WO2023067978A1

WO2023067978A1 - Leveling angle control system

Info

Publication number: WO2023067978A1
Application number: PCT/JP2022/035228
Authority: WO
Inventors: 佳典柴田
Original assignee: 株式会社小糸製作所
Priority date: 2021-10-20
Filing date: 2022-09-21
Publication date: 2023-04-27

Abstract

A leveling angle control system (100) for a vehicle headlight (30), comprising: a target leveling angle calculation unit (41) for calculating a target leveling angle θ for a vehicle headlight (30) at prescribed first point on the basis of point information pertaining to a prescribed second point that a vehicle (10) will reach upon travelling for a prescribed number of seconds or a prescribed distance from the prescribed first point; and a leveling angle control unit (42) for controlling the actual leveling angle at the first point such that the actual leveling angle of the vehicle headlight (30) approaches the target leveling angle θ.

Description

Leveling angle control system

The present disclosure relates to a leveling angle control system.

In recent years, headlights equipped with an auto-leveling function that automatically adjusts the vertical illumination range according to the vehicle's longitudinal tilt have become popular. For example, Patent Literature 1 discloses that the tilt angle of the vehicle is calculated by a gravity sensor and the optical axis of the headlight is controlled based on the tilt angle.

Japanese Patent Application Laid-Open No. 2000-85459

The technology disclosed in Patent Document 1 calculates the current tilt angle of the vehicle and adjusts the optical axis according to the current tilt angle. In such a place where the tilt angle changes abruptly, it is difficult to make the optical axis follow the abrupt change.

An object of the present disclosure is to appropriately change the optical axis of a vehicle headlamp in response to a sudden change in the inclination angle of a road even in a place where the inclination angle changes abruptly.

A leveling angle control system according to one aspect of the present disclosure includes:
A target leveling angle θ of a vehicle headlamp at a predetermined first point is obtained from point information of a predetermined second point reached by a vehicle traveling a predetermined number of seconds or a predetermined distance from the predetermined first point. a target leveling angle calculator that calculates based on
a leveling angle control unit that controls the actual leveling angle at the first point so that the actual leveling angle of the vehicle headlamp approaches the target leveling angle θ.

According to the present disclosure, it is possible to appropriately change the optical axis of the vehicle headlamp even in a place where the inclination angle of the road changes suddenly.

1 is a block diagram showing an example configuration of a leveling angle control system according to an embodiment of the present disclosure; FIG. FIG. 4 is a schematic diagram for explaining a measured angle of a vehicle; FIG. 4 is a schematic diagram showing an example of a method of acquiring information on a road surface angle using LiDAR; FIG. 4 is a schematic diagram showing an example of a method of acquiring information on a road surface angle using LiDAR; FIG. 4 is a schematic diagram showing an example of an image captured by a camera when the vehicle is heading uphill. FIG. 4 is a schematic diagram showing an example of an image captured by a camera when the vehicle is heading downhill; It is a schematic diagram for demonstrating an example of reinforcement learning. 4 is a flowchart showing an example of processing related to leveling angle control; It is an example of point data. 6 is a flowchart showing an example of processing related to reinforcement learning; 4 is a flowchart showing an example of processing related to calculation of a virtual leveling angle; 7 is a flowchart showing an example of processing related to leveling angle control during reinforcement learning;

Hereinafter, the present invention will be described based on the embodiments with reference to the drawings. The same or equivalent constituent elements and members shown in each drawing are denoted by the same reference numerals, and duplication of description will be omitted as appropriate. Also, the dimensions of each member shown in the drawings may differ from the actual dimensions of each member for convenience of explanation.

(Configuration of leveling angle control system)
First, a leveling angle control system according to an embodiment of the present disclosure will be described. The leveling angle control system according to the present embodiment is a system that controls the leveling angle at the current travel point based on point information of points ahead of the current travel point of the vehicle. FIG. 1 is a block diagram showing an example of the configuration of a leveling angle control system 100 (hereinafter also simply referred to as "system 100") according to this embodiment.

The system 100 is a system that controls the leveling angle of the vehicle headlights 30 . System 100 includes, for example, vehicle 10 and headlights 30 . The vehicle 10 includes, for example, a sensor section 11 and a vehicle control section 16 . Note that the sensor unit 11 may be provided in the headlamp 30 .

The sensor unit 11 includes, for example, a camera 12, a LiDAR (Light Detection And Ranging) 13, an acceleration sensor 14, and a position sensor 15. The camera 12 is provided so as to be able to image at least the front of the vehicle 10 . The LiDAR 13 is provided so as to acquire at least an image in front of the vehicle 10 . Data obtained by the camera 12 and the LiDAR 13 are output to the image processing unit 17, for example.

The acceleration sensor 14 is, for example, a three-axis acceleration sensor that detects acceleration in each direction of the mutually orthogonal x-axis, y-axis, and z-axis. The acceleration sensor 14 is attached to the vehicle 10 such that the x-axis is aligned with the longitudinal axis of the vehicle 10, the y-axis is aligned with the lateral axis of the vehicle 10, and the z-axis is aligned with the vertical axis of the vehicle 10. .

Based on the value measured by the acceleration sensor 14, it is possible to calculate the measured angle φ, which is the inclination angle of the vehicle 10 with respect to the horizontal plane. The measured angle φ is used, for example, for reinforcement learning of the learning model 52 described later. Further, the measured angle φ may be stored in the storage unit 50 in association with the position information, for example, and used for calculation of the target leveling angle θ by the target leveling angle calculation unit 41, which will be described later.

FIG. 2 is a schematic diagram for explaining the measured angle φ of the vehicle. The measured angle φ is the sum of the road surface angle θr, which is the inclination angle of the road surface with respect to the horizontal plane, and the vehicle angle θv, which is the inclination angle of the vehicle 10 with respect to the road surface. The acceleration sensor 14 detects, for example, a vector Gx, which is a detected value of the gravitational acceleration vector G in the x-axis direction, and a vector Gz, which is a detected value of the gravitational acceleration vector G in the z-axis direction. is used to calculate the measurement angle φ. Note that the calculation of the measurement angle φ is not limited to the above example, and other known methods may be used. Further, the calculation of the measured angle φ may be performed by the vehicle control unit 16 or the lamp control unit 40 described later based on the data detected by the acceleration sensor 14 .

Return to the description of Figure 1. The position sensor 15 is a sensor that acquires position information of the vehicle 10, and is, for example, a GPS (Global Positioning System) sensor or a GNSS (Global Navigation Satellite System) sensor. The position information of the vehicle 10 is stored as part of the location data 51 in the storage unit 50, for example.

The vehicle control unit 16 controls various operations such as traveling of the vehicle 10 . The vehicle control unit 16 includes, for example, a processor such as an ASIC (Application Specific Integrated Circuit), an FPGA (Field programmable Gate Array), or a general-purpose CPU (Central Processing Unit). Although not shown, the vehicle 10 includes, for example, a ROM (Read Only Memory) storing various vehicle control programs and a RAM (Random Access Memory) temporarily storing various vehicle control data. . The processor of the vehicle control unit 16 can load data designated from various vehicle control programs stored in the ROM onto the RAM and control various operations of the vehicle 10 in cooperation with the RAM.

In this embodiment, the vehicle control unit 16 functions as an image processing unit 17. Although details will be described later, the image processing unit 17, based on the data output from the camera 12 or the LiDAR 13, determines the point to which the vehicle 10 travels for a predetermined number of seconds or a predetermined distance from the current travel point. information can be identified.

The headlamp 30 is a lamp that is mounted on the vehicle 10 and illuminates the front of the vehicle 10 . The headlamp 30 includes, for example, a lamp control section 40, a storage section 50, and a leveling actuator 60. The lamp control unit 40 includes, for example, a processor such as ASIC, FPGA, or general-purpose CPU. The storage unit 50 is configured by, for example, a ROM, a RAM, or the like. The processor of the lamp control unit 40 can load data designated by the program stored in the ROM onto the RAM and control various operations of the headlamp 30 in cooperation with the RAM. Note that the storage unit 50 may be provided in the vehicle 10, or may be configured to be provided outside the vehicle 10 (for example, in a data center capable of communicating with the vehicle 10).

In the present embodiment, the lamp control unit 40 reads the program stored in the storage unit 50 to perform, for example, a target leveling angle calculation unit 41, a leveling angle control unit 42, a road surface angle information acquisition unit 43, and a learning processing unit. 44.

The target leveling angle calculation unit 41 calculates the target leveling angle θ of the headlight 30 at a predetermined first point from the first point by a predetermined number of seconds (eg, 1 second) or a predetermined distance (eg, 10 m) from the vehicle. It is calculated based on point information of a predetermined second point that 10 travels and reaches. Further, the target leveling angle calculator 41 may calculate the target leveling angle θ based on a learning model 52 described later obtained by reinforcement learning based on point information. The "point information" includes geographical information of the point and various types of information stored in association with the position information of the point (for example, point data 51 to be described later). The "point information" may include, for example, the measured angle φ, "information about the road surface angle θr" described later, the reference leveling angle, and the like.

The leveling angle control unit 42 controls the actual leveling angle at the first point so that the actual leveling angle of the headlamp 30 at the first point approaches the target leveling angle θ. The leveling angle control section 42 controls the actual leveling angle via the leveling actuator 60 .

The road surface angle information acquisition unit 43 acquires information regarding the road surface angle θr at the second point. The "information about the road surface angle θr" is not particularly limited, but is preferably, for example, information indicating whether the road surface is uphill or downhill or information indicating the road surface angle θr. These pieces of information can be acquired using the camera 12 or the LiDAR 13, for example.

Here, a method of obtaining information on the road surface angle θr will be described using FIGS. 3 to 6. FIG. FIG. 3 and FIG. 4 are schematic diagrams showing an example of a method of obtaining information on the road surface angle θr using the LiDAR 13. FIG. In the example of FIG. 3, the vehicle 10 is heading uphill. In this case, light emitted downward from the horizontal axis H of the LiDAR 13 (for example, light L3) always hits the ground E and is reflected. That is, when the front of the vehicle 10 slopes upward, the LiDAR 13 can detect the reflected light of all the light emitted downward from the horizontal axis H. Therefore, when the LiDAR 13 detects the reflected light of all the light emitted downward from the horizontal axis H, the road surface angle information acquisition unit 43 may be configured to determine that the second point is an upward slope. good. Note that the horizontal axis H is an axis parallel to the horizontal plane.

In the example of FIG. 3, part of the light emitted upward from the horizontal axis H of the LiDAR 13 (for example, light L2) hits the ground E and is reflected, but the other part (for example, light L1) , does not hit the ground E. That is, when the front of the vehicle 10 slopes upward, the LiDAR 13 detects only part of the light emitted upward from the horizontal axis H and reflected. Therefore, when the LiDAR 13 detects only part of the reflected light of the light emitted upward from the horizontal axis H, the road surface angle information acquisition unit 43 determines that the second point is an upward slope. may be configured.

In the example of FIG. 4, the vehicle 10 is heading downhill. In this case, the light emitted upward from the horizontal axis H (for example, the light L4) does not hit the ground E. That is, when the front of the vehicle 10 slopes down, the LiDAR 13 does not detect the reflected light of the light emitted upward from the horizontal axis H. Therefore, when the LiDAR 13 does not detect the reflected light of the light emitted upward from the horizontal axis H, the road surface angle information acquisition unit 43 may be configured to determine that the second point is a downward slope. good.

In the example of FIG. 4, part of the light emitted downward from the horizontal axis H (for example, light L6) hits the ground E and is reflected, but the other part (for example, light L5) is reflected by the ground. Does not hit E. That is, when the front of the vehicle 10 slopes downward, the LiDAR 13 detects only a part of the light emitted downward from the horizontal axis H as reflected light. Therefore, when the LiDAR 13 detects only part of the reflected light of the light emitted downward from the horizontal axis H, the road surface angle information acquisition unit 43 determines that the second point is downhill. may be configured.

In the examples of FIGS. 3 and 4, the road surface angle information acquisition unit 43 may calculate the road surface angle θr at the second point based on the three-dimensional image obtained by the LiDAR 13. A conventionally known image analysis method can be used without particular limitation for calculating the road surface angle θr.

Next, a method of obtaining information on the road surface angle θr using the camera 12 will be described. FIG. 5 is a schematic diagram showing an example of an image captured by the camera 12 when the vehicle 10 is heading uphill. In the example of FIG. 5 , the image acquired by the camera 12 includes a white or orange left line LL extending in the front-rear direction on the left side of the vehicle 10 and a front-rear line LL on the right side of the vehicle 10 as road markings that define the driving lane of the vehicle 10 . There is a white or orange right line RL extending in the direction.

When acquiring information about the road surface angle θr from the image captured by the camera 12, for example, the road surface angle information acquisition unit 43 uses image processing such as Hough transform to identify the left line LL and the right line RL. Next, the road surface angle information acquisition unit 43 determines whether or not at least one of the left line LL and the right line RL is curved. If at least one of the road surfaces is curved, the road surface angle information acquisition unit 43 determines the first vanishing point where the extension of one line closer to the vehicle 10 than the curve and the extension of the other line intersect, and from the curve. A second vanishing point at which an extension of one line far from the vehicle 10 intersects with an extension of the other line is identified. Then, when the angle formed by the left line LL and the right line RL having the first vanishing point as vertices is larger than the angle formed by the left line LL and the right line RL having the second vanishing point as vertices, road surface angle information The acquisition unit 43 determines that the second point is an upward slope.

In the example of FIG. 5, both the left line LL and the right line RL are bent with the line segment X as a boundary. In this case, an intersection point P1 between an extension of the left line LL closer to the vehicle 10 than the line segment X and an extension of the right line RL closer to the vehicle 10 than the line segment X is the first vanishing point. Similarly, an intersection point P2 between an extension of the left line LL farther from the vehicle 10 than the line segment X and an extension of the right line RL farther from the vehicle 10 than the line segment X is the second vanishing point. Then, the angle A formed between the left line LL and the right line RL having the vertex at the intersection point P1 and the angle B formed between the left line LL and the right line RL having the vertex at the intersection point P2 are compared, and "angle A > angle B", the road surface angle information acquiring unit 43 determines that the second point is an upward slope.

FIG. 6 is a schematic diagram showing an example of an image captured by the camera 12 when the vehicle 10 is heading downhill. In the example of FIG. 6, neither the left line LL nor the right line RL are curved, and the only specified vanishing point is the intersection point P3. In this way, when only one vanishing point can be specified, the road surface angle information acquiring unit 43 specifies the horizontal direction (horizontal direction ) is detected. In the example of FIG. 6, the line segment C is detected in the above range. In this case, the road surface angle information acquisition unit 43 determines that the second point is downward slope.

　In the examples of FIGS. 5 and 6, the road surface angle information acquisition unit 43 may calculate the road surface angle θr at the second point based on the image captured by the camera 12. For example, in the example of FIG. 5, the greater the slope of the uphill slope (the greater the road surface angle θr), the greater the distance between the first vanishing point and the second vanishing point in the image in the vertical direction. The road surface angle θr at the second point may be calculated from the separation distance in the vertical direction. Note that the method of calculating the road surface angle θr from the image captured by the camera 12 is not limited to the above example, and conventionally known methods can be used without particular limitations.

It should be noted that, with respect to the above-described method of acquiring information related to the road surface angle θr, it is preferable to combine two or more determination criteria and calculation methods as appropriate from the viewpoint of improving the accuracy of information to be acquired. Further, with respect to the method of acquiring information on the road surface angle θr described above, the image processing unit 17 may execute each process described as being executed by the road surface angle information acquisition unit 43 . In that case, the road surface angle information acquisition unit 43 may acquire information that has been determined, calculated, or the like by the image processing unit 17 .

Also, the information on the road surface angle θr may be configured to be calculated based on a machine-learned learning model. In this case, for example, an image captured by the camera 12 or a three-dimensional image acquired by the LiDAR 13 is input, and the second calculated based on the data detected by the acceleration sensor 14 when the vehicle 10 travels at the second point A learning model obtained by machine learning (for example, deep learning) using teacher data whose output is the measured angle φ or the road surface angle θr at two points can be used.

Return to the description of Figure 1. The learning processing unit 44 executes reinforcement learning for the learning model 52 . Reinforcement learning is repeatedly performed each time the vehicle 10 travels on a predetermined travel route including, for example, the first point and the second point. For example, the learning processing unit 44 gives a larger reward as the absolute value of the difference between the target leveling angle θ at the first point and the measured angle φ of the vehicle 10 at the second point measured by the acceleration sensor 14 is smaller. , is performed on the learning model 52 . The learning processing unit 44 performs, for example, Q-learning as reinforcement learning so that the Q-value increases at each point on the predetermined travel route. A specific example of the Q value will be explained in a later paragraph.

FIG. 7 is a schematic diagram for explaining an example of reinforcement learning. In the example of FIG. 7, points N−1, N, and N+1 are points on the travel route U. In the example of FIG. A point N is a point reached by the vehicle 10 traveling a predetermined number of seconds or a predetermined distance from the point N−1. Similarly, the point N+1 is a point reached by the vehicle 10 traveling from the point N for a predetermined number of seconds or a predetermined distance. The measured angles φ(N−1), φ(N), and φ(N+1) are the measured angles φ at the points N−1, N, and N+1, respectively.

Reinforcement learning is, for example, a system in which the smaller the absolute value of the difference between the target leveling angle θ at each point and the measured angle φ at the point next to each point, the larger the reward. is executed. In this reinforcement learning, the closer the target leveling angle θ(N−1) of the point N−1 is to the measured angle (φ) of the point N, the greater the reward for the point N−1. Similarly, the closer the target leveling angle θ(N) of the point N is to the measured angle (φ+1) of the point N+1, the greater the reward of the point N becomes.

As this type of reinforcement learning progresses, it becomes possible to control the leveling angle at the current travel point based on the measured angle φ at points ahead of the current travel point. Therefore, even if the tilt angle changes abruptly at the previous point, the optical axis can be appropriately changed in response to the sudden change in the tilt angle. As a result, even when the tilt angle changes abruptly, it is possible to suppress deterioration of forward visibility.

Return to the description of Figure 1. For example, the learning processing unit 44 sets a Q value comparison reference value and a reference leveling angle at each of a plurality of points on a predetermined travel route, and when the vehicle 10 travels on the predetermined travel route, the predetermined travel distance is set. Calculate the Q value at each of multiple points on the route, and if there is a point where the calculated Q value is greater than the comparison reference value, update the comparison reference value at that point to the calculated Q value Then, the reinforcement learning may be executed by updating the target leveling angle θ used in calculating the Q value as the reference leveling angle at the point. The comparison reference value and the reference leveling angle are stored in the storage unit 50 as point data 51 in association with the position information of each point, for example. Note that the initial value of the comparison reference value may be the same value at each point. Also, the initial value of the reference leveling angle may be the same value at each point, or the initial value may not be set.

The comparison reference value indicates the maximum value of the Q value at each point, and the reference leveling angle is the target leveling angle θ when the Q value indicates the maximum value at each point. The learning processing unit 44 may, for example, calculate the next target leveling angle θ based on the reference leveling angle. With such a configuration, it can be expected that the number of times of learning until the Q value converges at each point is reduced.

Further, when the vehicle 10 travels along a predetermined travel route, the learning processing unit 44 calculates a virtual leveling angle η at each of a plurality of points on the predetermined travel route, and converts the virtual leveling angle η to the target leveling angle. θ may be used to calculate the Q value, update the comparison reference value, and update the reference leveling angle. In this case, the leveling angle control unit 42 preferably does not control the actual leveling angle based on the target leveling angle θ at points where the comparison reference value does not exceed the predetermined threshold value. On the other hand, at points where the comparison reference value exceeds the predetermined threshold, it is preferable to use the reference leveling angle as the target leveling angle θ to actually control the leveling angle. The virtual leveling angle η can be calculated in the same manner as the target leveling angle θ.

Points with low comparison reference values are points where it is still difficult to control the appropriate leveling angle. Therefore, at such a point, the virtual leveling angle η is calculated instead of the target leveling angle θ, and reinforcement learning is performed using the virtual leveling angle η, while the actual leveling angle based on the point information of the previous point (For example, by configuring to perform control based on the measured angle φ of the current point as in the conventional method), for example, the leveling angle is randomly selected to an inappropriate value prevent it from being changed.

On the other hand, at points where the comparison reference value is high, the virtual leveling angle η is calculated instead of the target leveling angle θ, and reinforcement learning is performed using the virtual leveling angle η, thereby searching for the optimum leveling angle. By using the reference leveling angle as the target leveling angle θ, it is possible to control the leveling angle more appropriately than in the conventional configuration, even when the tilt angle changes abruptly at the previous point.

(Operation example of the leveling angle control system)
Next, an operation example of the system 100 according to this embodiment will be described with reference to FIGS. 8 to 12. FIG. It should be noted that the order of each process constituting each flowchart described below may be random as long as there is no contradiction or inconsistency in the contents of the process, and the processes may be executed in parallel.

FIG. 8 is a flowchart showing an example of processing related to leveling angle control. When a predetermined start condition is satisfied, the system 100 according to the present embodiment executes leveling angle control until a predetermined end condition is satisfied.

Although the predetermined start condition and end condition are not particularly limited, for example, the start condition is that the vehicle 10 has started traveling on a predetermined travel route, and the end condition is that the vehicle 10 has finished traveling on the predetermined travel route. may be In this case, the processes from step S2 to step S4, which will be described later, are repeatedly executed on a predetermined travel route.

The predetermined travel route may be set by the operation of the user of the vehicle 10, or may be set by the lamp control unit 40 by referring to the travel history of the vehicle 10 and setting a route that has been traveled frequently as the predetermined travel route. good. With such a configuration, it is possible to appropriately control the leveling angle on a route desired by the user or on a route frequently traveled by the user. It should be noted that the start and end of travel on the predetermined travel route may be determined, for example, based on the position information acquired by the position sensor 15, or may be determined based on the start operation and end operation of the vehicle 10 by the user. may

The predetermined start condition is that the absolute value of the difference between the road surface angle θr at the first point and the road surface angle θr at the second point is greater than or equal to a predetermined value, and the predetermined end condition is The absolute value of the difference between the road surface angle θr at the predetermined point and the road surface angle θr at the second point with respect to the predetermined point may be less than a predetermined value. According to such a configuration, in places where there is a large change in the inclination angle of the road, the leveling angle of the headlamp 30 can be appropriately changed in response to the change, and the change in the inclination angle of the road is small. In places, the load on the lamp control section 40 and the like can be reduced by controlling the leveling angle as in the conventional art. It should be noted that the processes from step S2 to step S4 may be repeatedly executed while the vehicle 10 is running without setting the predetermined start condition and end condition.

In the example of FIG. 8, if the start condition is not satisfied (No in step S1), it waits until the start condition is satisfied. If the start condition is satisfied (Yes in step S1), steps S2 to S4 are repeatedly executed until the end condition is satisfied.

In step S2, the lighting control unit 40 calculates the target leveling angle θ at the first point (current point of the vehicle 10). The target leveling angle θ at the first point is calculated based on the point information of the second point reached by the vehicle 10 traveling a predetermined number of seconds or a predetermined distance from the first point.

At step S2, for example, the target leveling angle θ can be calculated based on the information regarding the road surface angle θr at the second point. Specifically, when the information about the road surface angle θr is information indicating that the second point is an upward slope, the lighting control unit 40 calculates the target leveling angle θ so as to be larger than the measured angle φ at the first point. You may Conversely, if the information on the road surface angle θr indicates that the second point is on a downward slope, the lighting control unit 40 calculates the target leveling angle θ so as to be smaller than the measured angle φ at the first point. good too.

Further, when the information regarding the road surface angle θr indicates the value of the road surface angle θr at the second point, the lighting control section 40 may calculate the target leveling angle θ based on the road surface angle θr at the second point. . In this case, for example, a value obtained by correcting the road surface angle θr at the second point using the vehicle angle θv at the first point may be used as the target leveling angle θ at the first point. Further, when the measured angle φ at the second point is stored in the storage unit 50 as the point data 51, the lamp control unit 40 may adopt the measured angle φ at the second point as the target leveling angle θ.

Further, in step S2, the lamp control section 40 may calculate the target leveling angle θ of the first point based on the learning model 52 obtained by reinforcement learning based on the point information.

Next, in step S3, the lamp control unit 40 controls the leveling angle of the headlamp 30 so that the actual leveling angle at the first point approaches the target leveling angle θ. Through the processing in steps S2 and S3, control is performed to realize a leveling angle suitable for the inclination of the second point prior to arrival at the second point.

If the vehicle 10 has not traveled the predetermined distance from the first point or has not passed the predetermined time after passing the first point (No in step S4), it waits until it travels the predetermined distance or the predetermined time elapses. do. If a predetermined time has passed since the vehicle 10 traveled a predetermined distance from the first point or passed the first point (Yes in step S4), the position where the vehicle 10 is at that time is regarded as the first point, and the process proceeds to step S2. return. A series of processes from step S2 to step S4 are repeatedly executed until the end condition is satisfied, and the process related to the leveling angle control ends when the end condition is satisfied.

Next, a method of reinforcement learning for the learning model 52 by the system 100 will be described. First, an outline of reinforcement learning according to the present embodiment will be described. Reinforcement learning can be executed using, for example, an action-value function Q ^π (s, a) represented by the following formula (2) and a state-value function V ^π (s) represented by the following formula (3). can.

In the above equations (2) and (3), t is the time, s is the current state, s' is the next state, a is the action, and .pi. P and R are the probability that state s transitions to s' and the reward obtained at that time, respectively. γ indicates the discount rate of future rewards. The action-value function Q ^π (s, a) is the discount represents the reward sum. E is the expected value.

Here, we will explain the point information used for reinforcement learning. FIG. 9 is an example of the spot data 51. As shown in FIG. The point data 51 stores the measured angle φ, the comparison reference value, the reference leveling angle, and the virtual leveling angle η in association with the position information of each point on the predetermined travel route.

In this case, state s(N) at point N can be defined as follows, for example.
state s(N)=virtual leveling angle η(N)
Here, the virtual leveling angle η is a value that can be randomly set within a range of -3° to 2° with respect to the horizontal plane, and the resolution is 0.1°. Also, the initial value of the virtual leveling angle η is preferably −0.6°. Note that the above range, resolution, and initial value of the virtual leveling angle η are examples, and other numerical values may be adopted.

Action a is defined by changing the virtual leveling angle η(N) in the range of −3° to 2°. For the policy π, for example, the following can be adopted.
Point N+1 has a higher gradient than point N: The virtual leveling angle η(N) is randomly raised from the virtual leveling angle η(N−1) by 1° as the upper limit.
Point N+1 has a lower gradient than point N: The virtual leveling angle η(N) is randomly lowered from the virtual leveling angle η(N−1) by 1° as an upper limit.

The behavior a of this policy is represented by the following equation.
Action a(N)=Δη(N−1)=virtual leveling angle η(N)−virtual leveling angle η(N−1)

Reward R(N) can be defined, for example, as follows.
Reward R(N)=5°-|measured angle φ(N)-virtual leveling angle η(N-1)|
If the target value is controlled, the reward becomes maximum at 5°. If you deviate from the target value, the reward will be zero.
Also, the transition probabilities up to the next state s may similarly be assumed to be probable, or weighting may be performed centering on "-0.6° to the horizontal plane" which is the initial state.

Under the above premise, Q-learning, for example, can be used for reinforcement learning. Hereinafter, Q ^π (s, a) is also referred to as a Q value. With Q-learning, the system 100 attempts to optimize the leveling angle so that the further Q-value is always maximized. At that time, the search for the optimal action based on past experience and the new action aiming at reward acquisition is carried out according to the policy π as described above. As a result, learning model 52 learns the states and actions that maximize the Q value. By replacing the virtual leveling angle η(N) that maximizes the Q value with the target leveling angle θ(N), the target leveling angle θ is optimized.

The following is a more specific explanation of reinforcement learning. FIG. 10 is a flowchart illustrating an example of processing related to reinforcement learning. The reinforcement learning shown in FIG. 9 is repeatedly executed when the vehicle 10 travels on a predetermined travel route. As a result, a learning model 52 capable of appropriate leveling angle control on a predetermined travel route is obtained.

First, in step S11, the lamp control unit 40 sets the upper limit number of times of reinforcement learning. For the upper limit number of times, for example, the number of times that the Q value exceeds a predetermined threshold value may be set. Next, in step S12, the lamp control unit 40 detects the start of travel on a predetermined travel route. The predetermined travel route is a specific route set in advance as a target of reinforcement learning. Detection of the predetermined travel route may be performed, for example, based on the position information acquired from the position sensor 15, or may be performed based on the user's start operation of the vehicle 10, or the like.

Next, in step S13, the lamp control section 40 starts timing. Timing is performed to determine when to perform reinforcement learning. After that, a series of processes from step S14 to step S21 are repeated until the lamp control unit 40 detects the end of traveling on the predetermined travel route.

In step S14, the lamp control unit 40 determines whether or not it is time for reinforcement learning. In the example of FIG. 10, for example, it is determined that it is time for reinforcement learning each time a predetermined number of seconds (for example, one second) elapses from the start of timing in step S13. The timing of reinforcement learning may be set every time the vehicle 10 travels a predetermined distance. In this case, in step S13, measurement of traveled distance is started instead of timing.

If it is not the timing for reinforcement learning (No in step S14), wait until the timing for reinforcement learning. If it is the timing of reinforcement learning (Yes in step S14), in step S15, the lamp control unit 40 determines whether or not the traveling speed of the vehicle 10 is equal to or higher than a predetermined speed (eg, 30 km/h). If the running speed of the vehicle 10 is less than the predetermined speed (No in step S15), the process returns to step S14.

Even in a place where the inclination angle of the road suddenly changes, if the traveling speed of the vehicle 10 is slow, it is difficult to make the optical axis of the headlight 30 follow the sudden change even with the conventional configuration. do not have. Therefore, the system 100 according to this embodiment is particularly useful when the vehicle 10 travels at a high speed. To obtain a learning model 52 that is particularly useful when the vehicle 10 travels at a high speed by constructing the learning model 52 so that reinforcement learning is not executed with data when the vehicle 10 travels at a speed less than a predetermined speed. can be done.

When the running speed of the vehicle 10 is equal to or higher than the predetermined speed (Yes in step S15), in step S16, the lamp control unit 40 adjusts the virtual leveling angle η at the first point (current point) based on a predetermined measure. calculate. In the following description, the first point is point N, the second point reached by the vehicle 10 traveling a predetermined number of seconds from the first point is point N+1, and the vehicle 10 is traveling a predetermined number of seconds before the first point. The point at which it was located is also referred to as point N-1.

Here, the processing of step S16 will be described in detail using FIG. FIG. 11 is a flowchart showing an example of processing related to calculation of the virtual leveling angle η, and shows a specific example of the processing in step S16. First, in step S31, the lamp control unit 40 acquires information about the road surface angle θr at the point N+1. Information about the road surface angle θr can be obtained based on the image captured by the camera 12, the three-dimensional image obtained by the LiDAR 13, or the like, as described above.

If it is determined from the information acquired in step S31 that the point N+1 has a higher slope than the point N (Yes in step S32), in step S33 the lamp control unit 40 sets the virtual leveling angle η(N ) is randomly increased from the initial value or the virtual leveling angle η(N−1) with an upper limit of 1°. Further, when it is determined from the information acquired in step S31 that the point N+1 has a lower gradient than the point N (No in step S32), in step S34, the lamp control unit 40 sets the virtual leveling angle η (N) is randomly lowered from the initial value or the hypothetical leveling angle η(N−1) at the point N−1 with an upper limit of 1° down. With such a configuration, it can be expected that the number of times of learning until the Q value exceeds a predetermined threshold is reduced.

Note that when the value of the road surface angle θr at the point N+1 is acquired in step S31, or when the number of times the predetermined travel route has been traveled is two or more times, the measured angle φ(N+1) at the point N+1 is calculated in step S31. When the value is acquired, the virtual leveling angle η(N) may be calculated based on the value of the road surface angle θr and the value of the measured angle φ(N+1).

After steps S33 and S34, proceed to step S17 in FIG. In step S17, the lamp control section 40 acquires the measured angle φ(N+1) at the point N+1 that was measured upon arrival at the point N+1. Next, in step S18, the lamp control section 40 calculates the Q value at the point N based on the virtual leveling angle η(N) and the measured angle φ(N+1).

If the Q value calculated in step S18 is greater than the comparison reference value of point N included in the point data 51 (Yes in step S19), the lamp control unit 40 determines the value of point N in the point data 51 in step S20. The comparison reference value is updated to the Q value calculated in step S18. Also, in step S21, the lamp control unit 40 updates the reference leveling angle of the point N in the point data 51 to the value of the virtual leveling angle η(N) calculated in step S16. On the other hand, when the Q value calculated in step S18 is equal to or less than the comparison reference value of point N included in the point data 51 (No in step S19), the process returns to step S14.

A series of processes from step S14 to step S21 are repeatedly executed on a predetermined travel route. Upon detecting the end of travel on the predetermined travel route, the lamp control unit 40 updates the number of times of learning in step S22. When the number of times of learning reaches the upper limit number of times set in step S11 (Yes in step S23), the process related to reinforcement learning ends. If the number of times of learning has not reached the upper limit number of times set in step S11 (No in step S23), the process returns to step S12 to continue the process related to reinforcement learning.

When the processing related to reinforcement learning ends, it becomes possible to calculate the target leveling angle θ based on the obtained learning model 52, and it becomes possible to control the leveling angle based on the target leveling angle θ. . Further, even when the processing related to reinforcement learning is continued, for example, when a predetermined condition is satisfied, the target leveling angle θ calculated based on the learning model 52 at that time is used to determine the leveling angle. It is preferably arranged to control.

FIG. 12 is a flowchart showing an example of processing related to leveling angle control during reinforcement learning. Each process shown in the flowchart shown in FIG. 12 is performed in parallel with a series of processes from step S14 to step S20 in FIG. 10, for example.

First, in step S41, the lamp control unit 40 detects that the vehicle 10 has arrived at the point N. Next, in step S42, the lamp control section 40 reads the comparison reference value at the point N from the point data 51, and determines whether or not the comparison reference value is equal to or greater than a predetermined threshold.

If the comparison reference value for the point N is equal to or greater than the predetermined threshold (Yes in step S42), in step S43, the lamp control unit 40 reads the reference leveling angle at the point N from the point data 51, and converts it to the read reference leveling angle. Based on this, the target leveling angle θ(N) is calculated. In step S43, for example, the value of the read reference leveling angle is set as the target leveling angle θ(N).

Next, in step S44, the lamp control unit 40 controls the actual leveling angle at the point N based on the target leveling angle θ(N) calculated in step S43, and ends the process. On the other hand, if the comparison reference value of the point N is less than the predetermined threshold value (No in step S42), the processing of steps S43 and S44 is not executed and the process ends.

Even during reinforcement learning, at points where reinforcement learning has progressed to some extent, it is possible to change the optical axis of the headlight 30 more appropriately than in the conventional configuration in response to a sudden change in road gradient. , the learning model 52 is preferably used for leveling angle control. On the other hand, at points where reinforcement learning has not progressed, the accuracy of the target leveling angle θ calculated by the learning model 52 is not sufficient. Preferably, no actual leveling angle control is performed.

It should be noted that the present invention is not limited to the above-described embodiments, and can be modified, improved, etc. as appropriate. In addition, the material, shape, size, numerical value, form, number, location, etc. of each component in the above-described embodiment are arbitrary and not limited as long as the present invention can be achieved.

For example, the processing explained to be executed by the lamp control unit 40 in the description of each flowchart may be executed by the vehicle control unit 16 as long as there is no contradiction. Moreover, each data stored in the storage unit 50 may be stored in the storage unit of the vehicle 10 .

Also, the processing related to reinforcement learning may be executed in a server device that can communicate with the vehicle 10 . In this case, for example, position information of a plurality of positions on a predetermined travel route and the measured angles φ of each of the plurality of positions are transmitted from the vehicle 10 to the server device. Then, the server device executes reinforcement learning for the learning model 52 , and the obtained learning model 52 is transmitted to the vehicle 10 and stored in the storage unit 50 .

This application is based on a Japanese patent application (Japanese Patent Application No. 2021-171946) filed on October 20, 2021, the contents of which are incorporated herein by reference.

10: Vehicle 11: Sensor Unit 12: Camera 13: LiDAR
14: Acceleration sensor 15: Position sensor 16: Vehicle control unit 17: Image processing unit 30: Headlight (for vehicle) 40: Lamp control unit 41: Target leveling angle calculation unit 42: Leveling angle control unit 43: Road surface angle information Acquisition unit 44: Learning processing unit 50: Storage unit 51: Point data 52: Learning model 60: Leveling actuator 100: Leveling angle control system

Claims

A target leveling angle θ of a vehicle headlamp at a predetermined first point is obtained from point information of a predetermined second point reached by a vehicle traveling a predetermined number of seconds or a predetermined distance from the predetermined first point. a target leveling angle calculator that calculates based on
a leveling angle control unit that controls the actual leveling angle at the first point so that the actual leveling angle of the vehicle headlight approaches the target leveling angle θ;
Leveling angle control system for vehicle headlights.
The calculation of the target leveling angle θ and the control of the actual leveling angle are repeatedly executed on a predetermined travel route,
The target leveling angle calculator provides a larger reward as the absolute value of the difference between the target leveling angle θ at the first point and the measured angle φ of the vehicle at the second point measured by an acceleration sensor is smaller. Calculate the target leveling angle θ at the first point based on a learning model obtained by executing reinforcement learning set to provide
The leveling angle control system according to claim 1.
a learning processing unit that executes Q-learning as the reinforcement learning for the learning model;
a storage unit that stores comparison reference values of the Q value at each of a plurality of points on the predetermined travel route,
The learning processing unit calculates a Q value at each of a plurality of points on the predetermined travel route when the vehicle travels on the predetermined travel route, and the calculated Q value is higher than the comparison reference value. When there is a point with a large value, the comparison reference value at that point is updated to the calculated Q value, and the target leveling angle θ used when calculating the calculated Q value is set to the reference leveling angle Execute the reinforcement learning by storing in association with the point as
The target leveling angle calculator calculates the target leveling angle θ based on the reference leveling angle.
3. The leveling angle control system according to claim 2.
When the vehicle travels along the predetermined travel route, the learning processing unit calculates a virtual leveling angle η at each of a plurality of points on the predetermined travel route, and converts the virtual leveling angle η to the target leveling angle. Calculating the Q value, updating the comparison reference value, and storing the reference leveling angle using the angle θ,
wherein the leveling angle control unit does not control the actual leveling angle based on the target leveling angle θ at a point where the comparison reference value does not exceed a predetermined threshold;
4. The leveling angle control system according to claim 3.
The reinforcement learning is not executed when the running speed of the vehicle is less than or equal to a predetermined speed.
3. The leveling angle control system according to claim 2.
Furthermore, a road surface angle information acquisition unit that acquires information about the road surface angle at the second point,
The point information includes information about the road surface angle,
The target leveling angle θ or the virtual leveling angle η at the first point is calculated based on information about the road surface angle,
A leveling angle control system according to any one of claims 1 to 5.
The vehicle is equipped with LiDAR,
The road surface angle information acquisition unit,
When the LiDAR detects the reflected light of all light emitted downward from the horizontal axis of the LiDAR at the first point, or when the LiDAR detects the reflected light at the first point above the horizontal axis When only a part of the reflected light is detected for the light irradiated toward it, it is determined that the second point is an upward slope,
When the LiDAR detects only part of the reflected light of the light emitted downward from the horizontal axis at the first point, or when the LiDAR detects the light emitted downward from the horizontal axis at the first point Determining that the second point is a downward slope when reflected light of the light emitted upward is not detected,
7. A leveling angle control system according to claim 6.
The vehicle is equipped with a camera,
The road surface angle information acquisition unit identifies a vanishing point in the image acquired by the camera at the first point, and acquires information about the road surface angle at the second point based on the vanishing point.
7. A leveling angle control system according to claim 6.
The calculation of the target leveling angle θ and the control of the actual leveling angle are executed when the absolute value of the difference between the road surface angle at the first point and the road surface angle at the second point is greater than or equal to a predetermined value. ,
7. A leveling angle control system according to claim 6.