JP2000322682A - System, device and method for predicting traffic flow - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately predict a traffic flow at high speed. SOLUTION: This system is provided with traffic flow measuring means 1 and 2 fixedly installed for measuring traffic parameters at the spots of installation and a predictive means 4 for previously performing learning on the basis of measured data in the past, inputting data from the traffic flow measuring means 1 and 2 and finding the traffic flow parameter at a predictive spot (x) where the traffic flow measuring means 1 and 2 are not installed.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an information collecting technique suitable for a traffic control system and the like.

[0002]

2. Description of the Related Art Conventionally, as a method for predicting a vehicle speed at an arbitrary point between two points on a certain road, a method using an average of vehicle speeds at two points has been used. According to this method, for example, a curve, tunnel,
If there is a factor such as a vehicle on a slope or an outflow / inflow, a large error occurs between the vehicle speed at the actual predicted point and the predicted speed. If all factors such as curves are numerically calculated to solve this problem, a detailed investigation is required for prediction, and the amount of calculation is enormous.

[0003]

As described above, in the prior art, when the calculation amount is reduced, the prediction error of the vehicle speed increases, and conversely, the calculation amount becomes enormous in order to improve the prediction accuracy. There was a problem. Accordingly, it is an object of the present invention to eliminate this problem and to provide a traffic flow prediction system, a traffic flow prediction device, and a traffic flow prediction method with high speed and high accuracy.

[0004]

According to the present invention, there is provided a traffic flow prediction system according to the present invention, which is fixedly installed and measures traffic flow parameters at the installation point;
A learning unit that learns in advance based on past actual measurement data, receives data from the traffic flow measurement unit as input, and obtains a traffic flow parameter at a prediction point where the traffic flow measurement unit is not installed. .

According to a fifth aspect of the present invention, there is provided a traffic flow predicting apparatus for performing learning in advance based on input means for inputting traffic flow parameters by a fixedly installed traffic flow measuring means and past measured data. And a prediction means for obtaining a traffic flow parameter at a prediction point where the traffic flow measurement means is not installed in accordance with the input data from the input means.

According to a sixth aspect of the present invention, there is provided a traffic flow prediction method for measuring a traffic flow parameter at a predetermined measurement point, inputting the measured traffic flow parameter as input, and using a neural network learned in advance based on past measured data. It is characterized in that a traffic flow parameter at an arbitrary prediction point is obtained.

According to the present invention, since the learning is performed in advance based on the past measured data, the traffic flow parameters at the predicted point can be obtained at high speed and with high accuracy. In particular, when a neural network is applied to the prediction means, the amount of calculation after the end of learning becomes extremely small.

[0008]

Embodiments of the present invention will be described below. FIG. 1 is a conceptual diagram showing an embodiment of a traffic flow prediction system according to the present invention. Fixed vehicle sensors 1 and 2 as traffic flow measuring means are installed on the road to detect vehicles passing on the road. In general, there are two types, a single head type and a double head type. The former is for determining the ratio (occupancy ratio) of the vehicle occupying the road, and the latter is for determining the speed of the vehicle. May be applied. Reference numeral 3 denotes a mobile vehicle detector which can be installed at an arbitrary point on a road. The fixed vehicle sensors 1 and 2 and the mobile vehicle sensor 3 measure the traffic flow at each installation point for a predetermined period, and the measured data is stored in the traffic flow estimating device 4 as prediction means in advance. The traffic flow estimation device 4 is connected to the fixed vehicle sensors 1 and 2 by wire or wirelessly, and is provided so that measurement data can be collected in real time. The previously accumulated past measurement data,
The traffic flow at an arbitrary prediction point is predicted from the measurement data collected from the fixed vehicle sensors 1 and 2. More specifically, if the mobile vehicle sensor 3 is installed at the predicted point x and the actual traffic flow is measured and accumulated together with the fixed vehicle sensors 1 and 2 for a predetermined period, the mobile vehicle detection can be performed. Since there is some correlation between the measured data of the detector 3 and the measured data of the fixed vehicle sensors 1 and 2, even if the mobile vehicle sensor 3 is subsequently removed, the fixed vehicle sensor 1 The traffic flow at the predicted point x can be measured from the measurement data 2 and 2.

FIG. 2 shows a conceptual diagram when a neural network is used for the traffic flow estimation device 4. Data measured by the fixed vehicle sensors 1 and 2 is input to the input layer of the neural network, and data measured by the mobile vehicle sensor 3 is used as learning data of the neural network. Specifically, the weight of the connection between the layers is calculated by statistical processing based on measurement data for a predetermined period sufficient for learning of the neural network, learning of the neural network is performed, and the fixed vehicle sensor 1 is added to the learned neural network. 2 to obtain the estimated speed of the vehicle at the predicted point x.

Next, a learning method of the neural network will be described. FIG. 3 is a diagram in which the traffic flow estimation device 4 in the embodiment of FIG. 1 is replaced with a neural network. A vehicle is flowing in a direction from a point a to a point b, and fixed vehicle sensors 1 and 2 are respectively installed at the points a and b, and a movable vehicle sensor 3 is provided at a point x for a predetermined period. Install. In this embodiment, as the learning data of the neural network, data obtained by performing a statistical process on a group of spot velocity data measured by the vehicle sensors 1, 2, and 3 for a predetermined period is used. That is, the fixed vehicle sensor 1,
2 using the speed of points a and b measured by
Is predicted, and the result is compared with a value actually measured by the vehicle sensor 3.

An example in which 14 days of data are held as a speed data group measured by the vehicle sensors 1, 2, and 3 and 7 days are used for evaluation and the remaining 7 days are used for creating learning data. Take. When the vehicle sensors 1, 2, and 3 measure data in a 5-minute cycle, the number of samples of the learning data creation data is 60 (minutes) / 5 (minutes) × 24 (hours) × 7 (days) −1 =
2015 (Sample). Point a at a certain t minute,
The four parameters of the speeds Va (t) and Vb (t) of b and the speeds Va (t-5) and Vb (t-5) of the points a and b at (t-5) are used as input parameters. Prepare a class in which parameters are separated by a certain speed range. In the present embodiment, the speed width is set to 10 km / h, and the upper and lower limits are set to 0 to 1 in consideration of the calculation amount and the prediction accuracy.
30 km / h, and a representative value of each class is set as an intermediate value of each parameter. That is, the number of classes for a certain input parameter V (t) is (130-0) / 10 = 13 classes, and since the number of parameters is 4, the total class is 134. As an example, Va (t) = 52 km / h,
Va (t-5) = 35 km / h, Vb (t) = 40 km / h, Vb (t
-5) = 57 km / h, this data is class (Ca
(t), Ca (t-5), Cb (t), Cb (t-5)) = (55,35,45,55).

As described above, the data groups measured by the vehicle sensors 1 and 2 are represented by Va (t), Va (t-5), Vb (t) and Vb (t-
5) Classification by vehicle detector
Creates a histogram of the data Vx (t) measured by. The upper and lower limits of the histogram are 0 to 130 km / h,
The section width of the histogram is 10 km / h, and the representative value of each section is an intermediate value. In the present embodiment, data obtained by performing statistical processing on the speed data measured in this manner is used as learning data of the neural network. As a result of actually trying it out, as a result of statistically processing the 2015 learning data samples, one or more appearance frequencies were obtained for 134 classes.
It became 192 classes among the classes. Of these, the appearance frequency is 5
The above classes were 79 classes.
Nine classes were used as learning data. As an example, among the data obtained by performing statistical processing on the training data group,
Class (Ca (t), Ca (t-5), Cb (t), Cb (t-5)) = (45,45,3
The histogram in the case of (5, 35) is shown in FIG.

FIG. 5 is a conceptual diagram at the time of learning of a neural network, in which a representative value of an obtained class is used as input layer data of the neural network. However,
When input data is input to the input layer of the neural network, normalization is performed. Various methods can be considered for the normalization method, but this time, normalization is performed by standardization. Here, standardization refers to a case where there are n classes and each class is represented by k variables, and a class Ci is represented by Ci = (Xi1, Xi2,..., Xik ) I = 1... N. At this time, the average μj of the j-th variable xij
And the standard deviation σj are

[0014]

(Equation 1)

[0015]

(Equation 2) And the standardized class is

[0016]

(Equation 3) Becomes

In the above embodiment, only the speed data at the points a and b are used as input parameters. However, by performing normalization, it is possible to input parameters having different dimensions such as traffic volume and occupancy. The present invention can have versatility.

As the teacher layer data, the probability of the histogram in each class is input. That is, the output layer
O1 represents the probability that Vx (t) becomes 0 to 10 km / h, and the output layer O2
Is the probability that Vx (t) is 10 to 20 km / h, ...,
The output layer O13 indicates the probability that Vx (t) will be 120 to 130 km / h.

The learning method of the neural network has been described above. However, the number of samples, the class width, the number and types of input parameters, and the like may be changed. Next, a method of calculating a predicted speed from the value of the output layer is described. First, the average and the standard deviation used when creating the learning data are obtained from the speed data measured by the vehicle sensors 1 and 2 in the input layer. And input to the corresponding input layer. In this embodiment, the output layer
Oi (i = 1, 2,..., 13) is such that Vx (t) is (i-1) × 10
Since the probability of i × 10 km / h is shown, in this embodiment, the speed predicted by the neural network is used.
Vxn (t)

[0020]

(Equation 4) Defined by That is, the product of the output layer value and the representative value is the predicted speed. Specifically, when the output of FIG. 4 is obtained, the predicted speed is Vxn (t) = (0.02 × 5 + 0.15 × 15 + 0.27 × 25 + 0.48 × 35 +
0.07 × 45) /0.99=29.3 (km / h).

FIG. 6 shows an example of the result of speed prediction using the present invention. The solid line is the speed Vx (t) actually measured by the vehicle sensor 3, and the dotted line is the speed Vxn predicted according to the present invention.
(t). In addition, in order to show the effectiveness of the present invention, a broken line indicates a predicted speed Vxa (t) obtained by using the average value of the speed at point t and the point b at point t as the predicted value. Also, Vx
FIG. 7 shows the mean absolute error and mean square error of Vxn (t) and Vxa (t) with respect to (t). 6 and 7, it can be confirmed that the speed prediction according to the present invention is superior to the conventional prediction based on the average value.

The embodiment of the present invention has been described above.
An important point in applying the present invention is a method of creating learning data. If learning data having a large error from the actual one is used, the prediction accuracy is naturally low.

Therefore, as an example of a method of creating learning data, a method as shown in FIG. 8 can be considered. More specifically, speed data is collected by the vehicle sensor 3 for a certain period at the prediction point x before the speed is predicted. At the same time, vehicle sensors 1, 2 at upstream point a and downstream point b
The speed data is collected for a certain period of time. Learning data is created based on the collected speed data, and is applied at the time of speed prediction. In this method, since the actual measurement data of the predicted point x is actually used, the learning data includes factors such as curves, slopes, and tunnels, and highly accurate prediction can be performed without considering the situation of the predicted point. .

As another method, a method as shown in FIG. 9 can be considered. In this method, prediction of learning data is performed by using, as learning data, data of another point x ′ whose route condition is very similar to the predicted point x. Since the data at the points where the route conditions are very similar is used, there is an advantage that the trouble of collecting data for each predicted point can be omitted, while the prediction with a small error can be performed.

In the above embodiment, the information of the upstream point a and the downstream point b of the prediction point x is used as the input parameter as the parameter of the speed prediction. However, the prediction is performed only from the information of the upstream point a, or the information of the downstream point b is used. It is also possible to make predictions only from information.

In the above embodiment, the method of predicting the vehicle speed at an arbitrary point where no vehicle sensor is installed has been described.
As another application, it can also be used as a method for estimating the speed after a certain period of time. That is, in the above-described embodiment, the vehicle speed is predicted at a point "spatially" distant from the point at which the input parameter is sampled. However, it is very easy to turn this into a future vehicle speed prediction distant "temporarily". That is. In addition, changes can be made without departing from the spirit of the present invention.

As described above, in the present embodiment, the use of the neural network can realize high-speed vehicle speed prediction. Further, since learning of the neural network is performed by using measured data as learning data, there is an advantage that highly accurate prediction is possible.

By using the present invention as an example in a traffic control system for a highway, it is possible to predict the speed of a point where a vehicle sensor is not installed. Can be provided to the above driver. Further, it is possible to accurately calculate a travel time required to reach a certain destination from the predicted vehicle speed.

As another application field, the present invention can be used for controlling a traffic light on a general road. That is,
Traffic light control is performed based on the predicted vehicle speed, and smooth traffic can be encouraged.

[0030]

According to the present invention, traffic flow prediction can be performed at high speed and with high accuracy.

[Brief description of the drawings]

FIG. 1 is a conceptual diagram showing one embodiment of the present invention.

FIG. 2 is a conceptual diagram when a neural network is used as a prediction unit.

FIG. 3 shows an example of a prediction process using a neural network.

FIG. 4 is a sample of learning data of a neural network.

FIG. 5 is a conceptual diagram when learning a neural network.

FIG. 6 is an example of a result of predicting a vehicle speed at a point x according to the present invention.

FIG. 7 is an error when a vehicle speed is predicted according to the present invention.

FIG. 8 is an example of a case where actual measurement data of a predicted point in a certain period is used as learning data of the neural network.

FIG. 9 shows an example in which measured data of route sections a ′ and b ′ whose situation is very similar to prediction sections a and b are used as learning data of the neural network.

[Explanation of symbols]

 1-3 Vehicle detector 4 Traffic flow estimation device

Claims (6)

[Claims] Claims: 1. A fixedly installed traffic flow measuring means for measuring a traffic flow parameter at an installation point, and learning in advance based on past actually measured data; A traffic flow prediction system for calculating a traffic flow parameter at a prediction point where the traffic flow measurement means is not installed. 2. The traffic flow prediction system according to claim 1, wherein said prediction means obtains a traffic flow parameter at a prediction point using a neural network. 3. The learning means according to claim 2, wherein said prediction means learns data obtained by said traffic flow measurement means as input layer data of a neural network, and uses data actually measured at a prediction point as learning data of a neural network. The traffic flow prediction system described. 4. The traffic flow prediction system according to claim 1, wherein the traffic flow measurement means is installed at at least one location upstream or downstream of the prediction point. 5. An input means for inputting a traffic flow parameter by a fixedly installed traffic flow measuring means, and learning is performed in advance based on past measurement data, and the learning is performed in accordance with the input data by the input means. A traffic flow predicting device for calculating a traffic flow parameter at a prediction point where the traffic flow measuring means is not installed. 6. A traffic flow parameter at a predetermined measurement point is measured, and the measured traffic flow parameter is input, and a traffic flow parameter at an arbitrary predicted point is obtained by a neural network learned in advance based on past actually measured data. A traffic flow prediction method characterized in that: