and main trunk roads, for example. Further, road links (link numbers x = 11 to 24) indicated by dotted arrows are links for which VICS information cannot be acquired, and are mainly general roads other than highways and main trunk roads, for example. A road link indicated by a dotted arrow may become a prediction target road link. In other words, among the road links indicated by the dotted arrows, the road link from which the probe information is acquired becomes the “prediction target road link”, and the speed information at a predetermined time that arrives after the time at which the probe information is acquired is The information prediction apparatus 1 predicts.

  In addition, other road links other than the prediction target road link and whose speed information is estimated by the prediction system 11 are referred to as “estimation target road links”. The speed information of the estimation target road link is obtained by the same algorithm as the neural network (FIG. 5) for predicting the speed information of the prediction target road link based on the contents of the learning database 14 and the contents of the traffic information database 13. ,Presumed.

  In the present embodiment, it is assumed that the VICS information and the probe information as shown in FIG. 7 are obtained at a certain time “t” with respect to the road link shown in FIG. In FIG. 7, VICS information (link travel time) for road links with link numbers 1 to 10 is obtained, and probe information for road links with link number 20 is obtained. Neither VICS information nor probe information is obtained for other road links. In FIG. 7, the probe information about the road link with the link number 20 is indicated by “link travel time” in the same manner as the VICS information for easy notation.

The acquired “link travel time” is converted into “speed information” by the input information processing section 16 (see FIG. 2) (step S2 in FIG. 4), and the corresponding road in the traffic information database 13 is determined by the speed information. The “speed information” of the link is updated (step S3 in FIG. 4).
FIG. 8 is a diagram showing the traffic information database updated based on the VICS information and the probe information obtained at the time “t”. In FIG. 8, the road link whose speed information is updated is the link number 20 and the road links of the link numbers 1 to 10, and the speed information of the other road links is not updated, and the time “t” ", The value at the time before (time" t-1 "in the later embodiment) is left as it is.

FIG. 8 will be described. This traffic information database has data items of “speed information” 32, “related road link information” 33, and “link length” 34 for each link number, and the value of each data item for each road link is stored. It can be stored.
Among the data items, “speed information” 32 is an item in which speed information of each road link is set. The “related road link information” 33 indicates a road link (related road link) correlated with each road link, and here indicates a road link (link number) connected to each road link. .

The “related road link information” 33 is divided into five types of “reverse”, “A order”, “A reverse”, “B order”, and “B reverse”.
“Reverse” indicates a link number of a road link opposite to an arbitrary road link. For example, for a road link with link number 20, the road link with link number 17 becomes a “reverse” road link.
"A order" is a road link connected in the forward direction behind a certain road link. For example, for a road link with a link number 20, the road link with a link number 21 is a road link with an "A order". .
“A reverse” is a road link connected in the reverse direction behind a certain road link. For example, for a road link with link number 20, the road link with link number 24 is a road link with “A reverse”. Become.

“B order” is a road link connected in the forward direction in front of a certain road link. For example, for the road link with the link number 20, the road links with the link numbers 14 and 15 are “B order”. It becomes a link.
“B reverse” is a road link connected in the reverse direction in front of a certain road link. For example, for the road link with the link number 20, the road links with the link numbers 16 and 13 are “B reverse”. It becomes a link.

Of the data items, “link length” 34 indicates the link length of each road link. The values of “related road link information” 33 and “link length” 34 are set according to the road configuration of the target area. The values of “related road link information” 33 and “link length” 34 are not updated.
On the other hand, the value of “speed information” 34 is acquired every time traffic information (speed information) including VICS information and probe information is acquired, and the prediction system 11 predicts and estimates traffic information (speed information). Updated every time.

FIG. 9 shows the contents of a learning database in which a plurality of snapshots are accumulated for a certain road link. In this embodiment (FIG. 7), the road link (prediction target road link) from which the probe information has been acquired is a road link with a link number 20, and FIG. 9 shows data about the road link with the link number 20. Yes.
The snapshot is a part of the contents of the updated traffic information database (FIG. 8), and includes speed information (actually measured value) of the road link (prediction target road link) from which the probe information is acquired, and the speed information. This is a combination with speed information of a plurality of related road links (other road links) acquired at the time prior to the acquisition of (actually measured value). The data of this combination becomes learning data.

  Specifically, since the related road links for the road link with the link number 20 are the link numbers 13, 14, 15, 16, 17, 21, 24, the snaps accumulated for the road link with the link number 20 are stored. The shot is a combination of speed information (actually measured value) of the road link of the link number 20 and speed information of the related road link.

A plurality of snapshots are accumulated as learning data in this way (step S4 in FIG. 4), and the learning system 12 is stored in the weight database 15 based on the learning database 14 in which the learning data is accumulated. The weight (prediction parameter) is learned (optimized) (step S5), and the contents of the weight database 15 are updated (step S6).
FIG. 10 is a diagram showing an updated weight database. As shown in FIG. 10, the weight database learned and updated by the learning system 12 is set with the weight wi used when the speed information on each road link is obtained from the speed information on the related road link. The weight wi is set to “number of related road links + 1” for each road link. In FIG. 10, w ₀ is an independent parameter, and w _{1 and} subsequent weights are weights to be multiplied by the related road link speed information.

In the weight database, the weight (prediction parameter) of each road link is preferably set (first condition) so that the sum of the weight values for each link number is 1. That is, when describing the link number 1, it is preferable to set w ₀ = 0, w ₁ + w ₂ + w ₁₃ +... To + w ₉ + w ₁₀ + w ₁₁ = 0.
Moreover, it is preferable to use a random number as the initial value of each weight (prediction parameter) while satisfying the first condition or irrespective of the first condition, and the traffic information predicting apparatus 1 is a functional unit based on a computer program. It is preferable that the initial value of each weight is set by a random number (second condition) by a random number generator 17 (see FIG. 2) provided as By using random numbers in this way, the weight values do not have a strong correlation, and problems due to multicollinearity can be prevented.

The learning process by the learning system 12 will be further described. FIG. 12 is a time chart illustrating the traffic information prediction method of the present invention. The time when the probe information V ₂₀ related to the road link with the link number 20 is acquired is “t” in FIG. 12, the unit time from this time “t”, and the time after the elapse is “t + 1”. The time every time passes is “t + 2, t + 3...”. Further, the time before the unit time from the time “t” is “t−1”, and the time before each unit time is “t−2, t−3. The unit time may be the update unit time. For example, probe information acquired between time “t−1” and time “t” will be described as being acquired at time “t” in FIG. 12.

In addition, at a time before time “t” in FIG. 12, Vx (x = 1 to 24) indicates road link speed information acquired at the time, and the subscript x is a link number (1). To 24). That is, in the state of FIG. 12, at the time “t” when the learning process is started (current), the speed information V ₂₀ based on the actual measurement value is acquired for the road link with the link number 20, and the time “ At time “t−1” before “t”, the speed information Vx (x = 1 to 24) of all road links is in a state of being accumulated in the traffic information database (the state of FIG. 8).

When the traffic information database 13 is updated by acquiring the speed information V ₂₀ for the road link with the link number 20 at time “t” (step S3 in FIG. 4), the learning system 12 The speed information V ₂₀ (actually measured value) acquired for the road link (link number 20) to be a link and the related road acquired at time “t−1” prior to the acquisition of the speed information V ₂₀ A prediction parameter is learned using a combination (snapshot) of the link (link numbers 13, 14, 15, 16, 17, 21, 24) with speed information as learning data (steps S4 and S5 in FIG. 4). ).

FIG. 11 shows a procedure of learning processing by the learning system 12. First, the learning system 12 sets parameters for learning such as an allowable error for error determination, the number of intermediate layers constituting the neural network (FIG. 5), a learning count, and the like (step S11).
The learning system 12 is configured as a neuro engine that optimizes (learns) the neural network shown in FIG. That is, the learning system 12 uses the speed information that is an actual measurement value among the snapshots stored in the learning database 14 as a teacher signal, and from the speed information about the related road link of the road link having the speed information that becomes the teacher signal. Reconstruct an appropriate neural network to output this teacher signal.
That is, the learning system 12 calculates the optimum value of the weight (prediction parameter) from the learning data including the speed information that is the actual measurement value and the speed information about the related road link having the speed information. . The calculation of the optimum value of the weight (prediction parameter) can be performed by an operation for minimizing the residual sum of squares, that is, a least square method or the like, in addition to a method using a neural network.

The optimum value of the weight (prediction parameter) is calculated until the speed information calculated from the speed information about the related road link approaches the teacher signal and the error from the teacher signal becomes less than the set allowable error. Performed (step S13).
When the weights converge and the learning process ends, the obtained weights are reflected in the weight database 15 and the weight database 15 is updated (step S6 in FIG. 4).
When the weight database 15 is updated, the prediction system 11 starts a process of predicting speed information of the prediction target road link based on the updated weight database.

[2.3.4 Details of prediction processing]
In FIG. 12, in the present embodiment, probe information is acquired for the road link with the link number 20 at time “t”, so this road link becomes the prediction target road link. Then, other road links (link numbers 13, 14, 15, 16, 17, 21, and 24) other than the prediction target road link are the estimation target road links.

In FIG. 12, at time “t” when the probe information is acquired for the road link with the link number 20, probe information (speed information based on actual measurement values) is not acquired for the other road links (estimation target road links).
Therefore, the speed information at time “t−1” may be regarded as speed information at time “t” as it is. However, in this embodiment, the prediction system 11 uses the estimation target road link (link numbers 13, 14, 15). , 16, 17, 21, 24) The speed information at each time “t” is learned based on the speed information at time “t−1” and the actual measurement value at time “t” in step S5 of FIG. The estimated weight w _{i (t)} is used for estimation (step S7 in FIG. 4).

That is, the prediction system 11 reads out the speed information at the time “t−1” of the related road link correlated with each of the plurality of estimation target road links from the traffic information database 13 updated in step S3 of FIG. , The weight (prediction parameter) wi _(t) set for each of the estimation target road links (link numbers 13, 14, 15, 16, 17, 21, 24) is updated to the weight database ( 10), and using these, the estimated value of the speed information of each estimation target road link at time “t” is calculated.

Among the estimation target road links (link numbers 13, 14, 15, 16, 17, 21, 24), for example, estimated values (V ₁₃ , V ₁₅ ) of the speed information of the road links of link number 13 and link number 15 are used. According to the neural network shown in FIG. 5, the calculation formula for obtaining is as follows.
V13 _(t) = w0 _(t) + w1 _(t) V15 _(t-1) + w2 _(t) V8 _(t-1) + w3 _(t) V12 _(t-1)
+ W _{4 (t)} V _{7 (t-1)} + w _{5 (t)} V _{1 (t-1)} + w _{6 (t)} V _{11 (t-1)}
+ W _{7 (t)} V _{5 (t-1)} + w _{8 (t)} V _{14 (t-1)} + w _{9 (t)} V _{17 (t-1)}
+ W _{10 (t)} V _{16 (t-1)} + w _{11 (t)} V _{20 (t-1)}
V _{15 (t)} = w _{0 (t)} + w _{1 (t)} V _{13 (t-1)} + w _{2 (t)} V _{20 (t-1)}
+ W _{3 (t)} V _{16 (t-1)} + w _{4 (t)} V _{17 (t-1)} + w _{5 (t)} V _{14 (t-1)}
+ W _{6 (t)} V _{1 (t-1)} + w _{7 (t)} V _{11 (t-1)} + w _{8 (t)} V _{5 (t-1)}
+ W _{9 (t)} V _{8 (t-1)} + w _{10 (t)} V _{12 (t-1)} + w _{11 (t)} V _{5 (t-1)}

  With this calculation formula, the speed information at the time “t” of the estimation target road link (link numbers 13, 14, 15, 16, 17, 21, 24) is acquired as an estimated value, and the road link with the link number 20 The speed information at the time “t” of the (prediction target road link) is acquired as an actual measurement value (see FIG. 13). Then, these speed information values (estimated values) are set in the “speed information” of the traffic information database 13, and the traffic information database 13 is updated (step S7 in FIG. 4).

And the prediction process of the speed information of a prediction object road link is performed as follows.
The prediction system 11 reads the related road link speed information Vx for the prediction target road link (the road link of the link number 20) from the traffic information database 13 updated in step S7, and is set for the prediction target road link. The weight (prediction parameter) w _{i (t)} is read from the weight database 15 updated in FIG. 6, and the predicted value of the speed information of the prediction target road link is calculated using these.

The speed information Vx of the related road link for the prediction target road link (the road link with the link number 20) is V ₁₇ , V ₂₁ , V ₂₄ , V ₁₄ , V ₁₅ , V ₁₆ , V _13. According to the neural network of FIG. 5, an arithmetic expression for obtaining the predicted value V ₂₀ of the speed information is
V _{20 (t + 1)} = w _{0 (t)} + w _{1 (t)} V _{17 (t)} + w _{2 (t)} V _{21 (t)} + w _{3 (t)} V _{24 (t)}
+ W _{4 (t)} V _{14 (t)} + w _{5 (t)} V ₁₅ (t) + w _{6 (t)} V _{16 (t)} + w _{7 (t)} V _{13 (t)}

The weight (prediction parameter) w _{i (t)} is learned from the learning system 12 and acquired from the updated traffic information database 13 (see FIG. 10).
Note that the predicted value V _{20 (t + 1)} of the speed information calculated in FIG. 13 is a predetermined time that arrives later than the time “t” at which the probe information is acquired (time “t + 1” in the present embodiment). whereas a value at), speed information Vx related road link to be used in the arithmetic expression _(t) is the information stored in the traffic information database 13 is updated by the step S7 in FIG. 4, the The speed information of the related road link 33 acquired at a time “t” prior to a predetermined time (time “t + 1”).

  Since there is a correlation between the speed information of the prediction target road link (link number 20) and the speed information of the related road link (link numbers 13, 14, 15, 16, 17, 21, 24), FIG. In FIG. 13, the relationship between the speed information (measured value) of the prediction target road link at time “t” and the speed information of the related road link at time “t−1” is the prediction target road link at time “t + 1”. It is considered that the relationship between the speed information (predicted value) and the speed information of the related road link at time “t” is common.

Therefore, in the present embodiment, as described above, the learning system 12 acquires the speed information (actually measured value) acquired for the road link (link number 20) that becomes the prediction target road link and the speed information (actually measured value). A combination with speed information of related road links (link numbers 13, 14, 15, 16, 17, 21, 24) acquired at time “t−1” before time “t” is used as learning data. Therefore, the prediction parameter obtained by learning reflects the relationship at a different time between the time “t” when the measured value is acquired and the time “t−1” before that. Will be.
Then, the prediction system 11 uses the calculation formula of V ₂₀ =..., As shown in FIG. 13, the prediction target road link (the road link of the road link 20 at the time “t + 1” that arrives later). ) Based on the speed information of the related road links (link numbers 13, 14, 15, 16, 17, 21, 24) acquired at time “t” prior to the time “t + 1”. However, since the prediction parameters reflecting the relationship at different times as described above are used, it is possible to appropriately predict the speed information of the prediction target road link (step S8 in FIG. 4).

Then, the speed information related to the road link with the link number 20 at time “t + 1” is rewritten with the predicted value, and the traffic information database 13 is updated (step S9 in FIG. 4).
When the speed information based on the predicted value is acquired, a part of the contents of the traffic information database updated in step S9, that is, the speed information (predicted value) of the prediction target road link and the speed information (predicted value) A combination with speed information of a plurality of related road links at time “t” before “t + 1” at the time of acquisition is accumulated as a snapshot in the learning database 14 and used as learning data by the learning system 12 (step S4). ).

As described above, the predicted value at time “t + 2” can be obtained by executing step S4 to step S9 in FIG. 4, and further, the time “t + 3” can be obtained by repeatedly executing step S4 to step S9. Subsequent predicted values can also be obtained. That is, the speed information at the previous predetermined time can be predicted and estimated one after another.
Moreover, the process of step S3-step S9 can be repeatedly performed whenever the speed information based on a measured value is acquired by step S2.

As described above, the prediction system 11 uses the speed information at the predetermined time (time “t + 1” in the present embodiment) to arrive later as the predicted value (V ₂₀ ^E _{(t + 1)} ) for the road link with the link number 20. The predicted value (V ₂₀ ^E _{(t + 1)} ) is acquired and accumulated in the traffic information database 13.
Then, it is assumed that time elapses from the execution of this process, and the current time becomes the time t + 1. At that time, for the road link with the link number 20, an actual measurement value is acquired as, for example, probe information. In this case, the speed information of the road link with the link number 20 on the traffic information database 13 is obtained from the speed information (V ₂₀ ^E _{(t + 1)} ) at time t + 1 acquired by the prediction system 11 as the predicted value. Is replaced with speed information (V ₂₀ ^A _{(t + 1)} ). That is, at the current time t + 1, the speed information of the road link with the link number 20 at the time t + 1 is not the predicted value (V ₂₀ ^E _{(t + 1)} ) but the actually measured value (V ₂₀ ^A _{(t + 1)} ). Thus, only the speed information (V ₂₀ ^A _{(t + 1)} ) based on this actually measured value is newly accumulated in the traffic information database 13 and updated. The update process of the traffic information database 13 is executed by the function of the processing unit 18 provided in the prediction system 11.

At this time t + 1, the speed information of the road link with the link number 20 used for learning and prediction by the learning system 12 and the prediction system 11 is speed information (V ₂₀ ^A _{( t + 1)} ).
Furthermore, the speed information of the road link with the link number 20 that is used for learning and prediction by the learning system 12 and the prediction system 11 at the time after the time t + 1 after the time t + 1 (for example, the time t + 2) is already in the past. Speed information (V ₂₀ ^A _{(t + 1)} ) based on the actually measured value at time t + 1.

[Second Embodiment]
In the first embodiment (FIG. 12), the time “t” is obtained by using the speed information at the time “t−1” one unit time before the time “t” at which the speed information based on the actual measurement value is acquired. The learning parameter was learned by reflecting the relationship between and the time “t−1”.
In the second embodiment, as shown in FIG. 14, the speed information at time “t−1” one unit time before the time “t” at which the speed information based on the actual measurement value is acquired, and Using the speed information at time “t-2” prior to time “t−1”, the learning parameters are reflected by reflecting the relationship between time “t” and time “t-1” “t-2”. The case of learning will be described.

Also in this second embodiment, the learning system 12 executes a process of optimizing (learning) the neural network of FIG. 5 using learning data, and the prediction system 11 uses the neural network to predict the road to be predicted. Predict link speed information.
Similarly to the first embodiment, a case will be described in which the speed information V20 for the road link with the link number ₂₀ is acquired at time “t”.

  In this case, as in the first embodiment, the learning system 12 includes the “speed information based on the actual measurement value acquired for the road link (link number 20) that becomes the prediction target road link” and the “measurement value acquisition time”. A prediction parameter is learned using a combination (snapshot) with the related road link speed information at a time prior to “t” as learning data. The combination of the learning data (snapshot) The included “speed information of the related road link at the time prior to the actual measurement value acquisition time“ t ”” includes speed information at a time “t−1” prior to the acquisition time “t”; Speed information at a time “t-2” that is further before the previous time “t-1” is included.

That is, the related road links (link numbers 13, 14, 15, 16, ₁₇ at time “t−1” and time “t-2” before the time “t” when the speed information (actual measurement value) V ₂₀ is acquired. , 21, 24) is used to learn the prediction parameters using the learning data including the speed information.
In this case, the number of prediction parameters increases in the same manner as the number of speed information, but the optimal value of the prediction parameter is calculated from the learning data, as in the first embodiment. The calculation of the optimum value of the prediction parameter can be performed by, for example, an operation that minimizes the residual sum of squares, that is, a least square method or the like.

  In the case of this second embodiment, the learning data includes not only the speed information of the related road link at the time “t−1”, which is one unit time before the actual measurement value “t”, but more than that. Further, the speed information at the previous time “t-g (g = 2, 3, 4...)” Is also included, and the prediction parameter reflecting the temporal change of the speed information related to the related road link in the past. Learning is performed using.

Then, similarly to the first embodiment, the prediction system 11 estimates the speed information of the related road link at time “t” and uses the calculation formula of V ₂₀ =. As shown in FIG. 4, the speed information of the prediction target road link (the road link of the road link 20) at the time “t + 1” that arrives later is set to the time “t” before the time “t + 1” and the time before that. Prediction is performed based on the speed information of the related road link acquired at time “t−1”. At this time, since the prediction can be performed using the prediction parameter reflecting the temporal change in the past as described above, the speed information of the prediction target road link can be predicted more appropriately.

[Third embodiment]
In the first and second embodiments, the case has been described in which the speed information at the time “t + 1” after one unit time is predicted from the time “t” at which the speed information based on the actual measurement value is acquired.
In the third embodiment, as shown in FIGS. 16 and 17, a time “t + n (n = 2, 3, 3) later than the time“ t ”at which the speed information based on the actually measured value is acquired. 4 ...) ”will be described.
In the present embodiment, a case will be described in which speed information of a prediction target road link is predicted at a time “t + 3” that is two unit hours after a time “t” at which speed information based on an actual measurement value is acquired.

Also in the third embodiment, the learning system 12 executes a process of optimizing (learning) the neural network of FIG. 5 using learning data, and the prediction system 11 uses the neural network to predict the road to be predicted. Predict link speed information.
Similarly to the first embodiment, a case will be described in which the speed information V20 for the road link with the link number ₂₀ is acquired at time “t”.

In this case, in FIG. 16, the learning system 12 uses the speed information based on the actual measurement value acquired for the road link (link number 20) that becomes the prediction target road link, and the ΔT time from the time “t” when the actual measurement value is acquired. The prediction parameter is learned using the combination with the speed information of the previous related road link as learning data.
That is, the “ΔT time” corresponds to 3 unit times (n = 3) in the present embodiment, and the learning system 12 uses the speed information (measured value) acquired at time “t” and the time A combination with “t−n (n = 3)”, that is, speed information of the related road link at time “t−3” is used as learning data.

Then, the prediction system 11 uses the prediction parameter obtained by learning to predict speed information of the prediction target road link at a predetermined time that arrives after the time “t” from when the actual measurement value is acquired. .
That is, in FIG. 17, since the “ΔT time” corresponds to 3 unit times (n = 3) in the present embodiment, the prediction system 11 is time “t” when the measured value is acquired. The speed information at time “t + n (n = 3)” that comes later than that, that is, time “t + 3” is predicted.

  Then, the prediction system 11 uses the speed information used to predict the speed information at time “t + 3” as the speed information of the related road link acquired at the time “t” when the actual measurement value is acquired, that is, the time of the predicted value. The speed information of the related road link acquired at time “t” before Δt time (3 unit times (n = 3)) before “t + 3”.

Furthermore, in the third embodiment, similarly to the second embodiment, the prediction is performed using a prediction parameter that reflects a temporal change in the past.
That is, in the learning process by the learning system 12, the “related road link speed information acquired at a time before“ t ”at the time of actual measurement value acquisition” included in the combination (snapshot) of the learning data is included. Includes speed information at time “t-3” prior to the acquisition time “t” and speed information at time “t-4” prior to the previous time “t-3”. ing.
In the prediction process by the prediction system 11, the speed information of the prediction target road link at the time “t + 3” is acquired at the time “t” before the time “t + 3” and the time “t−1” before the time “t + 3”. Predict based on the speed information of the relevant road links.

In order to obtain the predicted value, the speed information of the related road link at the time “t” and the time “t−1” is necessary. For these speed information, the speed information at the time “t-3” is used as it is. Although it may be regarded as speed information at time “t” and time “t−1”, the measured value is acquired at time “t” in the third embodiment as well as in the first embodiment. Thus, the speed information estimated using the weight learned based on the actual measurement value can be obtained.
Thus, according to the third embodiment, it is possible to predict speed information in the future after ΔT time that is 2 unit time or more after speed information based on actual measurement values is acquired.

In each of the embodiments, the input (input signal x _i ) of the neural network (FIG. 5) for the prediction system 11 to predict the speed information of the prediction target road link is used as another road link other than the prediction target road link. Speed information. That is, the related road link of the prediction target road link does not include self (prediction target road link).

However, the self (prediction target road link) may be included in the related road link of the prediction target road link. That is, the input (input signal x _i ) of the neural network (FIG. 5) may be the speed information of the prediction target road link and the speed information of other road links.
In other words, the combination of the learning data may include speed information regarding a road link that can be a prediction target road link at a time before “t” when the actual measurement value is acquired.
In the case of FIG. 12, in the first embodiment, the speed information Vx of the related road link for the prediction target road link (link number 20) is the speed surrounded by a double line at time “t−1”. Information (V ₁₃ , V ₁₄ , V ₁₅ , V ₁₆ , V ₁₇ , V ₂₁ , V ₂₄ ) is added to the speed information of the prediction target road link (link number 20) at time “t−1”. (V ₂₀ ) is included.

In this case, the prediction system 11 uses the prediction parameter obtained by learning to obtain the speed information of the prediction target road link at the time “t + 1” that comes later, before the time “t + 1”. The speed information (V ₁₃ , V ₁₄ , V ₁₅ , V ₁₆ , V ₁₇ , V ₂₁ , V ₂₄ ) acquired at time “t” and the time “t + 1” before the time “t + 1”. Prediction is performed based on the speed information (V ₂₀ ) of the road link that becomes the prediction target road link acquired as an actual measurement value at “t”.

Thus, in order to predict speed information in the future of the prediction target road link, the significance in the case of learning including speed information related to the prediction target road link and predicting including speed information related to the prediction target road link is This is because, in the same road link (prediction target road link), it is considered that a correlation is recognized in the speed information even if the time is slightly different.
Therefore, if the combination of the learning data includes speed information regarding the road link that can be the prediction target road link, and learning the prediction parameter using such learning data, a more appropriate prediction parameter can be obtained. It is possible to predict speed information with higher accuracy.
As described above, when predicting future speed information of the prediction target road link, if the prediction target road link is included in the learning process and the prediction process, there is a strong correlation in the speed information. The problem of collinearity may occur. Therefore, in order to predict the future speed information of the prediction target road link, it is sufficient to learn without including the speed information about the prediction target road link, and to predict without including the speed information about the prediction target road link. In such a case, the occurrence of the above-described problem can be prevented.

  The embodiment disclosed this time should be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the meanings described above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

For example, other road links (related road links) used for predicting speed information of the prediction target road link are not limited to those connected to the prediction target road link, but are separated from the prediction target road link. However, it may be a road link (for example, a road link parallel to the estimation target road link) whose speed information changes in time.
Further, the traffic information processed by the traffic information estimation apparatus 1 of the present embodiment is not limited to speed information, but can be other traffic information such as traffic jam information in addition to / in place of the speed information.
The speed information is not limited to the standardized information, and the link travel speed itself may be used.

  1: traffic information prediction device, 11: prediction system (prediction unit), 12: learning system (learning unit), 13: traffic information database, 14: learning database, 15: weight database, 17: random number generation unit, 18: processing Part

Claims (9)

Based on traffic information associated road links associated with the prediction target road link, a traffic information prediction apparatus for predicting a traffic information of the prediction target road link,
A database in which road links that have a correlation in traffic information with road links that can be predicted road links are set as related road links;
As a learning data, a combination of traffic information based on an actual measurement value acquired for a road link that can be a prediction target road link and traffic information of the related road link acquired before the acquisition of the actual measurement value A learning unit that learns a prediction parameter for predicting traffic information of a road link to be predicted,
Using the prediction parameters obtained by learning, the traffic information of the prediction target road link at a predetermined time that arrives later is used to calculate the traffic information of the related road link acquired before the predetermined time . Prediction unit that predicts by data calculation ,
A traffic information prediction apparatus comprising: The traffic information prediction apparatus according to claim 1, wherein the related road link includes a road link connected to a prediction target road link. Wherein the combination of the learning data, the prediction with respect to a road link can be a target road link, the measured value according to claim 1 or 2 which also includes the traffic information which is acquired before the time of acquisition Traffic information prediction device. In the traffic information of the related road link acquired before the acquisition of the actual measurement value included in the combination of the learning data,
The traffic information prediction device according to any one of claims 1 to 3 , wherein the traffic information at a time before the acquisition time and the traffic information at a time before the previous time are included. The learning unit learns a combination of traffic information based on an actual value acquired for a road link that can be a prediction target road link and traffic information of the related road link before ΔT time from when the actual value is acquired. Using the data for learning, learning the parameter for prediction,
The prediction unit obtains the traffic information of the prediction target road link at a predetermined time that arrives after the ΔT time from the time of acquisition of the actual measurement value, using the prediction parameter obtained by learning. The traffic information prediction device according to any one of claims 1 to 4 , wherein prediction is performed based on traffic information of a road link that is sometimes acquired. The initial value of the prediction parameter which is updated by being learned by the learning section, the traffic information prediction apparatus according to any one of claims 1 to 5 and a random number generator for determining the random number. When traffic information based on an actual measurement value is acquired at a predetermined time that arrives later, a processing unit that replaces the traffic information at a predetermined time that arrives later predicted by the prediction unit with traffic information based on the actual measurement value is provided The traffic information prediction device according to any one of claims 1 to 6 . The computer program for functioning a computer as the traffic information prediction apparatus as described in any one of Claims 1-7 . Based on traffic information associated road links associated with the prediction target road link, the traffic information of the prediction target road link, a traffic information prediction method a computer to predict,
Road links that have a correlation in traffic information with road links that can be predicted road links are set in the database as related road links.
As a learning data, a combination of traffic information based on an actual measurement value acquired for a road link that can be a prediction target road link and traffic information of the related road link acquired before the acquisition of the actual measurement value The computer learns parameters for prediction for predicting traffic information of the prediction target road link,
Using the prediction parameters obtained by learning, the traffic information of the prediction target road link at a predetermined time that arrives later is used to calculate the traffic information of the related road link acquired before the predetermined time . A traffic information prediction method , wherein the computer predicts by calculation using data .

Images