JP4781104B2 - Driving action estimation device and driving support device - Google Patents

Description

  The present invention relates to a driving behavior estimation device, a driving support device, and a vehicle evaluation system, and, for example, relates to a device and system for estimating driving behavior, driving support, and vehicle evaluation using a driver model.

Various proposals have been made on the modeling of the driving operation of a vehicle driver (driver) and its application.
For example, the technique described in Patent Document 1 proposes a technique for evaluating the risk of an intersection road by a driver model using fuzzy rules or a neural network.

JP 2002-140786

In the technique described in Patent Document 1, since a driver model using a fuzzy rule or a neural network is configured, it is necessary to create a fuzzy rule, back propagation, etc., and a driver model is easily created. I can't.
In addition, in the prior art, it is possible to create a model for a general driver, but it is difficult to more accurately express the characteristics of the driving operation for each driver.
For this reason, a driver model representing characteristics for each of a plurality of drivers could not be created.
Furthermore, the conventional driver model is a model for evaluating the risk of the intersection road, and is not a driver model for estimating the driving behavior of the driving action.

  Accordingly, an object of the present invention is to estimate driving behavior using a driver model that can be easily created and that can more accurately express driving characteristics of the driver.

(1) In the invention described in claim 1, as the feature quantity detected as the vehicle travels, the time series data of N types of feature quantities are used as learning data, and the probability distribution in which each data exists in the N-dimensional space is described. A driver model for accelerator, a driver model for brake, a feature amount acquisition means for acquiring at least one feature amount excluding the specific feature amount x of the N types, and the acquired feature amount Maximum a posteriori probability calculating means for calculating the maximum a posteriori probability in the driver model for accelerator or the driver model for brake, and based on the calculated maximum a posteriori probability, the specific feature quantity x with respect to the acquired feature quantity Output means for outputting an estimated value, and the driver model for the accelerator and the driver model for the brake are the N types of feature values. Provided is a driving behavior estimation device characterized in that series data is used as learning data, and a probability distribution in which each data exists is described by a GMM (Gaussian mixture model) calculated by an EM algorithm.
(2) In the present invention of claim 2 Symbol placement, the specific feature amount x if an accelerator operation amount, the maximum a posteriori probability calculation means calculates the maximum posterior probability in the driver model for accelerator, the output means The estimated value output from is the next accelerator operation amount, and the driving behavior estimation device according to claim 1 is provided .
(3) In the present invention of claim 3 Symbol placement, the specific feature x is, if a brake operation amount, the maximum a posteriori probability calculation means calculates the maximum posterior probability in the driver model for the brake, the output means The estimated value output from is the following brake operation amount, and the driving behavior estimation device according to claim 1 is provided .
(4) In the claims 4 Symbol placing of the invention, a driver model for the accelerator, the driver model for brake when the learning data time-series data of the N types of features, at least the accelerator operation amount, a brake Time series data of feature values including an operation amount, a vehicle speed of the own vehicle, a distance between the vehicle ahead, a first-order differential value of the vehicle speed, a second-order differential value, and a first-order differential value and a second-order differential value of the inter-vehicle distance. A driving behavior estimation device according to claim 1, 2 or 3 is provided .
(5) In the invention of claim 5 Symbol mounting a driving action estimating device according to claim 2, a traveling data obtaining means for obtaining vehicle speed and inter-vehicle distance of the vehicle, the driving behavior with respect to the traveling data the acquired the accelerator operation amount estimated by the estimating device, to provide a driving support apparatus characterized by comprising a traveling control means for automatic follow-up running, the relative front vehicle by controlling an engine throttle and brake pedal.
(6) In the invention of claim 6 Symbol mounting a driving action estimating device according to claim 3, the traveling data obtaining means for obtaining vehicle speed and inter-vehicle distance of the vehicle, the driving behavior with respect to the traveling data the acquired accordance brake operation amount estimated by the estimating device, to provide a driving support apparatus characterized by comprising a traveling control means for automatic follow-up running, the relative front vehicle by controlling an engine throttle and brake pedal.

  In the present invention, the time series data of N types of feature quantities detected as the vehicle travels is used as learning data, and a driver model describing a probability distribution in which each data exists in the N-dimensional space is used, and a specific feature quantity x The maximum a posteriori probability in the driver model is calculated for the feature quantity excluding, and is output as the estimated value of the specific feature quantity x. Therefore, it is possible to easily create the driving behavior closer to the driving feature of the driver. Can be estimated.

Hereinafter, preferred embodiments of the driving behavior estimation device, the driving support device, and the vehicle evaluation system of the present invention will be described in detail with reference to FIGS.
(1) Outline of the embodiment In this embodiment, the driving behavior characteristics that are different for each driver are modeled, so that vehicle control and driving support according to the characteristics of the driver are performed, and safe and comfortable safe driving is supported. To do. In addition, a vehicle design evaluation system using objective evaluation criteria based on statistical data is constructed.
If the driver modeling process and the output calculation process using the model are simple, the above application can be realized easily and inexpensively.
Therefore, in this embodiment, by using a GMM (Gaussian mixture model) as a driver model, a driver model for each driver can be easily generated, and further, a driving operation can be performed by a calculation that maximizes a conditional probability. Easily estimate and output behavior.

That is, in the driving behavior estimation apparatus of the present embodiment, a Gaussian mixture model calculated by an EM (Expectation Maximization) algorithm using driving data including a plurality of types of feature quantities such as an accelerator operation amount, a vehicle speed, and an inter-vehicle distance as learning data is a driver model. Adopt as.
This Gaussian mixture model is composed of the parameters of the joint probability density function obtained by calculating the joint probability density distribution by the EM algorithm, and for each driver as necessary, for the driver's accelerator operation, for brake operation, It is generated for each estimated feature amount for the inter-vehicle distance maintenance range.

  Then, travel data Y (= y1, y2,...) Excluding a specific feature amount x among a plurality of feature amounts used in the driver model is measured, and a maximum posterior probability in the driver model for the travel data Y is calculated. Thus, the feature quantity x is estimated.

For example, a driver model of the driver A is generated in advance, and auto-cruise (ACC) that automatically travels following the vehicle ahead is executed.
That is, in ACC, the travel data Y such as the vehicle speed and the inter-vehicle distance, excluding the feature amount x = accelerator operation amount, is detected, and the maximum posterior probability in the instep driver model is calculated. This value is estimated as the accelerator operation amount that the driver A will actually operate under the same conditions, and accelerator control (engine throttle control) is executed according to the estimated accelerator operation amount.
Thereby, an accelerator operation close to the driving operation of the driver who generated the driver model is performed.

  In addition, when driving a vehicle with certain design value data (performance data) in a virtual space based on a road model, the vehicle is estimated by estimating feature quantities of various driving actions using a driver model generated in advance. Evaluate the performance.

(2) Details of Embodiment FIG. 1 shows a principle relating to generation of a driver model by the driving behavior estimation apparatus in the present embodiment and estimation of driving behavior based on the generated driver model.
Regarding generation of the driver model and estimation of driving behavior, the vehicle speed V as the feature quantity, the inter-vehicle distance F from the preceding vehicle, and these primary dynamic feature quantities ΔV and ΔF (first-order differential values), secondary Dynamic feature amounts ΔΔV and ΔΔF (second-order differential value), accelerator operation amount G and primary dynamic feature amount ΔG as a driver model for accelerator operation, and brake operation amount B and primary order as a driver model for brake operation. A case where the dynamic feature amount ΔB is used will be described.

  In the driving behavior estimation apparatus according to the present embodiment, the driving data 1 including the accelerator operation amount, the vehicle speed, the inter-vehicle distance, and the like is used as learning data, and the driver model 2 based on the GMM for each driver corresponding to the driving data is pre- Generate it.

  When estimating the driver's driving behavior (for example, accelerator operation amount), the corresponding driver model 2 is used, and the maximum posterior probability for the measured values (V, F, ΔV,...) 3 of the travel data 1 at time t. By calculating 4, the accelerator operation amount 5 that the driver will operate is estimated.

  In the driving behavior estimation device of this example, it is assumed that the driver determines the amount of operation of the accelerator pedal and the brake pedal based on the current vehicle speed, the inter-vehicle distance, and the primary and secondary dynamic feature amounts thereof. Is based.

Hereinafter, the principle of driver model generation and driving behavior estimation will be described in detail.
(A) Learning of Driver Model In the driver model 2 using the GMM, learning data is required, and the traveling data 1 is used as a feature amount.
The travel data 1 uses time-series data for every predetermined measurement interval s (s is arbitrary, but s = 0.1 seconds in this embodiment).
The travel data 1 is data that is actually driven by a driver for which a driver model is to be generated, and offline learning is possible by using the travel data 1 that has been measured and stored in advance. Further, the traveling data 1 measured and collected in real time when the driver is actually driving may be used.

In the driving behavior estimation apparatus according to the present embodiment, by generating a GMM for each driver, modeling that matches the characteristics of each driver is possible.
As described above, the driver model feature amount (running data 1) includes the vehicle speed, the inter-vehicle distance, and their primary and secondary dynamic feature amounts, the accelerator pedal operation amount, and the primary amount of the accelerator pedal operation amount. Of dynamic features are used.
In this way, by adding the dynamic feature amount to the feature amount and modeling, the time relationship before and after is taken into consideration, and a smooth and highly natural estimation result can be obtained.
In the description, the case where the primary and secondary dynamic feature quantities are used will be described. However, only the primary dynamic feature quantity may be used.

Similarly, a driver model for the brake pedal is also possible.
When generating a plurality of driver models for the accelerator pedal, the brake pedal, the inter-vehicle distance range, etc., data other than the accelerator pedal operation amount, the brake pedal operation amount, etc. (V, F, ΔV, ΔF,... ) May use the same data.

  In the present embodiment, the dynamic feature amount of the travel data is calculated from the measured values of the accelerator operation amount, the vehicle speed, and the inter-vehicle distance, but may be actually measured.

In this embodiment, the driver model 2 is generated by calculating a mixed Gaussian distribution (GMM) for the travel data 1.
That is, the joint probability density distribution for the running data 1 is calculated using the EM algorithm, and the parameters of the calculated joint probability density function = {λi, → μi, Σi | i = 1,2,3,. The driver model 2 is stored in a storage means such as a database.
Here, λi represents a weight, → μi represents an average vector group, Σi represents a variance-covariance matrix group, and M represents the number of mixtures. In addition, a symbol that is previously displayed as → μi means a vector.
As described above, the GMM according to the present embodiment uses the full-width covariance matrix in consideration of the correlation between the feature dimensions.

  As the EM algorithm, the mixed Gaussian distribution is estimated by the EM algorithm in accordance with, for example, Seiichi Nakagawa, “Voice Recognition by Probability Model” (The Institute of Electronics, Information and Communication Engineers 1988, P51 to P54).

(B) Estimation of driving behavior (accelerator pedal and brake pedal operation amount) The driver calculates the operation amount of the accelerator pedal and the brake pedal based on the current vehicle speed, the inter-vehicle distance, and their primary and secondary dynamic features. Based on the assumption that it is determined, the motor behavior such as the pedal operation amount is estimated.
That is, from the simultaneous distribution of the feature amount, the motion behavior such as the accelerator pedal operation amount having the highest probability under the given condition is estimated.

This is a problem of maximizing conditional probability, and is based on the calculation of maximum posterior probability.
That is, the accelerator pedal operation amount ∧G (t) and the brake pedal operation amount ∧B (t) are estimates of the value x (t) that maximizes the conditional probability under the condition where y (t) is given. The maximum posterior probabilities are calculated by the following equations (1) and (2), respectively.

∧G (t) = arg max p (G | ΔG, V (t), F (t), ΔV (t), ΔF (t), ΔΔV (t), ΔΔF (t)) Equation (1)
∧B (t) = arg max p (B | ΔB, V (t), F (t), ΔV (t), ΔF (t), ΔΔV (t), ΔΔF (t)) Equation (2)

Here, like ∧G (t), the one displaying ∧ before means an estimated value.
Also,
p (G | ΔG, V, F, ΔV, ΔF, ΔΔV, ΔΔF)
= {P (G, V, F, ΔV, ΔF, ΔΔV, ΔΔF, ΔG)} / {∫∫ ... ∫p (G, V, F, ΔV, ΔF, ΔΔV, ΔΔF, ΔG) dΔG, dV, dF , dΔV, dΔF, dΔΔV, dΔΔF}
p (B | ΔB, V, F, ΔV, ΔF, ΔΔV, ΔΔF)
= {P (B, V, F, ΔV, ΔF, ΔΔV, ΔΔF, ΔB)} / {∫∫ ... ∫p (B, V, F, ΔV, ΔF, ΔΔV, ΔΔF, ΔB) dΔB, dV, dF , dΔV, dΔF, dΔΔV, dΔΔF}
It is.

  In equations (1) and (2), t represents time, and G, B, V, F, and Δ represent accelerator pedal operation amount, brake pedal operation amount, vehicle speed, inter-vehicle distance, and dynamic feature amount, respectively.

  However, the value of the accelerator pedal and the brake pedal that maximizes the conditional probability is obtained by performing numerical integration with a certain step size (for example, 100 steps from 0 to 10,000) in the interval from the minimum value to the maximum value. The value of the accelerator pedal and the brake pedal when the probability is calculated and the probability becomes the maximum may be used as the estimation result.

FIG. 2 shows an outline of estimation of driving behavior based on the maximum posterior probability.
For simplicity, FIG. 2 shows a case in which ∧x (t) is estimated when a feature value y (t) at a certain time t is given.

FIG. 3 shows the configuration of the driving behavior estimation device.
The driving behavior estimation apparatus in the present embodiment includes a driver model generation unit 10 and a driving behavior estimation unit 11.
This driving behavior estimation device is realized by a computer system including a CPU, a ROM, a RAM, and the like.
In addition, as a driving action estimation apparatus, it can also be set as the structure which is not provided with the driver model production | generation part 10 by using the driver model produced | generated by the other apparatus.

The driver model generation unit 10 includes a driver information acquisition unit 101, a travel data acquisition unit 102, a joint probability density distribution calculation unit 103, and a joint probability density function parameter storage unit 104.
The driver information acquisition unit 101 is information for associating the generated driver model with the driver, and includes a driver ID. That is, it is a driver ID that identifies a driver when the travel data acquired by the travel data acquisition unit 102 is measured.

The travel data acquisition unit 102 acquires travel data as learning data for generating a driver model by GMM.
FIG. 4 shows the travel data acquired by the travel data acquisition unit 102.
As shown in FIG. 4, traveling environment data (a) and driver operation data (b) exist as traveling data.
However, these traveling data list what can be used as data, and are not necessarily all necessary data. Data is appropriately selected depending on the driver model to be generated.

As shown in FIG. 4A, the travel environment data includes travel status data and road status data.
The driving situation data is data that changes depending on driving and environment, and includes vehicle speed, distance between vehicles, weather, traffic jam (degree of traffic jam), brightness, and other data.
The road condition data is data that represents a road condition and does not change depending on the environment. The road condition data includes road type, pavement form, road width, number of lanes, friction coefficient, unevenness coefficient, curve curvature, cant, slope, line of sight, and other data.

As shown in FIG. 4B, the driver operation data includes a steering wheel operation amount, an accelerator pedal operation amount, a brake pedal operation amount, an inter-vehicle distance maintenance range amount, and other data.
This driver operation data is often driving behavior (estimated value of feature amount x) estimated using a generated driver model. Therefore, the number of driver operation data corresponding to the number of driver models to be generated is acquired. For example, when the driver model for operating the accelerator pedal and the driver model for operating the brake are created, the accelerator pedal operation amount and the brake pedal operation amount are acquired.
Both are also acquired when a driver model common to the accelerator operation and the brake operation is generated.

  The travel data acquired by the travel data acquisition unit 102 may be acquired in advance by collecting travel data measured and stored in advance, or at every predetermined sampling time while the driver is actually driving. The detected data may be acquired sequentially.

The joint probability density distribution calculation unit 103 (FIG. 3) calculates the joint probability density distribution in the Gaussian mixture model using the acquired traveling data as learning data.
The simultaneous probability density function parameter {λi, → μi, Σi} obtained as a result of the calculation by the simultaneous probability density distribution calculating unit 103 is associated with the driver ID acquired by the driver information acquiring unit 101, and the simultaneous probability density function parameter Store in the storage unit 104.
Note that the stored joint probability density function parameter is stored so as to be able to distinguish which (driver ID) and what (estimated driving behavior) driver model.

FIG. 5 is a flowchart showing driver model generation processing by the driver model generation unit 10 configured as described above.
The driver information acquisition unit 101 acquires driver information (step 10), and the driving data acquisition unit 102 acquires the driving data collectively or sequentially (step 11).
Note that the order of step 10 and step 11 may be reversed or parallel.

  Next, the joint probability density distribution calculation unit 103 calculates a joint probability density distribution using the acquired travel data as learning data (step 12), and associates the joint probability density function parameter with the driver information as a driver model to associate the joint probability density function. The data is stored in the function parameter storage unit 104 (step 13), and the process is terminated.

  In FIG. 3, the driving behavior estimation unit 11 includes a driver information acquisition unit 111, a travel data acquisition unit 112, a driver model selection unit 113, a maximum posterior probability calculation unit 114, and an estimated value output unit 115 of the feature amount x. Yes.

The driver information acquisition unit 111 acquires a driver ID for specifying the target of the driver model.
This driver information is acquired mainly by the input of the operator (driver himself or other operator).
The driver's (driver's) weight, height, and other information that can specify the driver are used as driver information, stored in association with the driver ID, and the driver information is acquired. The driver ID may be specified.

  The driving data acquisition unit 112 calculates driving behavior (feature amount x) estimated by the driver model from the driving data (N types of feature amounts) used when the driver model generation unit 10 generates a driver model to be used. Excluding travel data (N-1 types of feature values) is acquired.

  Based on the driver ID acquired by the driver information acquisition unit 111 and the driving data acquired by the driving data acquisition unit 112, the driver model selection unit 113 applies a driver model (simultaneous) from the simultaneous probability density function parameter storage unit 104. Select probability density function parameters.

  The maximum posterior probability calculation unit 114 applies the travel data acquired by the travel data acquisition unit 112 to the driver model selected by the driver model selection unit 113, and uses the above formulas (1), (2), etc. Calculate the probability.

  The estimated value output unit 115 of the feature quantity x outputs the value calculated by the maximum posterior probability calculation unit 114 as an estimated value of the feature quantity x.

FIG. 6 is a flowchart illustrating a process of estimating a specific driving action using the generated driver model.
First, driver information is acquired by the driver information acquisition unit 111 (step 20).
Then, the travel data acquisition unit 112 acquires travel data at the current time (time t) (step 21). The travel data acquired here is N-1 types of travel data excluding the feature amount x.
Note that the order of step 20 and step 21 may be reversed or may be processed in parallel.

Then, according to the driver information and the driving data, the corresponding simultaneous probability density function parameter storage unit 104 selects and reads the corresponding simultaneous probability density function parameter (driver model) (step 22).
In addition, in order to select a driver model, you may make it select a driver model from driver | operator information, without using driving | running | working data. In this case, you may make it select a driver model before acquisition of driving | running | working data.
In addition, depending on the device to which the driving behavior estimation device is applied, for example, an automatic following driving device (ACC device) described later, a driver model may be selected in advance. It can be omitted if necessary.

Next, the maximum posterior probability calculation unit 114 calculates the maximum posterior probability by applying the acquired travel data to the selected driver model (step 23).
Regarding the calculation of the maximum posterior probability, if the driving action (feature amount x) to be estimated is the accelerator operation amount, the above equation (1) is used, and if the brake operation amount is the brake operation amount, the equation (2) is used.

  Then, the calculated calculation result in the maximum posterior probability calculation unit 114 is output from the estimated value output unit 115 of the feature value x as the estimated value of the feature value x at the time t by the driver model (step 24), and the main routine Return to

(3) Driving Support Device Next, a driving support device that is an application example using the driving behavior estimation device described above will be described.
This driving assistance device is a device that performs automatic following driving (ACC) following a vehicle ahead.
In this driving support device, the driver is driving himself by automatically operating the inter-vehicle distance, the accelerator operation, and the brake operation during the ACC operation using the driver model generated from the driving data of the driving driver. ACC that is close to the driving sensation is performed, and the driver's uncomfortable feeling is eliminated.

(4) Overview of the driving support device Each driver has a habit when driving. In the conventional ACC operation, the vehicle automatically runs to keep the vehicle speed and distance between vehicles at a constant value. Therefore, the driver can use the accelerator and brake when adjusting the distance between the vehicle and the vehicle ahead. Because it is different from 癖, there are problems such as feeling uncomfortable.

  In the driving support device of the present embodiment, a driver model is generated in advance using the driving data as learning data from the actual driving operation of the driver in a state where the ACC is operable. As a result, the generated driver model reflects the driver's driving habits, such as the relationship between the vehicle speed and the inter-vehicle distance, and the amount of operation of the accelerator and brake when adjusting the distance.

That is, the driver's normal driving is learned and stored as a driver model from the driver's normal driving operation.
The driver model to be generated includes own vehicle data (accelerator operation amount, brake operation amount, vehicle speed,…), forward vehicle data (inter-vehicle distance, relative speed, vehicle type,…), road environment data (ambient brightness, line of sight, tunnel) , Rainfall amount, road surface μ, lane width, road congestion, etc.) and the like.

In addition to the geographical information of the surroundings, the road environment data includes information that changes with TPO, such as ambient brightness, weather, road surface conditions, and road congestion.
These estimate the darkness of the surroundings and the rainfall situation from a clock (time), a headlight switch, a wiper switch and the like.
Also equipped with illuminance meter, rain sensor, road surface μ detection sensor, lane recognition device, various sensors for observing surrounding congestion, image recognition device, etc., and actively acquiring surrounding information such as the type of vehicle ahead Alternatively, weather forecasts, VICS information, and the like may be acquired from the network.

When ACC is executed (during automatic following operation), the driver is released from the accelerator operation by monitoring and maintaining the inter-vehicle distance from the preceding vehicle in the same manner as in normal ACC.
That is, during execution of ACC, current road environment information is collected, and an accelerator operation amount, a brake operation amount, and the like are estimated using a driver model created in advance.
Based on this estimated value, the vehicle distance is adjusted along the estimated distance maintenance range, and the engine throttle and brake are controlled according to the estimated accelerator operation amount and the estimated brake operation amount to automatically follow the vehicle. .

Here, the inter-vehicle distance maintenance range estimator is calculated from the inter-vehicle distance that the driver normally maintains with the forward vehicle in a scene similar to the current situation, and thus the inter-vehicle distance in the scene that the driver prefers. The range is reflected.
The estimated accelerator operation amount is calculated from the accelerator operation that is usually performed when the driver adjusts the inter-vehicle distance in a scene that is similar to the current situation, so that the driver prefers the distance to the vehicle ahead in that scene. The manner of operation to close the distance (catch up quickly, catch up slowly, etc.) is reflected.
The estimated brake operation amount is calculated from the brake operation that is usually performed when the driver adjusts the inter-vehicle distance in a scene similar to the current situation, and this reduces the distance that the driver prefers to the vehicle ahead in that scene. The way of operating to increase the distance (vacating immediately, slowly releasing etc) is reflected.

As described above, the distance between the vehicles can be maintained by the operation of reproducing the driver's eyelids, and the uncomfortable feeling felt by the driver can be reduced.
In addition, the system can be made closer to the driver's sensibility by reproducing the habit of changing the inter-vehicle distance in response to external factors such as ambient brightness, weather, and road surface conditions.

(5) Details of the driving behavior estimation device.
FIG. 7 illustrates a configuration of a driving support device to which the driving behavior estimation device is applied.
The driving support device includes a travel data acquisition unit 12, an accelerator unit 13, a brake unit 14, an ECU (electronic control unit) 15, an ACC switch 16, and a navigation device 17.
In addition, about the structure of the driving assistance apparatus demonstrated by FIG. 7, not all are required, but it demonstrates each part and apparatus which can be used in order to perform an automatic following driving | running | working, It is possible to configure the driving support device by appropriately selecting according to the function or the like.

The travel data acquisition unit 12 includes a vehicle speed sensor 120 that detects the vehicle speed of the host vehicle, an inter-vehicle distance sensor 121 that detects an inter-vehicle distance from a preceding vehicle, an imaging device 122 that images the front of the vehicle, and GPS + vehicle-to-vehicle communication measurement. A distance unit 123 and a road environment information collection unit 124 are provided.
The GPS + inter-vehicle communication distance measuring unit 123 specifies the position (latitude, longitude) of the host vehicle by using a GPS device, and receives the coordinate values (latitude, longitude) of the preceding vehicle through inter-vehicle communication with the preceding vehicle. The inter-vehicle distance between the two vehicles is calculated.
The road environment information collection unit 124 collects road environment information from the headlight switch, the wiper switch, the road surface μ detection sensor, the lane recognition device, the VICS, the vehicle periphery monitoring sensor, and other devices.

The accelerator unit 13 includes an accelerator pedal 131, an accelerator pedal position sensor 132, and an engine throttle control device 133.
The brake unit 14 includes a brake pedal 141, a brake pedal position sensor 142, a brake control device 143, and a brake lamp 144.

The ECU 15 includes a forward vehicle recognition / tracking unit 151, an ACC availability determination unit 152, a driver model generation unit 153, and an ACC processing unit 154.
The ACC processing unit 154 includes an inter-vehicle distance monitoring / maintaining unit 154a, an accelerator operation unit 154b, a brake operation unit 154c, and a driving action estimation unit 154d.
The driving behavior estimation unit 154d calculates estimated values for the inter-vehicle distance maintenance range, the accelerator operation, and the brake operation.
The ECU 15 is configured by a computer system that includes CPU, ROM, RAM, and interface units.

The ACC switch 16 is a switch for the driver to select whether or not to execute ACC travel.
When the ACC switch 16 is off, a driver model is generated.
Further, when the ACC switch 16 is on, the driving operation amount is estimated using the generated driver model, and automatic follow-up traveling according to the estimated amount is performed.

The navigation device 17 includes a current position detection unit 171, map information 172, and the like.
The current position detection unit 171 detects the current position (latitude, longitude) of the vehicle using a GPS receiver or the like, and functions as the GPS of the travel data acquisition unit 12.

Next, the operation of the driving support apparatus configured as described above will be described.
FIG. 8 shows an automatic generation process of the ACC driver model.
First, the ECU 15 determines whether or not the ACC switch 16 is off (step 31).
If the ACC switch 16 is on (i.e., 31; N), the driver demands ACC travel, so the driver model is not generated and the ACC switch 16 is monitored to turn off.

On the other hand, when the ACC switch 16 is off (step 31; Y), the ECU 15 determines whether or not the preceding vehicle can be recognized (step 32).
The forward vehicle can be recognized by the forward vehicle recognition / tracking unit 151, and when the distance to the forward vehicle is detected by the inter-vehicle distance sensor 121, it is determined that the forward vehicle can be recognized. The forward vehicle recognition / tracking unit 151 recognizes and tracks the forward vehicle using a captured image captured in front of the vehicle captured by the imaging device 122.

If the preceding vehicle cannot be recognized (step 32; N), the inter-vehicle distance data cannot be obtained and the driver model for ACC cannot be generated, and the process returns to step 31.
On the other hand, when the vehicle ahead can be recognized (step 32; Y), the ECU 15 then determines from the map information of the navigation device 17 whether or not the scene is capable of ACC operation. For example, when driving on an expressway or Tokyo Metropolitan Expressway, if traveling on a toll road or bypass road on a road that does not have a merging lane for a predetermined distance, it is determined that the vehicle is traveling on a road capable of ACC operation. Is done.

If the scene is not capable of ACC operation (step 33; N), the ECU 15 returns to step 31 and repeats the process.
On the other hand, if the scene is capable of ACC operation (step 33; Y), the ECU 15 executes the driver model generation process described in FIG. 5 (step 34) and ends the process.
In this driver model generation process, a driver model for the inter-vehicle distance maintenance range, a driver model for accelerator operation, and a driver model for brake operation are generated.

  As described above, in the driving support device of the present embodiment, if the ACC switch 16 is turned on, the driver model is automatically generated based on the actual driving state in the environment where the ACC traveling is actually executed. Each generated driver model reflects the driver's driving habits, such as the relationship between the vehicle speed and the inter-vehicle distance, and the amount of operation of the accelerator and brake when adjusting the distance.

Next, the operation when the ACC traveling is actually executed using each driver model generated as described above will be described.
FIG. 9 is a flowchart showing the operation of the ACC process.
The ECU 15 monitors whether or not the ACC switch 16 is turned on (step 41).
When it is detected that the ACC switch 16 is turned on (step 41; Y), the ECU 15 collects road environment information (step 42). Here, the road environment information collected by the ECU 15 excludes the estimated operation amounts (accelerator operation amount, brake operation amount, and inter-vehicle distance maintenance range estimation amount) from the travel data used in the driver model generation process (step 34). Driving data.

Next, the ECU 15 determines whether or not the preceding vehicle recognition / tracking unit 151 can recognize the preceding vehicle to be followed (step 43).
If the vehicle ahead cannot be recognized (step 43; N), the automatic follow-up running cannot be maintained, so the ACC switch 16 is turned off and the driver is notified by voice or image (step 44). The process ends.

On the other hand, if the preceding vehicle can be recognized (step 43; Y), the ECU 15 executes the driving action estimation process described in FIG. 6 (step 45), and the inter-vehicle distance maintenance range estimation amount, the accelerator operation estimation amount, the brake Calculate the estimated operation amount.
Then, the current inter-vehicle distance is determined (step 46), and if it is within the inter-vehicle distance adjustment range, the ECU 15 maintains the current accelerator opening (step 37).
On the other hand, if the inter-vehicle distance is equal to or smaller than the inter-vehicle distance adjustment range, the ECU 15 controls the brake control device 143 according to the estimated brake value estimated by the brake operation driver model (step 48).
If the inter-vehicle distance is equal to or greater than the inter-vehicle distance adjustment range, the ECU 15 controls the engine throttle control device 133 according to the estimated value of the accelerator estimated by the accelerator operation driver model (step 49).

  Next, the ECU 15 determines whether or not the ACC switch 16 has been turned off (step 50). If the ACC switch 16 has not been turned off (step 50; N), the ECU 15 returns to step 42 and continues the ACC drive. Step 50; Y) The process is terminated.

  As described above, according to the driving support device of the present embodiment, the driver model reflecting the driver's driving habits, such as the relationship between the vehicle speed and the inter-vehicle distance, and the amount of operation of the accelerator and brake when adjusting the distance. Is generated, and the accelerator (engine throttle) and the brake are operated based on the accelerator operation amount, the brake operation amount, and the inter-vehicle distance adjustment range estimated by the driver model.

(6) Vehicle Evaluation System Next, a vehicle evaluation system that is a second application example using the driving behavior estimation apparatus described above will be described.
This vehicle evaluation system generates a driver model based on travel data when a plurality of drivers such as professional drivers and general drivers actually travel in various vehicle performances and travel conditions.
In this vehicle evaluation system, instead of evaluating the performance of the vehicle by actually driving the vehicle, the estimated value of the generated driver model (the amount of steering wheel operation) The operation amount of the accelerator pedal, the operation amount of the brake pedal, etc.) is used to evaluate various items such as the acceleration performance, deceleration performance, steering performance, and stable running performance of the vehicle.

(7) Details of Vehicle Evaluation System FIG. 10 shows an outline of the vehicle evaluation system.
As shown in FIG. 10, the vehicle evaluation system calculates the driver model 19 based on the GMM for estimating the pedal operation amount of the driver, and the vehicle acceleration speed a (t) based on the estimated pedal operation amount. A vehicle dynamics calculation unit 20 that updates the vehicle speed v (t) and the inter-vehicle distance F (t) using the vehicle acceleration a (t) and the position of the preceding vehicle, and a vehicle speed v (t ) And the inter-vehicle distance F (t), a dynamic feature quantity calculation unit 22 is calculated.

The driver model 19 uses driving data measured using actual driving, a drive simulator, or the like, and drivers that can be regarded as a population of target users for a vehicle to be evaluated, such as a professional driver or a general driver (for example, 100 A plurality of driver models are created from the driving data in the example).
As a driver model, an accelerator driver model 192 and a brake driver model 193 are generated, and the application determining unit 191 determines and selects which of them is applied.
In addition to the accelerator driver model 192 and the brake driver model 193, a steering wheel driver model for estimating the steering amount of the steering wheel of the driver is generated, and the application determination unit 191 determines the driver model for the steering wheel. May also be selectable.

  The driver model 19 is a part to which a driving behavior estimation device is applied, which receives a feature quantity and a dynamic feature quantity such as a vehicle speed and an inter-vehicle distance, and specifies values of an accelerator pedal and a brake pedal that the driver will operate. In this driver model 19, as explained in the driving behavior estimation device, the driver determines the operation amount of the accelerator pedal and the brake pedal based on the current vehicle speed, the inter-vehicle distance, and the primary and secondary dynamic feature amounts thereof. Based on the assumption that they have decided.

The vehicle dynamics calculation unit 20 calculates the vehicle model (evaluation) from the accelerator pedal operation amount G (t) and the brake pedal operation amount B (t) at a certain time t and the vehicle speed V (t-1) at the previous time. The acceleration speed is calculated using vehicle weight data (vehicle weight, engine performance, brake performance, gear ratio, etc.) regarding the target vehicle.
The vehicle model was implemented in MATLAB (a computer language that can handle matrix operations powerfully based on FORTRAN) with reference to the internal model of the driving simulator used for recording learning data.
In this vehicle model, the acceleration speed of the vehicle is calculated in consideration of the gear ratio, vehicle weight, road friction coefficient, and the like.

The travel environment updating unit 21 calculates the vehicle speed V (t + 1) and the inter-vehicle distance F (t + 1) at the next time (t + 1) using the acceleration speed a (t) at the time t input from the vehicle dynamics calculation unit 20. And update.
The vehicle speed and inter-vehicle distance at the next time (t + 1)
V (t + 1) = V (t) + a (t) × T
F (t + 1) = Df (t + 1) − (Dm (t) + V (t + 1) × T)
Calculated by
However, a (t) represents the acceleration speed output from the vehicle dynamics calculation unit 20, Df (t) represents the travel distance of the preceding vehicle up to time t, and Dm (t) represents the travel distance of the host vehicle up to time t. Indicates.
T is the system update time (sampling period), and in this embodiment, T = 0.1 seconds.

  In order to calculate the inter-vehicle distance, the preceding vehicle has a traveling distance at each time, and the inter-vehicle distance is calculated by obtaining the difference between the traveling distance of the preceding vehicle at the time t and the traveling distance of the own vehicle. Yes.

FIG. 11 shows the configuration of the driving behavior estimation device.
As illustrated in FIG. 11, the driving behavior estimation device includes data used for evaluation, an evaluation execution unit, and evaluation result data.
As the usage data, vehicle performance data 25, an evaluation driver model 26, simulation execution data 27, and a road model 28 are used, and travel evaluation data 36 is output and stored as evaluation result data.
The evaluation execution unit includes a road data development unit 29, a travel environment development unit 30, a driving behavior estimation unit 31, a vehicle dynamics calculation unit 32, a simulation travel processing unit 34, and a travel performance evaluation unit 35, which are CPU, ROM And a computer system including a RAM and the like.

The vehicle performance data 25 is performance data of the vehicle to be evaluated, and is composed of data such as vehicle weight, engine performance, brake performance, gear ratio, suspension spring constant, etc., as shown in FIG. The
As the driver model for evaluation 26, each driver model generated by the driving behavior estimation device described above is used.
As shown in FIG. 12B, the simulation execution data 27 is travel status data developed in a virtual space, and includes vehicle speed, inter-vehicle distance, weather, presence of traffic, and the like. As the traveling status data, data at time t = t1, t2, t3... Is used as time-series data.
The running road model 28 is data relating to a test running road developed in a virtual space, as shown in FIG.

  When the evaluation execution unit is compared with the driving behavior estimation device described with reference to FIG. 10, the driving behavior estimation unit 31 corresponds to the driver model 19, the vehicle dynamics calculation unit 32 corresponds to the vehicle dynamics calculation unit 20, and simulation travel processing The unit 34 and the travel environment development unit 30 correspond to the travel environment update unit 21 and the dynamic feature amount calculation unit 22.

Next, a vehicle design evaluation process in the vehicle evaluation system configured as described above will be described.
FIG. 13 is a flowchart showing the operation of the design evaluation process.
The vehicle performance data 25 is input to the vehicle dynamics calculation unit 32 (step 61), and the vehicle dynamics model is developed in the simulation space (step 62).
Next, the runway model 28 is input to the road data expansion unit 29 (step 63), and the simulation execution data 27 is input to the travel environment expansion unit 30 (step 64), so that the simulation execution environment is expanded in the simulation space (step 64). Step 65).
Then, the evaluation driver model 26 is input to the driving behavior estimation unit 31 and the simulation running is started from t = 0 (step 67).

  Next, the driving behavior estimation unit 31 calculates driver behavior estimation values (accelerator pedal operation amount G (t) and brake pedal operation amount B (t)) from the travel environment data (travel data) at time t (step 68). .

Then, in the vehicle dynamics calculation unit 32, the gear ratio and the vehicle are calculated from the accelerator pedal operation amount G (t) and the brake pedal operation amount B (t) at time t, and the vehicle speed V (t-1) at the previous time. The vehicle travel estimation data 33 such as the acceleration speed a (t) is calculated using the vehicle performance data 25 such as the weight and the friction coefficient of the road (step 69).
In addition to the acceleration speed a (t), the vehicle travel estimation data 33 to be calculated includes the own vehicle speed, the inter-vehicle distance, the center of gravity position, the tire angle, the yaw rate, the pitch rate, and the like as shown in FIG.

Then, the simulation travel processing unit 34 uses the vehicle travel estimation data 33 at time t calculated by the vehicle dynamics calculation unit 32 to calculate the vehicle speed V (t + 1) and the inter-vehicle distance F (t + 1) at the next time (t + 1). Calculate and update (step 70).
Further, the vehicle travel estimation data 33 at t = t + 1 is calculated by the simulation travel processing unit 34 (step 71).

  Then, the travel environment development unit 30 updates the simulation execution environment of t = t + 1 from the vehicle travel estimation data 33 (step 72), and the travel performance evaluation unit 35 calculates and stores the travel locus for the road data (step 73). To do.

  Then, it is determined whether or not the simulation running process for all data up to time tn of the simulation execution data 27 has been completed (step 74). If not completed (step 74; N), the process returns to step 68. Continue simulation using the driver model.

If the processing up to time tn has been completed (step 74; Y), traveling evaluation data is output from the traveling performance evaluation unit 35 and the processing is terminated (step 75).
As shown in FIG. 12E, the travel evaluation data output from the travel performance evaluation unit 35 includes an acceleration curve with respect to the accelerator opening as acceleration performance, a deceleration curve with respect to the brake operation amount as deceleration performance, and a handle as steering performance. A travel curve with respect to the operation amount, a travel locus with respect to the road direction as the stability travel performance, and the like are output.

(8) Simulation Experiment (8-1) Driver Model Learning with GMM Driving data was recorded using a driving simulator for GMM learning.
The course is a straight line, and the behavior data of the preceding vehicle is adopted so that all the vehicle speeds appear in order to give variation to the learning data.
The running was performed twice for 10 minutes, and the twice was used as learning data.

FIG. 14A shows the behavior of the preceding vehicle, and FIG. 14B shows the driving data recorded.
It can be seen that all vehicle speeds appear because variations in the behavior of the preceding vehicle are taken into account. Each model of the accelerator pedal operation and the brake pedal operation was learned as a 16-mix multidimensional mixed normal distribution (GMM) having a full-width covariance matrix.

(8-2) Simulation results and discussion In order to evaluate the constructed vehicle evaluation system, the behavior of the preceding vehicle not included in the learning data was prepared and recorded.
The course was a straight line, and the behavior of the preceding vehicle was recorded in a real environment. FIG. 14C shows the behavior of the preceding vehicle used for evaluation.
Using this preceding vehicle data, driving behavior was generated and compared with actual driving data.

The simulation conditions are as follows.
Learning data: 20 minutes (twice 10 minutes)
Feature amount: V, F, G, ΔV, ΔF, ΔG, ΔΔV, ΔΔF
Course; Straight line Δ Window width; 0.8 seconds Number of mixtures; 16
Update time: 0.1 seconds

FIG. 15 shows a simulation result using the vehicle evaluation system of the present embodiment under the above conditions.
FIG. 15 shows the vehicle speed result (a), the inter-vehicle distance result (b), and the accelerator pedal result (c). The solid line shows the simulation result, and the broken line shows the actual data.
As shown in FIG. 15, for example, regarding the accelerator pedal operation, the characteristics of the waveform of the actual accelerator operation signal are well captured, and it is considered that the modeling by the GMM is successful.

As mentioned above, although the embodiment in the driving action estimation device, the driving support device, and the vehicle evaluation system of the present invention has been described, the present invention is not limited to the described embodiment, and various types are included within the scope described in each claim. Deformation can be performed.
For example, in the described embodiment, the driver model based on the GMM learns the relationship between signals such as the vehicle speed and the inter-vehicle distance and the driving action signal from the learning data. Operation amount estimation is not successful.
For example, in follow-up driving, driving at an inter-vehicle distance exceeding 100 m or driving at an inter-vehicle distance of 1 m is not included in the learning data, and is estimated when the conditions are not included in the learning data. Is not performed well, and as a result, it is kept away from the preceding car or collides.
Therefore, in order to avoid such a situation, when the inter-vehicle distance is L1 or less (for example, 2 m or less), the estimation by the driver model is not performed, the full brake is applied, and the inter-vehicle distance is L2 or more (for example, 100 m or more). Sometimes it may be set so that the accelerator is fully depressed.

It is explanatory drawing showing the principle regarding the production | generation of the driver model by the driving action estimation apparatus in this embodiment, and the estimation of the driving action based on the produced | generated driver model. It is explanatory drawing showing the outline regarding the estimation of the driving action by the maximum posterior probability. It is explanatory drawing showing the structure of the driving action estimation apparatus. It is explanatory drawing showing the traveling data acquired by a traveling data acquisition part. It is the flowchart showing the driver model generation process by a driver model generation part. It is the flowchart showing the process which estimates a specific driving action using the produced | generated driver model. It is a block diagram of the driving assistance device to which the driving action estimation device is applied. 5 is a flowchart showing an automatic generation processing operation of an ACC driver model. It is a flowchart showing operation of ACC processing. It is a conceptual explanatory view showing the outline of a vehicle evaluation system. It is a block diagram of a driving action estimation apparatus. It is explanatory drawing showing the outline | summary of each data in a driving action estimation apparatus. It is a flowchart showing operation of design evaluation processing. It is explanatory drawing showing the behavior and driving data of the preceding vehicle for learning, and the behavior of the preceding vehicle for evaluation. It is explanatory drawing showing the simulation result using a vehicle evaluation system.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Driver model production | generation part 101 Driver information acquisition part 102 Traveling data acquisition part 103 Simultaneous probability density distribution calculation part 104 Simultaneous probability density function parameter storage part 11 Driving action estimation part 111 Driver information acquisition part 112 Driving data acquisition part 113 Driver model Selection unit 114 Maximum posterior probability calculation unit 115 Estimated value output unit of feature quantity x

Claims (6)

As a feature quantity detected as the vehicle travels, time-series data of N types of feature quantities is used as learning data, and a driver model for an accelerator that describes a probability distribution in which each data exists in an N-dimensional space , and a driver for a brake Model and
Feature amount acquisition means for acquiring at least one feature amount excluding the specific feature amount x of the N types;
Maximum posterior probability calculating means for calculating a maximum posterior probability in the driver model for accelerator or the driver model for brake for the acquired feature amount;
Output means for outputting an estimated value of the specific feature amount x with respect to the acquired feature amount based on the calculated maximum posterior probability;
With
The driver model for accelerator and the driver model for brake use the time series data of the N types of feature values as learning data, and a GMM (Gaussian mixture model) calculated by an EM algorithm as a probability distribution in which each data exists. ) Is described in (1). When the specific feature amount x is an accelerator operation amount, the maximum posterior probability calculating means calculates a maximum posterior probability in the driver model for accelerator,
2. The driving behavior estimation device according to claim 1, wherein the estimated value output by the output means is a next accelerator operation amount. When the specific feature amount x is a brake operation amount, the maximum posterior probability calculating means calculates a maximum posterior probability in a driver model for braking,
2. The driving behavior estimation apparatus according to claim 1, wherein the estimated value output by the output means is a next brake operation amount. The accelerator driver model and the brake driver model have at least the accelerator operation amount, the brake operation amount, the vehicle speed of the host vehicle, and the vehicle ahead when the time series data of the N types of feature amounts is used as learning data . 2. The time-series data of feature amounts including an inter-vehicle distance, a first-order differential value, a second-order differential value of the vehicle speed, and a first-order differential value and a second-order differential value of the inter-vehicle distance are used. The driving behavior estimation apparatus according to claim 2 or claim 3. A driving behavior estimation device according to claim 2 ;
Traveling data acquisition means for acquiring the vehicle speed and the distance between the vehicles,
Travel control means for performing automatic follow-up traveling with respect to the preceding vehicle by controlling the engine throttle and the brake pedal according to the accelerator operation amount estimated by the driving behavior estimation device with respect to the acquired travel data;
A driving support apparatus comprising: A driving behavior estimation device according to claim 3 ,
Traveling data acquisition means for acquiring the vehicle speed and the distance between the vehicles,
Travel control means for performing an automatic follow-up traveling with respect to the preceding vehicle by controlling an engine throttle and a brake pedal according to a brake operation amount estimated by the driving behavior estimation device with respect to the acquired travel data;
A driving support apparatus comprising:

Images