JP2973662B2 - Travel time estimation and prediction device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a travel time estimation / prediction apparatus used for providing travel time information to a driver of a traveling vehicle in an arbitrary section and controlling, managing or operating road traffic.

[0002]

2. Description of the Related Art Conventionally, this type of device is mainly an individual vehicle identification device (hereinafter referred to as an AVI) alone or a device based on an estimation / prediction means using a measured value from a vehicle sensor. . In the case of the former , statistical estimation models are the mainstream for estimation and prediction means.
It is. In the latter case, the travel time is predicted by a model describing a means for performing an estimation prediction from the measured value, the traffic volume, the occupancy or the speed as the traffic flow at the sensor installation point. As the specific means, a statistical model that extracts and uses a time-series statistical relationship from the past and a mathematical model that models the motion of a vehicle are mainly used. Further, the device is generally configured using a single model among these.

[0003]

In general, in an AVI-only device, the actual measured travel time of a vehicle can be used. Information at two points is needed.
Furthermore, when AVIs are densely installed in a wide area, it is necessary to check information at all installation points. Further, when it is necessary to capture travel time in a large number of sections, the travel time becomes large and the cost becomes high. In addition, the estimation / prediction means is based on past actual results. However, auxiliary means is necessary especially when recognition cannot be made at all due to bad weather or the like.

On the other hand, in the case of using the vehicle sensor and the estimation / prediction means, the apparatus is relatively inexpensive, but the actual travel time cannot be directly obtained. Since roads have different characteristics depending on their social roles and usage patterns, traffic conditions have various aspects, and travel time varies depending on each road and other conditions. On the other hand, traffic conditions on the road can be grasped with high accuracy by densely installing sensors and collecting a large number of measurement values, and it is thought that travel time can be derived by this,
Installing sensors all over the road is costly and environmentally unreasonable, and must be decentralized. Therefore, a model suitable for the target is required as the estimation and prediction means.

[0005] Since these models are constructed based on the theory under suitable conditions such as certain assumptions, sufficient results can be obtained if the theory matches the object to be processed. However, the present invention is not limited to the actual traffic phenomenon and the theory is always the same situation, even if the exact model under certain circumstances, under other circumstances might be a case of non-precise, of one whole It is difficult to cover with a model. On the other hand, a large number of models exist and are constructed in consideration of various viewpoints and complicated situations. As described above, when there are heterogeneous models or a plurality of models, it is difficult to select and apply the models.
Conventionally, the application of a single model selected in advance has been mostly used. Even in this case, it is necessary to verify the model, and it is necessary to periodically compare the model with the travel time of the actual object and correct the model in order to cope with the aging of the object.

In general, these estimation and prediction means need to be constructed by statistically or analytically deriving a law-based relationship based on previously collected measurement values and observed phenomena. However, travel time also depends on the autonomous factors of the vehicle and the influence of the surrounding environment, so the measured values and phenomena tend to be incomplete and unclear, and the construction work requires very advanced and specialized skills. Had become.

The present invention solves such a problem. The present invention estimates and predicts travel time with high accuracy from various viewpoints, reflects the travel time of an actual object, and can also be applied to characteristics and changes of the object. It is an object of the present invention to provide an excellent travel time estimation and prediction device that can flexibly respond.

[0008]

The present invention achieves the above object.
Vehicle sensors and individual vehicle identification within the section
Vehicle speed, travel time and travel speed measured by the sensor
Based on the degree, the travel time in the section is estimated or
Is provided with an estimating means for estimating .

In addition to the above, a plurality of estimated prediction models
From the estimated prediction means that stores the
Based on the past traffic conditions and the traffic conditions
It is provided with a matching means for comparing and selecting an optimal estimation prediction model.

Further, in addition to the above, a vehicle detector in a section
Based on the data from
The evaluation means is provided for determining whether to use the individual travel time estimation prediction result of the file or to combine a plurality of results and use the result as an overall result.

In addition to the above, a travel time meter in a section is provided.
Measured value and estimated predicted value and predetermined estimated predicted model
Based on the estimated data,
An adjusting means for changing the evaluation means is provided.

[0012]

According to the present invention, the following functions are provided.

Since the entire estimation and prediction means is constituted by the logical sum of the models of different types or subspaces as the estimation and prediction means, it is not necessary to cover all areas with a single model, and it is not necessary to cover traffic conditions, roads and the like. Models can be easily constructed according to various conditions.

In addition, when the model is applied, the matching condition is compared with the target situation at an arbitrary point in time, and the model is selected or matched. Single or multiple models can be selected or viewed and dynamically applied.

Further, the result of the estimation / prediction means is further evaluated using the result of the matching means, and selection or integration is performed to obtain an overall result. Therefore, an estimation / prediction result matching the matching with the matching condition can be obtained.

In addition, since the result of the model using the measured value of the vehicle sensor in the estimation and prediction means is adjusted so as to be adapted to the result of the measurement of the actual travel time by AVI, both results can be effectively utilized. for adjusting the means at an arbitrary timing, one or all of the estimated while reflecting the actual object, evaluation, and adjustment means, it is possible to change the applicable adjusted for each unit that reconstructs the entire target Can be flexibly adapted to

Therefore, if a plurality of estimated prediction means can be used, matching and evaluation means can be applied according to the traffic situation, and if AVI data can be used, the evaluation means uses this and the estimated prediction result. Make adjustments.

[0018]

FIG. 1A is a diagram showing the configuration of an embodiment of the present invention. In FIG. 1, reference numeral 1 is a simple schematic diagram of a road, and only the case where the vehicle travels from the left hand to the right hand in FIG. 1 will be considered. That is, in FIG. 1, the left hand side is upstream, and the right hand side is downstream. Reference numeral 2 denotes a vehicle detector, which is installed at intervals of several hundred meters on the road. Vehicle detector 2 is generally used for traffic volume,
The occupancy or the speed is measured. In this embodiment, the speed can be measured. 3 is an individual vehicle recognition device (Auto
In this embodiment, an imaging device is used, and the imaging devices are installed at intervals of several kilometers. When the downstream camera captures the license plate of the traveling vehicle at an arbitrary time by the camera, the number is compared with the number captured by the upstream camera and its time history, and if a matching number exists, the travel time is calculated from the time difference. And the travel speed is derived from the distance between the installation points. Reference numeral 4 denotes a processing unit that performs estimation and prediction. The processing unit 4 is configured by a microcomputer or the like, and a program stored in a memory provided in the processing unit 4 executes processing based on measurement values from the vehicle sensor 2 and the AVI 3.

Next, the operation of this embodiment shown in FIG. 1A will be described. In the following description, a section delimited by AVI is referred to as a travel section. Now, a plurality of vehicles are traveling from upstream to downstream, and the speed at each vehicle sensor installation point and the travel time of the travel section are the vehicle sensors 2 and AVI 3.
Is measured by Here, the average value is calculated every 5 minutes, and the average speed V _i , the average travel time T of the section, and the average travel speed V at the point i (i = 1, 2,..., N) are calculated.
Get. Using these, the processing unit 4 estimates and predicts the travel time (or speed), and adaptively corrects the calculation result in the processing unit 4 by comparing this result with the actual average travel time (or speed) based on AVI3. I do.

FIG. 1B is a diagram showing the configuration of the processing unit 4. As shown in FIG. The processing unit 4 arranges the estimating unit 41 and the predicting unit 42 in parallel, and furthermore, models 41a to 42c expressing a plurality of individual units, a matching unit 44, an evaluating unit 45,
And adjusting means 43 comprising two individual means 43a and 43b.

Here, the estimation / prediction means 4a will be described. Here, three models, ie, a model 41a in a traffic jam, a model 41b in a steady state, and a model 41c in a quiet time, are constructed as individual estimation means only for estimation. These estimation models are defined as follows using the time-series speed data from the vehicle sensor 3 as input.

[0022]

(Equation 1)

[0023]

(Equation 2)

[0024]

(Equation 3)

Here, Y _k : estimation result (k = 1, 2, 3) X: 5-minute time average speed data m: number of points (i = 1, 2, 3,..., M) n: number of time points ( j = 0, 1, 2, 3,..., n) f _k : congestion (k = 1), steady (k = 2), quiet (k =
3) Model k: Model number t: Present time The above-mentioned model has different model parameters according to each situation, but it is not necessary to consider competition or duplication between the models at all, and it is sufficient to construct each model.

Each model estimates based on the above equation.
Here, the present embodiment further introduces correction elements represented by (Equation 4) and (Equation 5) below. This is used for adaptive adjustment by the adjusting means 43, and has a proportional attribute and a bias attribute existing between the result of each model and the actual result.

[0027]

(Equation 4)

[0028]

(Equation 5)

Here, α _k is a proportional coefficient, an initial value is 1.0 β _k : a bias coefficient, and an initial value is 0.0 k: a model number (k = 1, 2, 3) Y _k ′ (t): results after correction also, the prediction of the formula _{Y k (t), Y k} ' a (t) Y _k (t
+1) and Y _k ′ (t + 1) (t + 1 represents the next point in time). However, it is constructed separately from the estimation. An estimation prediction result is obtained using these estimation means and prediction means.

Next, the matching means 44 will be described. The matching means 44 in the present embodiment makes a matching judgment between the situation targeted by each estimation or prediction model and the actual traffic situation. Further, the matching means 44 in the present embodiment is based on a distribution analysis of velocity data, and a learning vector quantization method which is one of the neural network methods is used for this. In this example, a data set R _k as a template called a reference vector is previously extracted from the velocity data of the sensor by learning using this method, and matching is performed based on the distance between the reference vector and the vector X formed by the input. . The learning and matching are performed by the following specific method in the learning vector quantization method. First, learning will be described. When the input (the data from the vehicle sensors 2 in this case) is given the reference vector R _c the distance between the input vector is minimized is selected by the following equation.

[0031]

(Equation 6)

[0032]

(Equation 7)

[0033]

(Equation 8)

Here, X (t): a vector based on the input at the time point t, (X ₁ (t), X ₁ (t−1),..., X ₁ (t−n),
_{..., X i (t-j} ), ..., X m (t-n)) R k, R c: a reference _{vector, (R 1 (t),} R 1 (t-1), ..., R 1 (T-n),
.., R _i (t−j),..., R _m (t−n)) D _k (t): distance between vectors at time t k: model number (k = 1, 2, 3) and reference vector Is learned by the following (Equation 9) and (Equation 10). 1) When k∈N

[0035]

(Equation 9)

2) Other than 1)

[0037]

(Equation 10)

Here, N: N = δ (t) γ (t), δ (t): monotone decreasing function 1) is R_cIs the reference vector in the vicinity of. For learning
Therefore, the obtained reference vector is close to the zero vector.
In order from 1 to 1: traffic jam, 2: steady, 3: quiet
assign. When the matching means is executed, the reference vector after learning
Distance D to Torr _k(K = 1, 2, 3) is used directly. One
Mari, D_kIs smaller, the situation of k (1: traffic jam, 2: fixed)
Always 3: Closed) and determine the matching result.
You. When data from the vehicle sensor 2 cannot be obtained in advance
May be set to an optimal initial value for the reference vector. Ma
Also, this vector is variable, and actual traffic conditions and aging
It can be adjusted by the above method according to the movement and the like. Book
In the embodiment, the same applies to both estimation and prediction.
This method can be used commonly for both of the above.
You.

Further, the evaluation means 45 in this embodiment integrates the results of the above three models using the matching results. The integration is based on the following (Equation 11) to (Equation 15).

[0040]

[Equation 11]

[0041]

(Equation 12)

[0042]

(Equation 13)

[0043]

[Equation 14]

[0044]

(Equation 15)

Here, Yi ″ (t): a value after evaluation of the estimation result Dt: a sum of distances as matching results Y: an estimation result by all models Wk: a proportional coefficient for adjustment, and an initial value is 1 .0 T: bias coefficient for adjustment, starting from 0.0 a: 1.0 or higher, the results of each model are weighted and evaluated by matching to traffic conditions and combined. The closer the situation is, the more the estimation result is reflected by the distance.The prediction is performed in the same manner as above, where t of Yi '(t) and Yi "(t) is t + 1.

Next, the operation of the adjusting means 43 will be described.
Here, the matching unit 44, the estimation prediction unit 4a, and the evaluation unit 45 are adjusted. In the present embodiment, the first adjustment 43a and the matching unit 44 for the estimation prediction unit 4a and the evaluation unit 45 are set. Is provided. In the first adjustment 43a, the proportionality and the bias coefficient of each model and the evaluation means 45 are adjusted based on the estimation or prediction error. In this embodiment, the following neural network method using the steepest descent method and the error back propagation method are used. First, (Equation 16) to (Equation 18) show the adjustment of the evaluation means 45 for estimation with respect to W _k and T.

[0047]

(Equation 16)

[0048]

[Equation 17]

[0049]

(Equation 18)

Here, E: error Y _t (t): actual value (value based on AVI3) h ′ (I): 1.0 (∵h ′ (I) = aI + T; a =
1.0) L _c : learning constant (eg, 0.9) S _c : stabilization coefficient (eg, 0.9) W _k ': adjusted W _k Δ: correction amount The prediction error cannot be known until the next time point Therefore, it is derived from the previous inference result and the current value. T is Y _k ″
With (t) set to 1.0 and W _{k set} to 1.0, the same calculation may be performed. Αk in each estimation / prediction model is obtained by adjusting the above equation as in the following (Equation 19) to (Equation 21), and βk is Y
_{The same operation is performed with k} (t) set to 1.0.

[0051]

[Equation 19]

[0052]

(Equation 20)

[0053]

(Equation 21)

The adjustment relating to the prediction is performed by using Y (t) in the above equation.
The (t) of the terms (Yk, Yk ', Yk', Dk, Dt) other than (t) may be set to (t-1) in the same manner, but these mean the predicted value one time before and the matching result. because,
This is because adjustment to the predicted value is impossible without the actual value and must be made at the next point in time. However, the above example is based on the assumption that an actual value is obtained at the next time . If not, no adjustment is performed. As described above, based on the estimation or prediction error, the proportionality and the bias coefficient of each estimation / prediction unit 4a and the evaluation unit 45 are adjusted. This is a real-time or short-term adjustment to the current traffic flow.

In the second adjustment 43b, the matching means 44 is adjusted. This adjustment is for an aging change such as a change in traffic conditions in the travel section.
The purpose is longer than 3a. In this example, the reference vector, which is a variable element of the matching means 44, is updated for a fixed period, here, every month. However, in consideration of continuity from the past, it is performed using speed data for the past two months. Therefore, since the data for at least one month overlaps, it is possible to include the secular variation from the past, and if the traffic condition changes, the matching means 44 and the result by the matching means 44 will change, and at the same time, each model and the evaluation result will also change. Changes gradually with adjustment.

Next, the operation of each means in this embodiment will be described with reference to FIGS. FIG. 2 is a flowchart showing the operation at the time of estimation, and FIG. 3 is a flowchart showing the operation at the time of prediction. The estimation and the prediction operate separately, and a series of steps (steps 21 to 26 and steps 31 to 36) are executed once, and are repeated every cycle, in this embodiment, every 5 minutes. In the matching step, the matching means 44 uses the data from the vehicle sensor 2 to make a comparison judgment with the traffic situation (steps 21 and 32). The estimated prediction derives each travel time in each model (steps 22 and 33). In the evaluation step, a final estimation / prediction result is derived from the judgment result by matching and the result of each model using the evaluation means 45 (steps 23 and 34). In the step of the adjustment 1, the first adjustment 43a of the adjustment means 43 is used, and the estimation / prediction result and the actual result (AV
The estimation and prediction means 4a and the evaluation means 45 are adjusted by the value of I3 (step 24, step 3).
1). In steps 25 and 35, the adjusting means 45
The execution timing of the second adjustment 43b is determined, and if so, the matching means 44 is updated by the second adjustment 43b (steps 26 and 36). These operations are repeatedly performed on the actual system. However, for the prediction, the value of AVI3 is obtained at the next point in time, so the first adjustment 43a is performed prior to the first adjustment 43a.

In addition to the above example, only one model is selected from the result of matching, and this value is evaluated as a final result, and only that model is adjusted. Alternatively, optimal matching, estimation prediction, evaluation, and adjustment means may be set in accordance with the characteristics of the section, such as making adjustment coefficients inherent in the model itself and adjusting by the same means. Further, as an estimation prediction means other than the above examples, a model using different data, or, furthermore, a statistical model, a quantitative model based on mathematical analysis, and the like may be used.
What is necessary is just to construct matching, evaluation, and adjustment means.

Thus, according to the present invention, a plurality of traffic
Easily implement, including various estimated prediction models for the situation
Can be constructed, and the estimated
Travel time from multiple perspectives, considering match, fit and choice
Estimation prediction becomes possible.

[Brief description of the drawings]

1A is a diagram illustrating a configuration of an embodiment of the present invention; FIG. 1B is a diagram illustrating an internal configuration of a processing unit in FIG. 1A;

FIG. 2 is a flowchart showing an operation at the time of estimation in the embodiment.

FIG. 3 is a flowchart showing an operation at the time of prediction in the embodiment.

[Explanation of symbols]

 2 Vehicle Detector 3 AVI 4 Processing Unit 41 Estimating Means 42 Predicting Means 43 Adjusting Means 44 Matching Means 45 Evaluation Means

──────────────────────────────────────────────────続 き Continuation of the front page (72) Inventor Yuji Hasegawa 3-1-1 Tsunashima Higashi, Kohoku-ku, Yokohama-shi, Kanagawa Prefecture Inside Matsushita Communication Industrial Co., Ltd. (56) References JP-A-63-49999 (JP, A) JP-A-3-19100 (JP, A) JP-A-2-253499 (JP, A) (58) Fields investigated (Int. Cl. ⁶ , DB name) G08G 1/00-1/16

Claims (1)

(57) [Claims] (1) provided at arbitrary intervals on a road;
And a vehicle detector that measures the vehicle speed
Are installed on the road where the
And identify the same vehicle, the two points of the vehicle
Measure the travel time and travel speed in the section as the end point
The vehicle individual recognition device and the travel time and
Estimation prediction that stores multiple estimation prediction models of travel speed and travel speed
Measuring means, the vehicle sensor and the vehicle individual recognition device
Whether the obtained data and the estimated prediction model match
Matching means for determining whether or not
Use one estimated prediction model based on the results of
Or use multiple estimated prediction models together
The vehicle detector and the vehicle
Based on the data obtained from both individual recognition devices,
Constant prediction means, said matching means and said evaluation means
Adjusting means for adjusting, the estimation prediction means receives the result of the evaluation means,
Estimate travel time in the section from the estimated prediction model
Alternatively, a travel time estimation / prediction device characterized by performing prediction.