Estimation device, estimation method, and program

Abstract

To estimate behavior of the other moving body with high accuracy when a moving body merges with the other moving body.SOLUTION: An estimation device, which estimates movement of the other moving body with respect to a moving body, is provided with an obtaining part that obtains a prediction parameter that is a second parameter, by inputting an actual parameter that is a first parameter to a first learnt model which has learnt, as leaning data, the time-series first parameter representing a relation between the moving body and the other moving body in a first situation in which the other moving body has performed predetermined movement and has learnt, as correct-answer data, the second parameter representing a state of the other moving body in the first situation and to a second learnt model which has learnt, as learning data, the first parameter in a second situation in which the other moving body has not performed the predetermined movement and has learnt, as correct-answer data, the second parameter in the second situation; and an estimating part that estimates whether the other moving body performs the predetermined movement or not, by comparing the prediction parameter with the actual parameter at a target time of the prediction parameter.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an estimation device, an estimation method, and a program. [Background technology]

[0002] Conventionally, there are known techniques for controlling the travel of a moving object when the moving object changes lanes from the driving lane to an adjacent lane. For example, Patent Document 1 describes a technique for making a moving object merge between a second vehicle and a third vehicle in accordance with a control policy learned based on passively collected data related to the operation of the vehicle. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2019-31268 Summary of the Invention [Problem to be solved by the invention]

[0004] The technology described in Patent Document 1 learns a control policy by applying a passive actor-critic reinforcement learning method to passively collected data related to vehicle operation. However, in the conventional technology, data on other moving objects with which a moving object merges may not be effectively utilized. As a result, there are cases where the behavior of other moving objects when a moving object merges cannot be estimated with high accuracy.

[0005] The present invention has been made in consideration of the above circumstances, and one of its objectives is to provide an estimation device, an estimation method, and a program that can estimate with high accuracy the behavior of other moving bodies when a moving body merges with another moving body by effectively utilizing data on the other moving bodies with which the moving body is to merge. [Means for solving the problem]

[0006] The estimation device, estimation method, and program according to the present invention employ the following configuration. (1): An estimation device according to one embodiment of the present invention is an estimation device that estimates the behavior of another moving body relative to a moving body, and includes: a first trained model that uses a first parameter of a time series representing a relationship between the moving body and the other moving body in a first scene in which the other moving body performed a predetermined behavior as learning data, and a second parameter representing the state of the other moving body in the first scene as correct answer data, and that has been trained to output distribution information of the second parameter when the first parameter is input; an acquisition unit that acquires a predicted parameter, which is the second parameter, by inputting an actual parameter, which is the first parameter, into a second trained model that uses the first parameter in a second scene in which the other moving body did not perform the predetermined behavior as learning data, and the second parameter in the second scene as correct answer data, and that has been trained to output distribution information of the second parameter when the first parameter is input; and an estimation unit that estimates whether the other moving body will perform the predetermined behavior by comparing the predicted parameter with the actual parameter at a target time point of the predicted parameter.

[0007] (2): In the aspect (1) above, the first parameters include at least the relative speed between the other moving body and the moving body, the relative position between the other moving body and the moving body, and the speed of the moving body.

[0008] (3): In the above aspect (1), the distribution information of the second parameter output from the first trained model is first distribution information regarding the relative speed and relative position between the other moving body and the moving body predicted when the other moving body performs the specified action, and the distribution information of the second parameter output from the second trained model is second distribution information regarding the relative speed and relative position between the other moving body and the moving body predicted when the other moving body does not perform the specified action.

[0009] (4): In the above aspect (3), the estimation unit repeatedly estimates whether or not the other moving body will perform the specified behavior as a probability value, and the estimation unit estimates the probability value at the current time point based on the probability value at the previous time point when the estimation was made, the first distribution information at the current time point when the estimation is made, and the second distribution information at the current time point when the estimation is made.

[0010] (5): In the aspect (4) above, the estimation unit estimates the probability value at the first time point in the repeated estimation using a third trained model that has been trained to output a probability value indicating whether or not the other moving body will perform the specified behavior when a parameter representing the relationship between the moving body and the other moving body is input.

[0011] (6) In the aspect (4) above, the estimation device further includes a driving assistance unit that assists driving of the moving object based on the probability value.

[0012] (7): In the above aspects (1) to (6), the predetermined behavior is a behavior that indicates that, when the moving body changes lanes from the driving lane to an adjacent lane, the other moving body traveling in the adjacent lane allows the moving body to go ahead.

[0013] (8): Another aspect of the present invention is an estimation method executed by a computer that estimates the behavior of another moving body relative to a moving body, in which the computer inputs an actual parameter, which is the first parameter, into a first trained model, the first parameter being training data, which is a time series representing the relationship between the moving body and the other moving body in a first scene in which the other moving body performed a predetermined behavior, and a second parameter being correct data, which is a time series representing the state of the other moving body in the first scene, and being trained to output distribution information of the second parameter when the first parameter is input, and a second trained model, the first parameter being training data, which is the first parameter in a second scene in which the other moving body did not perform the predetermined behavior, and the second parameter in the second scene being correct data, and being trained to output distribution information of the second parameter when the first parameter is input, to obtain a predicted parameter, which is the second parameter, and estimates whether the other moving body will perform the predetermined behavior by comparing the predicted parameter with the actual parameter at the target time of the predicted parameter.

[0014] (9): Another aspect of the present invention provides a program executed by a computer that estimates the behavior of another moving body relative to a moving body, the program causing the computer to input an actual parameter, which is the first parameter, into a first trained model that uses as learning data a first parameter of a time series representing the relationship between the moving body and the other moving body in a first scene in which the other moving body performed a predetermined behavior, and as correct data a second parameter representing the state of the other moving body in the first scene, and that has been trained to output distribution information of the second parameter when the first parameter is input, and a second trained model that uses as learning data the first parameter in a second scene in which the other moving body did not perform the predetermined behavior, and as correct data the second parameter in the second scene, and that has been trained to output distribution information of the second parameter when the first parameter is input, and to obtain a predicted parameter, which is the second parameter, and to estimate whether the other moving body will perform the predetermined behavior by comparing the predicted parameter with the actual parameter at the target time of the predicted parameter. [Effects of the Invention]

[0015] According to aspects (1) to (9), by effectively utilizing data on other moving bodies that the moving body is to merge with, it is possible to estimate with high accuracy the behavior of the other moving bodies when the moving body merges with the other moving body. [Brief explanation of the drawings]

[0016] [Figure 1] 1 is a configuration diagram of a vehicle M on which a driving assistance device 100 according to an embodiment is mounted. [Figure 2] 10 is a diagram showing an example of a merging lane recognized by a merging recognition unit 110. FIG. [Figure 3] A figure showing an example of the configuration of training data and correct answer data used to generate the first trained model 152 and the second trained model 154. [Figure 4] A figure showing an example of the input / output configuration of a first trained model 152 and a second trained model 154 generated by learning. [Figure 5] 10 is a diagram showing an example of distribution information of prediction parameters acquired by the parameter acquisition section 120. FIG. [Figure 6] 10 is a flowchart illustrating an example of a flow of processing executed by the estimation device according to the present embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0017] Hereinafter, with reference to the drawings, an embodiment of an estimation device, an estimation method, and a program of the present invention will be described. The estimation device according to this embodiment estimates whether another vehicle will perform a predetermined action by comparing a predicted distribution of the state of the other vehicle output by a first trained model generated based on learning data collected under the assumption that the other vehicle has performed a predetermined action with an actual parameter of the other vehicle. The predicted distribution is output by a first trained model generated based on learning data collected under the assumption that the other vehicle has not performed a predetermined action, and a second trained model generated based on learning data collected under the assumption that the other vehicle has not performed a predetermined action. Hereinafter, as an example, the estimation device according to this embodiment will be described as being applied to estimate whether another vehicle traveling in a to-be-merged lane will yield to the other vehicle as a predetermined action when merging.

[0018] [Overall configuration] FIG. 1 is a configuration diagram of a vehicle M on which a driving assistance device 100 according to an embodiment is mounted. The vehicle M may be, for example, a two-wheeled, three-wheeled, or four-wheeled vehicle, and its drive source may be an internal combustion engine such as a diesel engine or a gasoline engine, an electric motor, or a combination of these. The electric motor operates using power generated by a generator connected to the internal combustion engine, or discharged power from a secondary battery or a fuel cell. The vehicle M is an example of a "mobile body."

[0019] The vehicle M is equipped with, for example, a camera 10, a radar device 12, a LIDAR (Light Detection and Ranging) 14, an object recognition device 16, an HMI (Human Machine Interface) 30, vehicle sensors 40, a navigation device 50, a driver monitor camera 60, a driving operator 70, a driving assistance device 100, a driving force output device 200, a braking device 210, and a steering device 220. These devices and equipment are connected to each other by multiplexed communication lines such as a CAN (Controller Area Network) communication line, serial communication lines, a wireless communication network, etc. Note that the configuration shown in FIG. 1 is merely an example, and some of the configuration may be omitted, or other configurations may be added.

[0020] The camera 10 is, for example, a digital camera using a solid-state imaging element such as a CCD (Charge Coupled Device) or a CMOS (Complementary Metal Oxide Semiconductor). The camera 10 is attached to any location of a vehicle (hereinafter referred to as the host vehicle M) in which the vehicle system 1 is installed. When capturing an image of the front, the camera 10 is attached to the top of the front windshield, the back of the rearview mirror, or the like. The camera 10, for example, periodically and repeatedly captures images of the surroundings of the host vehicle M. The camera 10 may be a stereo camera.

[0021] The radar device 12 emits radio waves such as millimeter waves around the vehicle M and detects radio waves reflected by an object (reflected waves) to detect at least the position (distance and direction) of the object. The radar device 12 is attached to any location on the vehicle M. The radar device 12 may detect the position and speed of an object using an FM-CW (Frequency Modulated Continuous Wave) method.

[0022] The LIDAR 14 irradiates the surroundings of the vehicle M with light (or electromagnetic waves with wavelengths similar to light) and measures the scattered light. The LIDAR 14 detects the distance to the target based on the time between light emission and light reception. The irradiated light is, for example, pulsed laser light. The LIDAR 14 is attached to any location on the vehicle M.

[0023] The object recognition device 16 performs sensor fusion processing on the detection results from some or all of the camera 10, the radar device 12, and the LIDAR 14 to recognize the position, type, speed, etc. of the object. The object recognition device 16 outputs the recognition results to the driving assistance device 100. The object recognition device 16 may output the detection results from the camera 10, the radar device 12, and the LIDAR 14 directly to the driving assistance device 100. The object recognition device 16 may be omitted from the vehicle system 1.

[0024] The HMI 30 presents various information to the occupants of the vehicle M and accepts input operations by the occupants. The HMI 30 includes various display devices, a speaker, a buzzer, a vibration generator (vibrator), a touch panel, switches, keys, and the like.

[0025] The vehicle sensor 40 includes a vehicle speed sensor that detects the speed of the host vehicle M, an acceleration sensor that detects acceleration, a yaw rate sensor that detects angular velocity around a vertical axis, a direction sensor that detects the direction of the host vehicle M, and the like.

[0026] The navigation device 50 includes, for example, a GNSS (Global Navigation Satellite System) receiver, a guidance control unit, and a storage unit storing map information. The GNSS receiver identifies the position of the vehicle M based on signals received from GNSS satellites. The position of the vehicle M may be identified or supplemented by an INS (Inertial Navigation System) that uses the output of the vehicle sensors 40. The guidance control unit, for example, determines a route from the position of the vehicle M identified by the GNSS receiver (or an arbitrary input position) to a destination input by the occupant by referring to map information, and causes the HMI 30 to output guidance information so that the vehicle M travels along the route. The map information is, for example, information that represents road shapes using links indicating roads and nodes connected by the links. The map information may include road curvature, POI (Point of Interest) information, and the like. The navigation device 50 may transmit the current position and destination of the vehicle M to a navigation server via a communication device and acquire the route from the navigation server.

[0027] The driver monitor camera 60 is a digital camera that uses a solid-state imaging element such as a CCD or CMOS. The driver monitor camera 60 is attached to any location on the vehicle M in a position and orientation that allows it to capture an image of the head of an occupant sitting in the driver's seat of the vehicle M from the front. The driver monitor camera 60 captures an image of the interior of the vehicle M, including the driver, from its installed position and outputs the image to the driving assistance device 100.

[0028] The driving operators 70 include, for example, an accelerator pedal, a brake pedal, a steering wheel, a shift lever, and other operators. The driving operators 70 are fitted with sensors that detect the amount of operation or the presence or absence of operation, and the detection results are output to some or all of the driving force output device 200, the braking device 210, and the steering device 220.

[0029] The driving force output device 200 outputs a driving force (torque) for the vehicle to travel to the driving wheels. The driving force output device 200 includes, for example, a combination of an internal combustion engine, an electric motor, a transmission, etc., and an ECU (Electronic Control Unit) that controls these. The ECU controls the above components according to information input from the driving assistance device 100 or information input from the driving operator 70.

[0030] Braking device 210 includes, for example, a brake caliper, a cylinder that transmits hydraulic pressure to the brake caliper, an electric motor that generates hydraulic pressure in the cylinder, and an ECU. The ECU controls the electric motor according to information input from driving assistance device 100 or information input from driving operator 70, so that a brake torque corresponding to the braking operation is output to each wheel. Braking device 210 may include a backup mechanism that transmits hydraulic pressure generated by operation of a brake pedal included in driving operator 70 to the cylinder via a master cylinder. Note that braking device 210 is not limited to the configuration described above, and may also be an electronically controlled hydraulic brake device that controls an actuator according to information input from driving assistance device 100 to transmit hydraulic pressure from a master cylinder to the cylinder.

[0031] The steering device 220 includes, for example, a steering ECU and an electric motor. The electric motor applies a force to a rack and pinion mechanism to change the direction of the steered wheels. The steering ECU drives the electric motor to change the direction of the steered wheels in accordance with information input from the driving assistance device 100 or information input from the driving operator 70.

[0032] [Driving assistance devices] The driving assistance device 100 includes, for example, a merging recognition unit 110, a parameter acquisition unit 120, a merging estimation unit 130, a driving assistance unit 140, and a storage unit 150. The merging recognition unit 110, the parameter acquisition unit 120, the merging estimation unit 130, and the driving assistance unit 140 are realized by, for example, a hardware processor such as a CPU (Central Processing Unit) executing a program (software). Furthermore, some or all of these components may be realized by hardware (including circuitry) such as an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a GPU (Graphics Processing Unit), or may be realized by a combination of software and hardware. The program may be stored in advance in a storage device (a storage device having a non-transitory storage medium) such as the HDD or flash memory of the driving assistance device 100, or may be stored in a removable storage medium such as a DVD or CD-ROM, and installed in the HDD or flash memory of the driving assistance device 100 by inserting the storage medium (non-transitory storage medium) into a drive device.

[0033] The storage unit 150 is realized by, for example, a read-only memory (ROM), a flash memory, an SD card, a random access memory (RAM), a hard disk drive (HDD), a register, etc. The storage unit 150 stores, for example, a first trained model 152, a second trained model 154, and a third trained model 156. An apparatus including at least the parameter acquisition unit 120, the confluence estimation unit 130, and the first trained model 152 and the second trained model 154 stored in the storage unit 150 is an example of an "estimation device." Furthermore, in this embodiment, the first trained model 152, the second trained model 154, and the third trained model 156 are stored in the memory unit 150 and are used by the functional units of the driving assistance device 100, but alternatively, some or all of these trained models may be stored, for example, on a cloud outside the driving assistance device 100 and used by the driving assistance device 100 via a network.

[0034] The merging recognition unit 110 recognizes that the lane in which the host vehicle M is traveling is a merging lane that requires a lane change to an adjacent lane. FIG. 2 is a diagram showing an example of a merging lane recognized by the merging recognition unit 110. FIG. 2 shows that the lane L1 in which the host vehicle M is traveling is a merging lane that requires a lane change to an adjacent lane L2. For example, the merging recognition unit 110 can recognize that the lane in which the host vehicle M is traveling is a merging lane when it recognizes, based on road dividing line information recognized by the camera 10, that the width of the lane in which the host vehicle M is traveling decreases to zero in the direction of travel and that a lane change to an adjacent lane is possible. Furthermore, for example, the merging recognition unit 110 can recognize that the lane in which the host vehicle M is traveling is a merging lane by using the navigation device 50 to compare the current position of the host vehicle M identified by the GNSS receiver with map information.

[0035] When the merging recognition unit 110 recognizes that the driving lane of the host vehicle M is a merging lane, the merging recognition unit 110 then determines whether or not another vehicle is present in an adjacent lane (i.e., a lane to be merged) into which the host vehicle M will merge. More specifically, for example, the merging recognition unit 110 determines whether or not another vehicle is present within a predetermined distance d (i.e., within a predetermined range) in the extension direction of the adjacent lane, with the center of gravity C of the host vehicle M as a reference, based on information recognized by the object recognition device 16. If it is determined that no other vehicle is present within the predetermined distance d, the driving assistance unit 140 causes, for example, the HMI 30 to display recommendation information recommending that the host vehicle M change lanes from the driving lane to the adjacent lane. On the other hand, if it is determined that another vehicle is present within the predetermined distance d, the merging recognition unit 110 recognizes the other vehicle as a processing target for which processing will be executed by the parameter acquisition unit 120 and the merging estimation unit 130, which will be described below. In the case of the scene shown in FIG. 2, the merging recognition unit 110 recognizes the other vehicle M1 as a processing target for which processing is to be executed by the parameter acquisition unit 120 and the merging estimation unit .

[0036] The parameter acquisition unit 120 acquires, in a time series (at a predetermined control cycle), first parameters representing a relationship between the host vehicle M and another vehicle for the other vehicle recognized as a processing target by the merging recognition unit 110. The first parameters include, for example, at least the relative speed between the other vehicle and the host vehicle M, the relative position between the other vehicle and the host vehicle M, and the speed of the host vehicle M. For example, the parameter acquisition unit 120 can calculate the relative speed and relative position between the other vehicle and the host vehicle M based on images of the other vehicle captured in time series by the camera 10. Furthermore, for example, the parameter acquisition unit 120 can acquire the speed of the host vehicle M from the vehicle sensor 40. In the case of the scene shown in FIG. 2, the merging recognition unit 110 acquires, for the other vehicle M1, the relative speed between the other vehicle M1 and the host vehicle M, the relative position between the other vehicle M1 and the host vehicle M, and the speed of the host vehicle M. Hereinafter, the relative speed between the other vehicle and the host vehicle M and the relative position between the other vehicle and the host vehicle M may be referred to as the "relative speed of the other vehicle" and the "relative position of the other vehicle", respectively.

[0037] When the parameter acquisition unit 120 acquires first parameters for another vehicle recognized as a processing target by the merging recognition unit 110, the parameter acquisition unit 120 inputs the acquired first parameters into the first trained model 152 and the second trained model 154 to acquire distribution information of second parameters representing the state of the other vehicle. Here, the second parameters representing the state of the other vehicle represent, for example, the relative speed and relative position of the other vehicle in the next step of the control cycle. The first trained model 152 and the second trained model 154 will be described in detail below.

[0038] FIG. 3 is a diagram showing an example of the configuration of training data and supervised data used to generate the first trained model 152 and the second trained model 154. In the table shown in FIG. 3, the input corresponds to the training data, and the output corresponds to the supervised data. The symbol s represents the state of the other vehicle. More specifically, the state s of the input training data represents the relative speed and relative position of the other vehicle and the speed of the host vehicle M, and the state s of the output supervised data represents the relative speed and relative position of the other vehicle. The symbol t represents a time point in a control cycle. For example, the record for t=5 in the table shown in FIG. 3 indicates that data representing the state s5 of the other vehicle at time point 5 in the control cycle is associated with data representing the states s0 to s4 of the other vehicle from time points 0 to 4 in the control cycle.

[0039] Based on the above data configuration, the learning data and correct answer data used to generate the first trained model 152 are data collected on the assumption that, when the host vehicle M changes lanes from the driving lane to an adjacent lane, another vehicle traveling in the adjacent lane intentionally performs an action indicating that it allows the host vehicle M to proceed (i.e., a yielding action). For example, in the scene shown in FIG. 2, the subject driving the other vehicle M1 intentionally performs a yielding action in response to the host vehicle M's approach to the adjacent lane, and the state data of the other vehicle acquired when the intentional yielding action was performed is stored as the learning data and correct answer data. Here, the intentional yielding action means, for example, the subject's action of slowing down the other vehicle M1 in response to the host vehicle M's approach to the adjacent lane to create space between the other vehicle M1 and the leading vehicle M2. The yielding action is an example of a "predetermined action."

[0040] On the other hand, the training data and correct answer data used to generate the second trained model 154 are data collected on the assumption that when the host vehicle M changes lanes from the driving lane to an adjacent lane, another vehicle traveling in the adjacent lane takes an action indicating that it does not allow the host vehicle M to proceed (i.e., a behavior of not yielding to the right of way). For example, in the scene shown in FIG. 2, the subject driving the other vehicle M1 intentionally takes an action of not yielding to the right of way in response to the host vehicle M's approach to the adjacent lane, and the state data of the other vehicle acquired when the subject took the intentional behavior of not yielding to the right of way is stored as the training data and correct answer data. Here, the behavior of intentionally not yielding to the right of way means, for example, the subject's action of accelerating (or not decelerating) the other vehicle M1 in response to the host vehicle M's approach to the adjacent lane to close the space between the other vehicle M1 and the leading vehicle M2.

[0041] In this way, the first trained model 152 is generated based on training data that stores state data of other vehicles acquired when the subject intentionally performs an action of yielding to others, and correct answer data, and the second trained model 154 is generated based on training data that stores state data of other vehicles acquired when the subject intentionally performs an action of not yielding to others, and correct answer data. Using these training data and correct answer data, learning is performed using a method such as LSTM (Long Short Term Memory), which is a type of RNN (Recurrent Neural Network) that takes time series into consideration, to generate the first trained model 152 and the second trained model 154.

[0042] 4 is a diagram showing an example of the input / output configuration of the first trained model 152 and the second trained model 154 generated by learning. As shown in FIG. 4, the first trained model 152 and the second trained model 154 output a predicted distribution regarding the relative speed and relative position of the other vehicle in the next step, in response to input of the relative speed and relative position of the other vehicle in the current step in the control cycle and the speed of the host vehicle M. More specifically, in this embodiment, it is assumed that the relative speed and relative position of the other vehicle in the next step follow a normal distribution, and the first trained model 152 and the second trained model 154 are trained to output the expectation value and variance of the normal distribution representing the relative speed and relative position of the other vehicle in the next step.

[0043] In other words, the first trained model 152 is distribution information indicating what relative speed and relative position the other vehicle is likely to assume in the next step if the other vehicle yields to the host vehicle M. On the other hand, the second trained model 154 is distribution information indicating what relative speed and relative position the other vehicle is likely to assume in the next step if the other vehicle does not yield to the host vehicle M. The relative speed and relative position of the other vehicle in the current step in the control cycle, and the speed of the host vehicle M, which are input to the first trained model 152 and the second trained model 154, are examples of "actual parameters," and the expected value and variance of the normal distribution representing the relative speed and relative position of the other vehicle in the next step, which are output from the first trained model 152 and the second trained model 154, are examples of "predicted parameters."

[0044] FIG. 5 is a diagram illustrating an example of distribution information of predicted parameters acquired by the parameter acquisition unit 120. As an example, FIG. 5 illustrates an example in which the first trained model 152 and the second trained model 154 each output distribution information regarding relative positions of the predicted parameters as a normal distribution. The expected value of the normal distribution output at this time represents the expected value of the destination of the other vehicle, and the variance represents the individual difference and sensor error (i.e., including the error of the camera 10 and the error of the vehicle sensor 40) regarding the movement of the other vehicle. The merging estimation unit 130 estimates whether the other vehicle will yield by comparing the predicted parameters acquired by the parameter acquisition unit 120 with the actual parameters measured at the target time of the predicted parameters (i.e., the actual parameters measured in the next step). The estimation process performed by the merging estimation unit 130 will be described in detail below.

[0045] First, the merging estimation unit 130 estimates the probability at an initial point in time (i.e., an initial probability) of whether or not the other vehicle recognized as a processing target by the merging recognition unit 110 will yield, using the third trained model 156. The third trained model 156 is a trained model that has been trained to output, when a first parameter representing the relationship between the host vehicle M and the other vehicle is input, a probability value indicating whether or not the other vehicle will yield. The training data used to generate the third trained model 156 is data of the first parameter at a predetermined point in time (i.e., non-time series), and the correct answer data is information indicating whether or not the other vehicle has yielded, assuming the state of the first parameter at the predetermined point in time (e.g., 1 if it has yielded, 0 if it has not). Unlike the first trained model 152 and the second trained model 154, the training data and correct answer data used to generate the third trained model 156 are data collected by a subject driving a vehicle so as to yield or not yield based on his or her own judgment. Hereinafter, P(yield = 1|s0) represents the initial probability that another vehicle will yield, output by the third trained model 156 in response to the input of the first parameter s0 at the initial time point 0, and P(yield = 0|s0) represents the initial probability that another vehicle will not yield, output by the third trained model 156 in response to the input of the first parameter s0 at the initial time point 0.

[0046] In this case, if the initial probability output by the third trained model 156 is equal to or greater than a threshold value (e.g., 0.8), the driving assistance unit 140 may, for example, cause the HMI 30 to display recommendation information recommending that the host vehicle M change lanes from the driving lane to an adjacent lane, without the merging estimation unit 130 performing processing using the first trained model 152 and the second trained model 154. Also, for example, if the initial probability output by the third trained model 156 is less than another threshold value (e.g., 0.2), the driving assistance unit 140 may, for example, cause the HMI 30 to display recommendation information recommending that the host vehicle M not change lanes from the driving lane to an adjacent lane, without the merging estimation unit 130 performing processing using the first trained model 152 and the second trained model 154.

[0047] Next, after calculating the initial probability for the other vehicle, the merging estimation unit 130 updates the probability that the other vehicle will yield in time series (i.e., for time point k (k=1, 2, . . .)) by Bayesian estimation according to the following procedure. More specifically, the merging estimation unit 130 inputs the first parameter s0 at time point 0 to the first trained model 152 and the second trained model 154, and obtains the predicted distributions p(s1|yield=1, s0) and p(s1|yield=0, s0) at time point 1. The merging estimation unit 130 then obtains the actual probability value by substituting the actual parameters (i.e., the relative position and relative speed of the other vehicle) measured at time point 1 for the obtained predicted distributions p(s1|yield=1, s0) and p(s1|yield=0, s0). 5 shows the probability value obtained by substituting the relative position of the other vehicle M1 at time point 1, and in the case of FIG. 5, as an example, it shows a situation where p(s1|yield=1, s0)>p(s1|yield=0, s0) holds. In this way, the "comparison" of the predicted parameters and the actual parameters described above means that the actual parameters actually measured in the next step are substituted into the predictive distribution defined by the predicted parameters to obtain the probability value.

[0048] Next, the merging estimation unit 130 calculates α=p(s1|yield=1, s0)P(yield=1|s0) by multiplying the previous probability value P(yield=1|s0) that the other vehicle will yield by the current step probability value p(s1|yield=1, s0) that the other vehicle will yield, and calculates β=p(s1|yield=1, s0)P(yield=0|s0) by multiplying the previous probability value P(yield=0|s0) that the other vehicle will not yield by the current step probability value p(s1|yield=0, s0). The values ​​of α and β mean the probability values ​​of the current step taking into account the previous probability values. Next, the merging estimation unit 130 normalizes the values ​​of α and β as probability values, and calculates the probability value P(yield=1|s0) at time point 1 taking into account the state s0 at time point 0. 0:1 ) = α / (α + β) and P(yield = 0|s 0:1 )=β / (α+β). The merging estimation unit 130 calculates the obtained probability value P(yield=1|s0:1 ) and P(yield = 0|s 0:1 ) is used as the previous value in estimating the probability value for time point 2.

[0049] The probability value calculation process described above can be generalized to calculate the probability value by using the state s 0:t-1 The previous probability value P(yield = 1|s 0:t-1 ) and the previous probability value P(yield = 0|s 0:t-1 ) and the state s at time t-1 t-1 Based on this, if the other vehicle were to give way, the predicted distribution p(s t |Given=1,s t-1 ) is calculated, and if the other vehicle does not give way, a predicted distribution p(s t |yield=0,s t-1 Next, the merging estimation unit 130 calculates α=p(s t |Given=1,s t-1 )P(yield=1|s 0:t-1 ) and β=p(s t |Given=1,s t-1 )P(yield=1|s 0:t-1 ) and normalize the values ​​of α and β to obtain the probability value P(yield = 1|s 0:t ) = α / (α + β) and the probability value P(yield = 0|s 0:t )=β / (α+β). This allows for accurate estimation of whether or not the other vehicle will yield to the right of way, taking into account the time-series state of the other vehicle.

[0050] The driving support unit 140 calculates the probability value P(yield=1|s) of the current step that another vehicle will yield, which is updated in time series by the merging estimation unit 130. 0:t ) is greater than or equal to the threshold, and the probability value P(yield = 1|s 0:t) is determined to be equal to or greater than the threshold, the HMI 30 displays recommendation information that recommends that the host vehicle M change lanes from the driving lane to the adjacent lane. In addition, for example, the driving support unit 140 calculates the probability value P(yield=0|s 0:t ) is greater than or equal to the threshold, and the probability value P(yield = 0|s 0:t ) is determined to be equal to or greater than the threshold, the HMI 30 may display recommendation information recommending that the host vehicle M not change lanes from the driving lane to an adjacent lane.

[0051] Next, the flow of processing executed by the estimation device according to this embodiment will be described with reference to Fig. 6. Fig. 6 is a flowchart showing an example of the flow of processing executed by the estimation device according to this embodiment. The processing of the flowchart shown in Fig. 6 is repeatedly executed while the host vehicle M is traveling.

[0052] First, the merging recognition unit 110 determines whether or not it has recognized that the lane in which the host vehicle M is traveling is a merging lane, for example, based on road dividing line information recognized by the camera 10 (step S100). If it is determined that it has not recognized that the lane in which the host vehicle M is traveling is a merging lane, the merging recognition unit 110 returns the process to step S100. On the other hand, if it is determined that it has recognized that the lane in which the host vehicle M is traveling is a merging lane, the merging recognition unit 110 determines whether or not it has recognized another vehicle within a predetermined range of the merging lane (step S102). If it is determined that it has not recognized another vehicle within the predetermined range of the merging lane, the merging recognition unit 110 proceeds to step S114, which will be described later.

[0053] On the other hand, if it is determined that another vehicle has been recognized within a predetermined range of the merging lane, the parameter acquisition unit 120 acquires first parameters of the recognized other vehicle (i.e., the relative speed between the other vehicle and the host vehicle M, the relative position between the other vehicle and the host vehicle M, and the speed of the host vehicle M) (step S104). Next, the merging estimation unit 130 inputs the acquired first parameters into the third trained model 156 to acquire an initial probability that the other vehicle will yield to the other vehicle (step S106).

[0054] Next, the driving assistance unit 140 determines whether the acquired initial probability is equal to or greater than a threshold (step S108). If it is determined that the acquired initial probability is equal to or greater than the threshold, the driving assistance unit 140 proceeds to step S114. On the other hand, if it is determined that the acquired initial probability is less than the threshold, the merging estimation unit 130 calculates a probability value of the current step that another vehicle will give way, using the first trained model 152, the second trained model 154, and the probability value of the previous step (the initial probability if this is the first time) (step S110).

[0055] Next, the driving assistance unit 140 determines whether the calculated probability value of the current step is equal to or greater than a threshold (step S112). If it is determined that the calculated probability value of the current step is equal to or greater than the threshold, the driving assistance unit 140 causes the HMI 30 to display recommendation information recommending that the host vehicle M change lanes from the driving lane to an adjacent lane (step S114). On the other hand, if it is determined that the calculated probability value of the current step is less than the threshold, the merging estimation unit 130 sets the calculated probability value of the current step as the probability value of the previous step and executes the process of step S110 again. This ends the process of this flowchart.

[0056] In the above embodiment, the driving assistance unit 140 assists the driving of the host vehicle M based on the probability value estimated in time series by the merging estimation unit 130. However, the present invention is not limited to such a configuration, and the present invention can also be applied to autonomous driving of the host vehicle M. For example, the driving assistance device 100 may include a driving control unit instead of the driving assistance unit 140, and the driving control unit may control the steering so that the host vehicle M changes lanes when the estimated probability value is equal to or greater than a threshold.

[0057] Furthermore, in the above embodiment, the predetermined behavior is estimated as whether another vehicle will yield to the other vehicle when the host vehicle M merges. However, the present invention is not limited to such a configuration, and the present invention can be applied to control other than merging. For example, the present invention can also be applied to the predetermined behavior of estimating whether an oncoming vehicle will yield to the other vehicle when the host vehicle M turns right or left at an intersection. In this case, the first trained model 152 is generated based on training data and correct answer data collected under the assumption that the oncoming vehicle yields to the other vehicle, and the second trained model 154 is generated based on training data and correct answer data collected under the assumption that the oncoming vehicle does not yield to the other vehicle. Furthermore, for example, the present invention can also be applied to the predetermined behavior of estimating whether an oncoming vehicle will yield to the other vehicle when the host vehicle M is stopped in a parking lot. In this case, the first trained model 152 is generated based on training data and correct answer data collected under the assumption that the oncoming vehicle yields to the other vehicle, and the second trained model 154 is generated based on training data and correct answer data collected under the assumption that the oncoming vehicle does not yield to the other vehicle.

[0058] Furthermore, in the above embodiment, the estimation device is mounted on the host vehicle M. However, the present invention is not limited to such a configuration. For example, the function of the estimation device may be mounted on an infrastructure camera installed in a merging lane to monitor vehicles. In this case, in the flowchart of FIG. 6, step S100 is omitted, and if the distance between a vehicle traveling in the merging lane and another vehicle traveling in the merging lane is within a predetermined range in the processing of step S102, the processing from step S104 to step S112 is executed. If it is determined in step S112 that the probability that the other vehicle will yield is equal to or greater than a threshold, the infrastructure camera may wirelessly transmit information recommending a lane change to the vehicle traveling in the merging lane.

[0059] According to the present embodiment described above, when a vehicle merges, whether or not another vehicle will yield is estimated by comparing the predicted distribution of the state of the other vehicle output by the first trained model generated based on training data collected under the assumption that the other vehicle yields to the other vehicle and the second trained model generated based on training data collected under the assumption that the other vehicle does not yield to the other vehicle with the actual parameters of the other vehicle. This makes it possible to estimate with high accuracy the behavior of the other moving object when the moving object merges by effectively utilizing data on the other moving object with which the moving object is to merge.

[0060] The above-described embodiment can be expressed as follows. a storage medium storing computer-readable instructions for estimating the behavior of other moving objects relative to the moving object; a processor connected to the storage medium; The processor executes the computer-readable instructions to: a first trained model that is trained to output a first parameter representing a time series of a relationship between the moving object and the other moving object in a first scene in which the other moving object performs a predetermined action as training data and a second parameter representing a state of the other moving object in the first scene as correct answer data when the first parameter is input; a second trained model that is trained to output the second parameter when the first parameter is input, using the first parameter in a second scene in which the other moving object did not perform the predetermined action as training data and the second parameter in the second scene as correct answer data; By inputting the actual parameters as the first parameters, the predicted parameters as the second parameters are obtained; by comparing the predicted parameters with actual parameters of the other moving object at the time of outputting the predicted parameters, it is estimated whether the other moving object will perform the predetermined action; The estimation device is configured as follows.

[0061] The above describes the form for carrying out the present invention using an embodiment, but the present invention is not limited to such an embodiment, and various modifications and substitutions can be made within the scope that does not deviate from the gist of the present invention. [Explanation of symbols]

[0062] 10 Camera 12 Radar equipment 14 LIDAR 16 Object recognition device 30 HMI 40 Vehicle Sensors 50 Navigation equipment 60 Driver monitor camera 70 Driving controls 100 Driving assistance device 110 Confluence recognition unit 120 Parameter acquisition unit 130 Confluence Estimation Unit 140 Driving Support Department 150 Storage section 152 First trained model 154 Second trained model 156 Third trained model 200 Driving force output device 210 Brake equipment 220 Steering device

Claims

1. An estimation device that estimates the behavior of another moving object relative to a moving object, a first trained model that is trained to output distribution information of a second parameter when the first parameter is input, the first parameter being a time series representing a relationship between the moving object and the other moving object in a first scene in which the other moving object performed a predetermined action as training data and the second parameter being a correct answer data representing a state of the other moving object in the first scene; a second trained model that is trained to output distribution information of the second parameter when the first parameter is input, using the first parameter in a second scene in which the other moving object did not perform the predetermined action as training data and the second parameter in the second scene as correct answer data; an acquisition unit that acquires a predicted parameter that is distribution information of the second parameter by inputting an actual parameter that is the first parameter; an estimation unit that estimates whether the other moving object will perform the predetermined behavior by comparing the predicted parameter with an actual parameter at a target time point of the predicted parameter; the estimation unit repeatedly estimates whether the other moving object will perform the predetermined action as a probability value, the estimation unit estimates the probability value at the current time point based on the probability value at the previous time point when the estimation was performed, distribution information of the second parameter output from the first trained model at the current time point when the estimation is performed, and distribution information of the second parameter output from the second trained model at the current time point when the estimation is performed; Estimation device.

2. the first parameters include at least a relative speed between the other moving body and the moving body, a relative position between the other moving body and the moving body, and a speed of the moving body; The estimation device according to claim 1 .

3. The distribution information of the second parameter output from the first trained model is first distribution information regarding the relative speed and relative position between the other moving body and the moving body predicted when the other moving body performs the predetermined action, and the distribution information of the second parameter output from the second trained model is second distribution information regarding the relative speed and relative position between the other moving body and the moving body predicted when the other moving body does not perform the predetermined action. The estimation device according to claim 1 .

4. the estimation unit estimates the probability value at an initial time point in the repeated estimation using a third trained model that has been trained to output, when a parameter representing a relationship between the moving body and the other moving body is input, a probability value indicating whether or not the other moving body will perform the predetermined behavior. The estimation device according to claim 1 .

5. a driving assistance unit that assists the driving of the moving object based on the probability value; The estimation device according to claim 1 .

6. The predetermined behavior is a behavior that indicates, when the moving body changes lanes from a driving lane to an adjacent lane, that the other moving body traveling in the adjacent lane allows the moving body to go ahead. The estimation device according to any one of claims 1 to 5.

7. An estimation method executed by a computer for estimating the behavior of another moving object relative to a moving object, the method comprising: a first trained model that is trained to output distribution information of a second parameter when the first parameter is input, the first parameter being a time series representing a relationship between the moving object and the other moving object in a first scene in which the other moving object performed a predetermined action as training data and the second parameter being a correct answer data representing a state of the other moving object in the first scene; a second trained model that is trained to output distribution information of the second parameter when the first parameter is input, using the first parameter in a second scene in which the other moving object did not perform the predetermined action as training data and the second parameter in the second scene as correct answer data; By inputting the actual parameter as the first parameter, a predicted parameter as distribution information of the second parameter is obtained; By comparing the predicted parameters with actual parameters at a target time point of the predicted parameters, it is estimated whether the other moving object will perform the predetermined behavior; the estimation is performed by repeatedly estimating, as a probability value, whether or not the other moving object will perform the predetermined behavior; The probability value at the current time point is estimated based on the probability value at the previous time point when the estimation was performed, distribution information of the second parameter output from the first trained model at the current time point when the estimation is performed, and distribution information of the second parameter output from the second trained model at the current time point when the estimation is performed. Estimation method.

8. A program executed by a computer for estimating the behavior of another moving object relative to a moving object, the program comprising: a first trained model that is trained to output distribution information of a second parameter when the first parameter is input, the first parameter being a time series representing a relationship between the moving object and the other moving object in a first scene in which the other moving object performed a predetermined action as training data and the second parameter being a correct answer data representing a state of the other moving object in the first scene; a second trained model that is trained to output distribution information of the second parameter when the first parameter is input, using the first parameter in a second scene in which the other moving object did not perform the predetermined action as training data and the second parameter in the second scene as correct answer data; obtaining a predicted parameter as the second parameter by inputting an actual parameter as the first parameter; by comparing the predicted parameters with actual parameters at a target time point of the predicted parameters, it is estimated whether or not the other moving object will perform the predetermined behavior; the estimation is performed by repeatedly estimating, as a probability value, whether or not the other moving object will perform the predetermined behavior; The probability value at the current time point is estimated based on the probability value at the previous time point when the estimation was performed, distribution information of the second parameter output from the first trained model at the current time point when the estimation is performed, and distribution information of the second parameter output from the second trained model at the current time point when the estimation is performed. program.

Images