JP2012059058A - Risk estimation device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reuse the learning result even when there is a change in types of states detected.SOLUTION: An information classification part 32 classifies types of sensor information, a learning part 34 learns corresponding relation between the sensor information and a collision risk prediction degree for each classification of the sensor information, based on the learning data. An information acquisition part 26 acquires a variety of sensor information. A risk estimation part 38 calculates corresponding collision risk prediction degree for each classification, based on the learned corresponding relation and the variety of sensor information acquired, and estimates a comprehensive value of the collision risk prediction degree.

Description

  The present invention relates to a risk estimation device and a program, and more particularly, to a risk estimation device and a program for estimating a risk prediction degree based on a traveling state of a vehicle and an operation state of a driver.

  Conventionally, the risk information included in the external environment is learned by learning the correlation with the one-dimensional state calculated from the multidimensional feature amount included in the external environment, using the risk information extracted from the operation information of the moving body as the teacher information. A technique for automatically recognizing the degree is known (Patent Document 1).

JP 2008-238831 A

  However, in the technique described in Patent Document 1, since the multidimensional feature value is converted into a one-dimensional state, the past learning result is reused when the combination of feature values is changed. There is a problem that it is not possible to re-learn.

  The present invention has been made in order to solve the above-described problems, and provides a risk estimation device and program capable of reusing learning results even when the type of detected state is changed. The purpose is to provide.

  In order to achieve the above object, the risk estimation apparatus according to the present invention includes a plurality of types including a traveling state of a vehicle, an operation state when a driver operates the vehicle, or an environmental state indicating a traveling environment around the vehicle. Stored in the storage means, the state detection means for detecting the state of the condition, the storage means for storing the correspondence relationship between the state belonging to the classification and the risk prediction degree, learned for each classification into which the plurality of types of states are classified. Based on the correspondence relationship and the plurality of types of states detected by the state detection means, the corresponding risk prediction degree is obtained for each classification, and a comprehensive value of the risk prediction degree is estimated. And estimating means.

  The program according to the present invention is learned for each classification into which a plurality of types of states including a vehicle driving state, an operation state when a driver operates the vehicle, or an environmental state indicating a driving environment around the vehicle are classified. A computer including a storage unit that stores a correspondence relationship between a state belonging to the classification and a risk prediction degree is detected by the state detection unit that detects the correspondence relationship stored in the storage unit and the plurality of types of states. A program for obtaining a corresponding risk prediction degree for each of the classifications based on the plurality of types of states and functioning as an estimation means for estimating a comprehensive value of the risk prediction degree.

  According to the present invention, the state detecting means detects a plurality of types of states including a traveling state of the vehicle, an operation state when the driver operates the vehicle, or an environmental state indicating a traveling environment around the vehicle. Then, based on the correspondence relationship stored in the storage unit and the plurality of types of states detected by the state detection unit, the estimation unit obtains the corresponding risk prediction degree for each classification, and the overall risk prediction degree is obtained. Estimate the correct value.

  In this way, because the risk prediction degree is estimated using the correspondence between the state and the risk prediction degree learned for each detected state classification, the type of the detected state has changed Even so, the learning results can be reused.

  The correspondence relationship stored in the storage means according to the present invention can be learned as learning data for each classification, the state belonging to the classification and the risk level obtained when the state is detected. .

  In addition, the above-described classification of the plurality of types of states is performed based on the mutual information amount between the plurality of types of states detected when the degree of risk equal to or higher than the threshold is obtained, and the mutual information between the levels of risk and the respective types of states. It may be classified based on at least one of an amount, a mutual information amount between a plurality of types of states, and a correlation coefficient between a plurality of types of states. As a result, combinations of states having a strong relationship with the degree of risk or combinations of states having a strong relationship between the states can be classified into the same class.

  The correspondence stored in the storage unit according to the present invention can be learned by a learning device that learns the correspondence based on the learning data. At this time, the plurality of risk estimation devices can use the learning result in the learning device.

  Further, the learning device can further classify a plurality of types of states.

  The estimation means according to the present invention can estimate the overall value of the risk prediction degree by weighted addition of the risk prediction degree obtained for each classification.

  The estimation means according to the present invention can estimate the maximum value of the risk prediction degree obtained for each classification as a comprehensive value of the risk prediction degree.

  As described above, according to the risk estimation apparatus and program of the present invention, the risk prediction degree is estimated using the correspondence relationship between the state and the risk prediction degree, which is learned for each detected state classification. Therefore, even if there is a change in the type of state to be detected, there is an effect that the learning result can be reused.

It is a block diagram which shows the structure of the collision risk estimation apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows the map for converting into a state from sensor information. It is a flowchart which shows the content of the classification process routine in the collision risk estimation apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the content of the classification process routine in the collision risk estimation apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the content of the learning process routine in the collision risk estimation apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the content of the danger estimation process routine in the collision risk estimation apparatus which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the structure of the collision risk estimation apparatus which concerns on the 2nd Embodiment of this invention. It is a flowchart which shows the content of the learning update process routine in the collision risk estimation apparatus which concerns on the 2nd Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. A case where the present invention is applied to a collision risk estimation device mounted on a vehicle will be described as an example.

  As shown in FIG. 1, the collision risk estimation apparatus 10 according to the first embodiment includes a GPS sensor 12 that detects a vehicle position (latitude and longitude) as a traveling state of the vehicle, and a traveling environment around the vehicle. An inter-vehicle distance measuring unit 14 that measures a front inter-vehicle distance as a state, an acceleration sensor 16 that detects acceleration as a traveling state of the host vehicle, and a brake amount (brake as an operation state when the driver operates the host vehicle) Brake sensor 18 that detects the amount of depression of the vehicle, a vehicle speed sensor 20 that detects the vehicle speed as the traveling state of the host vehicle, a GPS sensor 12, an inter-vehicle distance measuring unit 14, an acceleration sensor 16, a brake sensor 18, and a vehicle speed sensor 20. A computer 22 for estimating a collision risk prediction degree based on the output from the alarm device and outputting an alarm sound from the alarm device 24 according to the estimation result. .

  The GPS sensor 12 outputs a host vehicle position signal indicating the detected position of the host vehicle. The inter-vehicle distance measurement unit 14 includes, for example, a laser radar, measures the inter-vehicle distance from a preceding vehicle ahead, and outputs an inter-vehicle distance signal indicating the measured inter-vehicle distance. The acceleration sensor 16 outputs an acceleration signal indicating the detected acceleration. The brake sensor 18 detects the brake amount operated by the driver, and outputs a brake signal indicating the detected brake amount. The vehicle speed sensor 20 outputs a vehicle speed signal indicating the detected vehicle speed of the host vehicle.

  The GPS sensor 12, the inter-vehicle distance measuring unit 14, the acceleration sensor 16, the brake sensor 18, and the vehicle speed sensor 20 input each signal to the computer 22 at certain intervals (for example, every 100 ms).

  The computer 22 includes a CPU, a RAM, and a ROM that stores a program for executing a classification processing routine, a learning processing routine, and a risk estimation processing routine, which will be described later, and is functionally configured as follows. ing. The computer 22 includes an information acquisition unit 26 that acquires signals from the GPS sensor 12, the inter-vehicle distance measurement unit 14, the acceleration sensor 16, the brake sensor 18, and the vehicle speed sensor 20 and stores them in an information storage unit 28 described later. Based on the information storage unit 28 that stores various sensor information including vehicle position, inter-vehicle distance, acceleration, brake amount, and vehicle speed in correspondence with the collision risk described later in time series, and the acceleration of the host vehicle, A risk determination unit 30 that determines the collision risk, an information classification unit 32 that classifies various sensor information types including the position of the host vehicle, the inter-vehicle distance, acceleration, brake amount, and vehicle speed, and a time series of various sensor information Based on the data and the collision risk determined at each time point, a learning unit 34 that learns the correspondence between various sensor information and the collision risk prediction level, and a learning result by the learning unit 34 A learning result storage unit 36 for storing, and a.

  The information acquisition unit 26 repeatedly acquires various signals that are continuously output in time series from the GPS sensor 12, the inter-vehicle distance measurement unit 14, the acceleration sensor 16, the brake sensor 18, and the vehicle speed sensor 20. The data is stored in the storage unit 28.

  Based on the current acceleration stored in the information storage unit 28, the risk determination unit 30 determines the current collision risk according to the following equation (1) and corresponds to various sensor information detected at the same time. And stored in the information storage unit 28.

However, _{r t} collision risk at time t, is _{a t} longitudinal acceleration at time t, Th is the threshold.

  In addition, you may make it determine the present collision risk according to the following (2) Formula.

However, Th _max and Th _min are threshold values.

  Also, instead of a continuous value as calculated by the above equation (2), a discrete value obtained by dividing the 0 to 1 section into N stages at equal intervals, or a discrete value obtained by dividing the N stages at predetermined intervals, It is good also as a collision risk.

  The information classification unit 32 classifies the types of various sensor information based on the data for a certain period stored in the information storage unit 28 as described below.

  First, for each type of sensor information, the mutual information amount between the corresponding collision risks is calculated according to the following equation (3). If the mutual information amount is smaller than the threshold, the sensor information The types are classified into sensor information classes that are not related to the collision risk.

However, S ⁱ represents the type of sensor information.

  Next, for various types of sensor information other than the sensor information not related to the collision risk, the mutual information amount between the various types of sensor information detected under the condition satisfying the collision risk r> 0 is calculated as sensor information. Each combination of types is calculated according to the following equation (4), and combinations of sensor information types for which the calculated mutual information amount is equal to or greater than a threshold are classified as the same class. However, one type of sensor information is allowed to belong to a plurality of classes.

  According to the above classification, combinations of sensor information types that are strongly related to reward (collision risk) and also strongly related to sensor information can be classified as the same class.

  The learning unit 34 performs learning for each classification as described below based on data for a certain period stored in the information storage unit 28.

  First, in the target classification, the type of sensor information belonging to the classification is converted into one state for each of the time series data. As a conversion method from various sensor information to a state, various sensor information is discretized, and various sensors at each time according to a map in which each grid corresponding to a combination of discrete values is in the same state as shown in FIG. Transform information into one state.

  The mapping from the vector representing various sensor information to the state is learned by SOM (Self Organizing Map), and the SOM node corresponding to the various sensor information is obtained as the converted state. Also good.

  Then, the converted time-series data and the collision risk time-series data are used as learning data for the classification, the collision risk prediction degree that is the future collision risk, and the state converted as described above. Learn the correspondence of.

  In the present embodiment, the state value V expressed by the following equation (5) is used as the collision risk prediction degree, and the correspondence between the converted state and the state value V is learned.

However, s _t the state at time t, γ is the discount rate, r _t is the collision risk at time t. E [] represents an expected value.

Using the time-series data in the converted state and the time-series data of the collision risk as learning data, for example, using the TD error learning method, the state value V corresponding to each state st is _expressed according to the following equation (6). Update and learn.

Incidentally, by using a Monte Carlo method may be learned by updating the state value V for each state s _t. In this case, it is sufficient to the following equation (7), updates the state value V for each state s _t.

  However, T is a constant, and it suffices to abort the update with a sufficiently large T.

Learning unit 34, as described above, for each classification, learning a correspondence relationship between the state s _t and state value V, the correspondence between the state s _t and state value V of each classification learned, the learning result Store in the storage unit 36. Incidentally, as a learning result may be obtained a table representing the correspondence between the state s _t and state value V, it may be determined function representing the relationship between the state s _t and state value V. Further, the correspondence between the state _st and the state value V is an example of the correspondence between the state belonging to the classification and the risk prediction degree.

The computer 22 is further estimated by a risk estimation unit 38 that estimates the degree of collision risk prediction based on the current various sensor information and the correspondence between the learned state _st and state value V for each classification. And an alarm control unit 40 that controls the output of the alarm device 24 in accordance with the degree of collision risk prediction.

Danger estimating unit 38, first, for each classification, the current sensors information belonging to the classification, according to the map shown in FIG. 2, it is converted to a state s _t. Then, the risk estimation unit 38 calculates the state value V (s _t ) for each classification based on the converted state s _t and the correspondence between the learned state s _t and the state value V, The degree of collision risk prediction for the classification.

Further, the risk estimation unit 38 estimates the overall value V ^* (t) of the collision risk prediction degree at the current time t according to the following equation (8) based on the collision risk prediction degree calculated for each classification. To do.

Here, s _t ⁱ is the state value V at time t in the i-th classification, and w _i is a weight for the i-th classification. The weight is normally set to 1 for all classifications, but may be set for each classification according to the degree of accident damage.

  When the alarm control unit 40 determines that the total value of the estimated collision risk prediction degree is equal to or greater than the threshold value, the alarm device 24 causes the driver to output an alarm sound.

  Next, the operation of the collision risk estimation apparatus 10 according to the first embodiment will be described. First, during traveling of a vehicle equipped with the collision risk estimation device 10, signals output from the GPS sensor 12, the inter-vehicle distance measurement unit 14, the acceleration sensor 16, the brake sensor 18, and the vehicle speed sensor 20 are output from the computer 22. The information is acquired and stored in the information storage unit 28, and the collision risk is determined based on the acquired acceleration signal and stored in the information storage unit 28. When data for a certain period of time is accumulated in the information storage unit 28, the computer 22 executes a classification process routine shown in FIGS.

First, in step 100, an initial value 0 is substituted into a variable i for identifying the type of sensor information, and in step 102, the mutual information amount I (S of the sensor information type _Si and the collision risk r is calculated. _i ; r) is calculated.

Then, in step 104, the mutual information I _(S i; r) may determine whether a higher threshold Th _r, mutual information I _(S i; r) is equal to or more than the threshold value Th _r In step 106, 1 is substituted into the element A (S _i ) of the vector indicating whether or not there is a relationship with the collision risk, and the fact that the sensor information type S _i is related to the collision risk is registered. . On the other hand, if the mutual information amount I (S _i ; r) is less than the threshold Th _r , in step 108, 0 is set to the element A (S _i ) of the vector indicating whether or not there is a relationship with the collision risk. Substituting and registering that the sensor information type S _i is not related to the collision risk.

  In the next step 110, the constant N is a number indicating the type of sensor information, and it is determined whether or not the variable i is N-1 or more. If the variable i has not reached N-1, the variable i is incremented in step 112, and the process returns to step 102. On the other hand, when the variable i reaches N−1, 0 is substituted into the variable i in step 114, and the set S of sensor information types related to the collision risk is set as an empty set.

In step 116, it is determined whether A (S _i ) is 1. If the sensor information type S _i is registered as having a relationship with the collision risk, the sensor information S _i is added to the sensor information set S in step 118. On the other hand, if it is registered that the type of sensor information S _i is not related to the collision risk, the process proceeds to step 120.

  In step 120, it is determined whether or not the variable i is N-1 or more. If the variable i has not reached N−1, the variable i is incremented in step 122 and the process returns to step 114. On the other hand, when the variable i reaches N−1, the process proceeds to step 123.

In step 123, the variable i for identifying the type of sensor information, the j, with substitutes an initial value 0, the set G _i reserved for each type Si sensor information, the empty set.

In the next step 124, it is determined whether or not the sensor information type S _i and the sensor information type S _j are included in the set S. When both the sensor information type S _i and the sensor information type S _j are included in the set S, in step 126, the sensor information type S _i detected when the collision risk is greater than zero. The mutual information I (S _i ; S _j | r> 0) between the sensor information of the sensor information and the sensor information of the sensor information type S _j is calculated, and it is determined whether or not it is equal to or greater than the threshold Th _s . If the mutual information amount I (S _i ; S _j | r> 0) is equal to or greater than the threshold Th _s , the sensor information type S _j is added to the set G _i for the sensor information type Si in step 130. Then, the process proceeds to step 132.

If the mutual information amount I (S _i ; S _j | r> 0) is less than the threshold Th _s in step 128, the process proceeds to step 132.

If at least one of the sensor information type S _i and the sensor information type S _j is not included in the set S in step 124, the process proceeds to step 132.

  In step 132, it is determined whether or not the variable j is N-1 or more. If the variable j has not reached N-1, the variable j is incremented in step 134 and the process returns to step 124. On the other hand, when the variable j reaches N−1, the routine proceeds to step 136.

  In step 136, it is determined whether or not the variable i is N-1 or more. If the variable i has not reached N−1, the variable i is incremented in step 138, 0 is substituted for the variable j, and the process returns to step 124. On the other hand, when the variable i reaches N-1, the classification processing routine is terminated.

The set G _i (i = 0,..., N−1) obtained by the above classification processing routine is a classification result obtained by classifying the types of sensor information.

  Further, in the computer 22, a learning process routine shown in FIG. 5 is executed.

  First, in step 150, signals output from the GPS sensor 12, the inter-vehicle distance measuring unit 14, the acceleration sensor 16, the brake sensor 18, and the vehicle speed sensor 20 are acquired and stored in the information storage unit 28. In step 152, the collision risk is determined based on the acquired acceleration signal and stored in the information storage unit 28.

  In the next step 154, it is determined whether or not the number of data stored in the information storage unit 28 has become a predetermined number or more. When a predetermined number or more of data is accumulated in the information storage unit 28, the process proceeds to step 156. . On the other hand, if the number of data stored in the information storage unit 28 does not reach the predetermined number, the process returns to step 150.

In step 156, based on the time series data of various sensor information stored in the information storage unit 28 and the time series data of the collision risk determined in step 152, the sensor information belonging to the class is classified for each class. learning a correspondence relationship between transformed state s _t and state value V from, in step 158, it stores the learning result in the learning result storage unit 36, and ends the learning processing routine.

The learning processing routine is periodically executed, and the correspondence relationship between the state _st and the state value V for each classification is periodically updated.

  Further, the computer 22 executes a risk estimation processing routine shown in FIG.

First, in step 160, signals output from the GPS sensor 12, the inter-vehicle distance measuring unit 14, the acceleration sensor 16, the brake sensor 18, and the vehicle speed sensor 20 are acquired. In the next step 162, for each classification of sensor information, based on the signals acquired in step 160 of the various sensor information belonging to the classification and the correspondence between the state _st and the state value V learned for the classification. Thus, the state value is calculated for each classification to obtain a collision risk prediction degree, and a comprehensive value of the collision risk prediction degree is estimated from the collision risk prediction degree for each classification.

  In step 164, it is determined whether or not the overall value of the collision risk prediction degree estimated in step 162 is equal to or greater than a threshold value. If the overall value of the estimated collision risk prediction degree is less than the threshold value, the danger estimation processing routine is terminated. On the other hand, if the overall value of the estimated collision risk prediction is equal to or greater than the threshold value, it is determined that there is a risk of collision if the host vehicle travels as it is. An alarm sound is output and the danger estimation processing routine is terminated.

  The risk estimation processing routine is repeatedly executed every predetermined time, and a comprehensive value of the collision risk prediction degree is repeatedly estimated.

  As described above, according to the collision risk estimation device according to the first embodiment, the collision risk is learned using the correspondence relationship between the state and the state value learned for each type of detected sensor information. Since the degree of prediction is estimated, even if there is a change in the type of sensor information to be detected, the learning result for the classification composed of the type of sensor information that has not been changed can be reused. .

  In addition, since the relationship between sensor information corresponding to actions often taken by the driver when avoiding danger and the collision risk prediction degree is learned, there is a high probability that the driver must avoid the danger (collision risk prediction degree). The situation can be detected automatically.

  In addition, in order to learn the correspondence between the state and the state value for each classification of the type of sensor information to be detected, the sensor information classification that can be acquired in common even for vehicles with different sensor combinations, Learning results can be shared. Further, when the combination of sensors is changed, it is only necessary to re-learn only the classification that has been affected, so that the reusability of the learning result is improved.

  In addition, by classifying combinations of sensor information types that are strongly related to collision risk and also have strong relationships between sensor information types, the reusability and sharing of learning results, In addition, it is possible to improve the learning efficiency when there are many types of sensor information.

  Next, a second embodiment will be described. In addition, about the part which becomes the structure similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  The second embodiment is different from the first embodiment in that learning is performed and the learning results are accumulated in a server different from the in-vehicle collision risk estimation device.

  As shown in FIG. 7, the collision risk estimation system 200 according to the second embodiment includes a collision risk estimation device 210 and a server 212 mounted on the vehicle, and the collision risk estimation device 210 and the server 212. Is connected via a base station 214 installed on the roadside and the Internet 216. The server 212 is an example of a learning device. Moreover, the collision risk estimation device 210 is mounted on a plurality of vehicles.

  The computer 222 of the collision risk estimation apparatus 210 includes an information acquisition unit 26, an information storage unit 28, a risk determination unit 30, and time series data of various sensor information stored in the information storage unit 28 and the time of collision risk. The series data is transmitted to the server 212 via the base station 214 and the Internet 216, and from the server 212, a communication unit 232 that receives a learning result corresponding to the type of sensor information of the host vehicle, a learning result storage unit 36, The risk estimation unit 38 and the alarm control unit 40 are provided.

  The communication unit 232 uses the time series data of the various sensor information and the time series data of the collision risk stored in the information storage unit 28 as a list of the types of sensor information of the own vehicle (for example, the position of the own vehicle, the inter-vehicle distance, Together with a list representing acceleration, brake amount, and vehicle speed), the data is transmitted to the server 212 via the base station 214 and the Internet 216.

The communication unit 232 transmits a list of types of sensor information of the own vehicle and a transmission request for the learning result to the server 212 via the base station 214 and the Internet 216, for each classification of various sensor information of the own vehicle. The correspondence relationship between the learned state _st and the state value V is received from the server 212 and stored in the learning result storage unit 36.

  The server 212 includes a CPU, a RAM, and a ROM that stores a program for executing a learning update processing routine described later, and is functionally configured as follows. The server 212 receives time series data of various sensor information, time series data of collision risk, and a list of sensor information types transmitted from the collision risk estimation device 210 and is transmitted from the collision risk estimation device 210. A communication unit 240 that transmits a learning result corresponding to the list of types of sensor information to the collision risk estimation device 210 via the base station 214 and the Internet 216, and various sensor information received from the collision risk estimation device 210. A data storage unit 242 that stores time series data of series data and collision risk, an information classifying unit 244 that classifies types of various sensor information detected by various vehicles, time series data of various sensor information, A learning unit 246 that learns a correspondence relationship between various sensor information and the collision risk prediction degree based on the collision risk degree determined at the time point; The learning result storage unit 248 that stores the learning result of the response relationship, and the learning result corresponding to the list of types of sensor information transmitted from the collision risk estimation device 210 is selected from the learning result stored in the learning result storage unit 248 And a learning result selection unit 250.

  The communication unit 240 receives time series data of various sensor information, time series data of collision risk, and a list of types of sensor information from the collision risk estimation device 210 of various vehicles, and stores them in the data storage unit 242. .

  The data storage unit 242 stores various types of sensor information time series data, collision risk time series data, and a list of types of sensor information transmitted from the collision risk estimation device 210 of various vehicles.

  The information classification unit 244 identifies all types of sensor information that can be detected by various vehicles from the list of types of sensor information stored in the data storage unit 242, and determines all types of sensor information as the first implementation. It classifies like the form of.

  The learning unit 246 performs learning for each time-series data of the sensor information transmitted from the collision risk estimation device 210 as described below.

  First, a list of corresponding sensor information types stored in the data storage unit 242 is acquired, and among the sensor information classifications classified by the information classification unit 244, the types of sensor information in the sensor information type list And the learning result for the specified classification is acquired from the learning result storage unit 248.

Then, the time series data of the corresponding sensor information and the time series data of the collision risk stored in the data storage unit 242 are acquired as learning data, and the learning result for the acquired classification is obtained for each identified classification. used, learning a correspondence relationship between transformed state s _t and state value V from the sensor information belonging to the classification, it updates the learning result stored in the learning result storage unit 248.

  When the learning result selection unit 250 receives the list of types of sensor information and the transmission request for the learning result from the collision risk estimation device 210, the learning result selection unit 250 has received the sensor information classified by the information classification unit 244. A classification composed of sensor information types in the sensor information type list is specified, and a learning result for the specified classification is selected and acquired from the learning result storage unit 248.

  Next, the operation of the collision risk estimation system 200 according to the second embodiment will be described.

  First, while the vehicle equipped with the collision risk estimation device 210 is running, signals output from the GPS sensor 12, the inter-vehicle distance measurement unit 14, the acceleration sensor 16, the brake sensor 18, and the vehicle speed sensor 20 are output from the computer 222. The information is acquired and stored in the information storage unit 28, and the collision risk is determined based on the acquired acceleration signal and stored in the information storage unit 28. When data for a certain period of time is accumulated in the information storage unit 28, the computer 222 converts the time series data of various sensor information and the time series data of the collision risk stored in the information storage unit 28 into the sensor information of the host vehicle. Are transmitted to the server 212 via the base station 214 and the Internet 216.

  The server 212 stores the data received from the collision risk estimation device 210 for each vehicle in the data storage unit 242.

  Further, the server 212 identifies all types of sensor information from the list of types of sensor information stored in the data storage unit 242, and, as in the classification processing routines of FIGS. Classify types.

Further, the server 212 executes a learning update processing routine shown in FIG. First, in step 260, it is determined whether learning data (various sensor information time series data, collision risk time series data, and sensor information type list) has been received from the collision risk estimation apparatus 210. When the learning data is received, the process proceeds to step 262, and it is determined whether or not each of the sensor information classifications configured in the sensor information type list of the learning data has already been classified. If it has already been classified, in step 264, for each of the classifications that have a classification result in the learning result storage unit 248, the learning result for the classification is selected from the learning result storage unit 248, and the selected learning result is displayed. used, learning a correspondence relationship between transformed state s _t and state value V from the sensor information belonging to the classification, it updates the learning result stored in the learning result storage unit 248. On the other hand, if there is no classification result for each of the classifications of sensor information configured in the list of types of sensor information in the learning data, the process proceeds to step 266.

  The selection result of the learning result in step 264 in response to the reception of the learning data from the collision risk estimation device 210 is stored, and the selection of the learning result in step 264 is performed by learning from the collision risk estimation device 210. It may be performed only once when the learning result is updated for the first time according to the data.

  In step 266, sensor information is classified for the list of types of sensor information in the learning data. At this time, the sensor information is classified as in the classification processing routine shown in FIGS.

  In step 268, it is determined whether or not there is a classification having no learning result for each of the sensor information classifications classified in step 266. If there is already a learning result for each of the sensor information classifications classified in step 266, it is determined that there is no new classification and has already been learned, and the learning update processing routine is terminated.

On the other hand, if it is determined in step 268 that there is a classification having no learning result, it is determined that a new classification has been added. In step 270, for each of the classifications, a sensor belonging to the classification of the learning data. based on the information, to learn the correspondence between the transformed state s _t and state value V from the sensor information belonging to the classification. In step 272, the learning result is stored in the learning result storage unit 248, and the learning update processing routine is terminated.

In addition, the collision risk estimation apparatus 210 transmits a list of types of sensor information of the host vehicle and a learning result transmission request to the server 212 via the base station 214 and the Internet 216. At this time, the server 212 stores the learning result (correspondence between the state _st and the state value V) for the classification configured in the received list of types of sensor information, and the learning result stored in the learning result storage unit 248. The acquired learning result is transmitted to the collision risk estimation apparatus 210 via the base station 214 and the Internet 216. The collision risk estimation device 210 stores the received learning result in the learning result storage unit 36. Note that the learning result selection result corresponding to the transmission request from the collision risk estimation device 210 is stored, and the learning result is selected only once when the collision risk estimation device 210 uses the learning result for the first time. It may be.

  Further, the collision risk estimation device 210 periodically transmits a list of types of sensor information of the host vehicle and a transmission request for learning results to the server 212, and updates the learning results stored in the learning result storage unit 36. Update to one.

  In addition, the collision risk estimation device 210 estimates the overall value of the collision risk prediction degree and controls the alarm device 24 in the same manner as the risk estimation processing routine of FIG.

As described above, according to the collision risk estimation system according to the second embodiment, the learning data is collected in the server connected to the plurality of collision risk estimation devices, and the state _st and the state value V are obtained. By learning the correspondence and accumulating the learning results, the learning results can be easily shared among the plurality of collision risk estimation devices.

  In addition, in order to learn the correspondence between the state and the state value for each classification of the type of sensor information to be detected, the sensor information classification that can be acquired in common even for vehicles with different sensor combinations, The learning result accumulated in the server can be shared.

  In the first embodiment to the second embodiment described above, a case where the vehicle position, the inter-vehicle distance, the acceleration, the brake amount, and the vehicle speed are detected as the operation state, the traveling state, and the traveling environment state. Although described as an example, the present invention is not limited to this. For example, own vehicle position, inter-vehicle distance, distance and angle from own vehicle to specific point (intersection, pedestrian, bicycle, etc.), road structure (road width, number of lanes, presence of signal, etc.), vehicle speed, acceleration, roll, pitch, Arbitrary combinations of angular acceleration in the yaw direction, accelerator and brake depression amounts, steering angles, and first order differential amounts, second order differential amounts, first order integral amounts, and combinations of these information. May be detected.

  Moreover, based on the mutual information amount between the types of the plurality of types of sensor information detected when the collision risk level equal to or greater than the threshold is determined, and the mutual information amount between the collision risk level and each type of sensor information The case of classifying the types of sensor information has been described as an example, but the present invention is not limited to this. For example, mutual information amount between multiple types of sensor information detected when a collision risk level equal to or higher than a threshold is determined, mutual information amount between collision risk level and each type of sensor information, no condition The types of sensor information may be classified based on at least one of the mutual information amount between the types of the plurality of types of sensor information and the correlation coefficient between the types of the types of sensor information.

  Moreover, although the case where the collision risk is determined based on the acceleration has been described as an example, the present invention is not limited to this, and the collision risk is determined based on the brake operation amount or the handle operation amount. Also good. The collision risk may be determined based on any combination of acceleration, brake operation amount, and steering wheel operation amount.

  Moreover, although the case where the overall value of the collision risk prediction degree is estimated by weighted addition of the collision risk prediction degree calculated for each classification has been described as an example, the present invention is not limited to this. The maximum value of the collision risk prediction degree calculated for each classification may be estimated as a comprehensive value of the collision risk prediction degree.

  Moreover, although the case where the alarm sound is output when the total value of the collision risk prediction degree is equal to or greater than the threshold has been described as an example, the present invention is not limited thereto, and a sound may be output. The warning message may be presented by visual information. Or you may make it always show the comprehensive value of the estimated collision risk prediction degree by visual information.

  The program of the present invention can be provided by being stored in a recording medium.

10, 210 Collision risk estimation device 12 GPS sensor 14 Inter-vehicle distance measurement unit 16 Acceleration sensor 18 Brake sensor 20 Vehicle speed sensor 22, 222 Computer 26 Information acquisition unit 28 Information storage unit 30 Risk level determination unit 32 Information classification unit 34 Learning unit 36 Learning Result storage unit 38 Risk estimation unit 40 Alarm control unit 200 Collision risk estimation system 212 Server 244 Information classification unit 246 Learning unit 248 Learning result storage unit 250 Learning result selection unit

Claims (8)

A state detecting means for detecting a plurality of types of states including a driving state of the vehicle, an operating state when the driver operates the vehicle, or an environmental state indicating a driving environment around the vehicle;
A storage unit that stores the correspondence relationship between the state belonging to the classification and the risk prediction degree, which is learned for each classification of the plurality of types of states,
Based on the correspondence stored in the storage unit and the plurality of types of states detected by the state detection unit, the corresponding risk prediction degree is obtained for each classification, and the overall risk prediction degree is obtained. An estimation means for estimating a typical value;
Risk estimation device including   The correspondence relationship stored in the storage means is learned as learning data for each classification, a state belonging to the classification and a degree of risk obtained when the state is detected. The risk estimation apparatus according to 1.   Mutual information amount between the plural types of states detected when the degree of risk equal to or higher than a threshold is determined for the classification of the plural types of states, and the mutual information amount between the risk and each type of state The risk estimation device according to claim 2, wherein the risk level estimation device is classified based on at least one of mutual information between the plurality of types of states and a correlation coefficient between the plurality of types of states.   The risk estimation device according to claim 2 or 3, wherein the correspondence relationship stored in the storage means is learned by a learning device that learns the correspondence relationship based on the learning data.   The risk level estimation device according to claim 4, wherein the learning device further classifies the plurality of types of states.   The said estimation means estimates the total value of the said risk prediction degree by weighting and adding the said risk prediction degree calculated | required for every said classification, The claim 1 of any one of Claims 1-5. Risk estimation device.   The risk estimation according to any one of claims 1 to 5, wherein the estimation means estimates a maximum value of the risk prediction degree obtained for each classification as a comprehensive value of the risk prediction degree. apparatus. A state belonging to the classification, learned for each classification of a plurality of types of states including a traveling state of the vehicle, an operation state when the driver operates the vehicle, or an environmental state indicating a traveling environment around the vehicle, and A computer including a storage means storing a correspondence relationship with the degree of risk prediction;
Based on the correspondence stored in the storage unit and the plurality of types of states detected by the state detection unit detecting the plurality of types of states, the corresponding risk prediction degree is obtained for each classification. A program for functioning as an estimation means for estimating a comprehensive value of the risk prediction degree.

Images