KR102600238B1 - Method and apparatus for autonomous driving for unstructured environments - Google Patents

Abstract

The present invention relates to an autonomous driving method and device for an unstructured driving environment. The present invention relates to a look-ahead point using imitation learning so that a vehicle can head to a drivable area recognized based on vision. Point) and provide an autonomous driving method and device that provides deep learning-based motion-planning.

Description

Autonomous driving method and apparatus for unstructured driving environment {METHOD AND APPARATUS FOR AUTONOMOUS DRIVING FOR UNSTRUCTURED ENVIRONMENTS}

The present invention relates to an autonomous driving method and device for an unstructured driving environment. The present invention relates to a look-ahead point using imitation learning so that a vehicle can head to a drivable area recognized based on vision. It relates to an autonomous driving method and device that selects a point and provides deep learning-based motion-planning.

The content described below is merely for the purpose of providing background information related to embodiments of the present invention, and does not necessarily constitute prior art.

Recently, research on autonomous driving in unstructured driving environments is in progress. Autonomous driving in an unstructured driving environment has greater changes in the width and curvature of the drivable area than in a structured driving environment, and is more complicated and difficult due to the presence of obstacles of various sizes and shapes.

Conventional research for driving in unstructured environments includes optimization based on global paths, graph search, and tree expansion.

However, such a method i) takes a long time to calculate when the environment is complex or large, ii) it is not clear to determine which point in the global path should be selected as the destination of the avoidance path, and iii) the Global positioning system ( If localization data obtained through GPS (GPS) or simultaneous localization and mapping (SLAM) is not accurate, it is difficult to follow the path accurately, and there is a risk of colliding with obstacles.

Meanwhile, there are methods such as candidate path selection and artificial field generation by model-based motion-planning.

However, the parameters in each objective function must be adjusted differently depending on the diverse and complex environment, and if the boundary of the drivable area is inaccurately recognized or a shadow is misrecognized as a drivable area, vehicle movement may become unstable or There is a problem that increases the possibility of colliding with obstacles.

Autonomous driving technology for unstructured driving environments that can address these limitations is needed.

Meanwhile, the above-mentioned prior art is technical information that the inventor possessed for deriving the present invention or acquired in the process of deriving the present invention, and cannot necessarily be said to be known art disclosed to the general public before the application for the present invention. .

One object of the present invention is to provide an autonomous driving method and device for an unstructured driving environment based on deep learning.

One object of the present invention is to provide an autonomous driving method and device for an unstructured driving environment that uses a look-ahead point for a drivable area recognized based on vision.

The object of the present invention is not limited to the problems mentioned above, and other objects and advantages of the present invention that are not mentioned can be understood through the following description and will be more clearly understood through examples of the present invention. It will also be appreciated that the objects and advantages of the present invention can be realized by means and combinations thereof as set forth in the claims.

An autonomous driving method for an unstructured driving environment according to an embodiment of the present invention includes determining a drivable area of a vehicle from an input image, and candidate look-ahead points for the drivable area using a driving policy network. determining a look-ahead point corresponding to a target arrival point of the vehicle based on the candidate look-ahead point and a reference look-ahead point for the drivable area, and the driving based on the determined look-ahead point. It may include training the policy network.

An autonomous driving device for an unstructured driving environment according to an embodiment of the present invention includes a processor and a memory storing at least one instruction, wherein the at least one instruction, when executed by the processor, causes the processor to: Determine the drivable area of the vehicle from the input image, determine candidate lookahead points for the drivable area using a driving policy network, and determine the vehicle's drivable area based on the candidate lookahead point and the reference lookahead point for the drivable area. It may be configured to determine a look-ahead point corresponding to the target arrival point and train a driving policy network based on the determined look-ahead point.

Other aspects, features, and advantages in addition to those described above will become apparent from the following drawings, claims, and detailed description of the invention.

According to an embodiment, autonomous driving performance in an unstructured driving environment can be improved by training a deep neural network with driving data acquired in various and complex environments.

According to the embodiment, by using the look-ahead point as an imitation learning action, the imitation learning iterative algorithm can be applied, and effective and accurate autonomous driving imitation is possible.

According to an embodiment, the number of repetitions of imitation learning can be reduced by reflecting the difference between the candidate lookahead point determined by the driving policy network and the reference lookahead point in training.

This allows learning about more diverse environments with the same data generation time and effort, thereby improving the safety of autonomous driving.

The effects of the present invention are not limited to those mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

1 is a schematic flow diagram of autonomous driving for an unstructured driving environment according to an embodiment.
Figure 2 is a block diagram of an autonomous driving device for an unstructured driving environment according to an embodiment.
3 is a flowchart of an autonomous driving method for an unstructured driving environment according to an embodiment.
Figure 4 is an exemplary configuration diagram of a driving environment recognition network according to an embodiment.
Figure 5 is an exemplary configuration diagram of a driving policy network according to an embodiment.
Figure 6 is an example pseudocode of one step of an autonomous driving method for an unstructured driving environment according to an embodiment.
7 is an example pseudocode of one step of an autonomous driving method for an unstructured driving environment according to an embodiment.
FIG. 8 is an example diagram illustrating lookahead point determination by an autonomous driving method for an unstructured driving environment according to an embodiment.
9 is an example pseudocode of one step of an autonomous driving method for an unstructured driving environment according to an embodiment.
Figure 10 is an example diagram for explaining learning data update according to an embodiment.

Hereinafter, the present invention will be described in more detail with reference to the drawings. The present invention may be implemented in many different forms and is not limited to the embodiments described herein. In the following embodiments, parts that are not directly related to the description are omitted in order to clearly explain the present invention, but this does not mean that such omitted elements are unnecessary when implementing a device or system to which the spirit of the present invention is applied. . In addition, the same reference numbers are used for identical or similar components throughout the specification.

In the following description, terms such as first, second, etc. may be used to describe various components, but the components should not be limited by the terms, and the terms may be used to separate one component from another component. It is used only for distinguishing purposes. Additionally, in the following description, singular expressions include plural expressions, unless the context clearly indicates otherwise.

In the following description, terms such as “comprise” or “have” are intended to designate the presence of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification, but are intended to indicate the presence of one or more other It should be understood that this does not exclude in advance the possibility of the presence or addition of features, numbers, steps, operations, components, parts, or combinations thereof.

The present invention provides an autonomous driving method using deep learning-based motion-planning to select a look-ahead point using imitation learning so that the vehicle can head to a drivable area recognized by vision alone for autonomous driving in an unstructured environment. and devices.

The present invention is a method of overcoming the limitations of model-based motion-planning mentioned above and is a method of selecting a look-ahead point for driving toward a drivable area while avoiding obstacles in real time. Autonomous driving performance in unstructured driving environments is improved without the need for high-precision global maps such as 3D-point clouds, global routes composed of waypoint units, and localization through GPS or SLAM.

The present invention will be described in detail below with reference to the drawings.

1 is a schematic flow diagram of autonomous driving for an unstructured driving environment according to an embodiment.

The road on which vehicles travel has structured driving environment information such as lanes, center lines, traffic lights, and signs. An unstructured driving environment is a place where there is no such structured driving environment information, and the unstructured driving environment largely includes information about the drivable area and obstacles.

Referring to FIG. 1 , an autonomous driving process for an unstructured driving environment according to an embodiment will be schematically described.

Autonomous driving according to an embodiment captures the driving environment using a camera consisting of a lens (for example, two lenses) attached to the top of the vehicle. Here, the camera captures the front driving direction of the vehicle. Additionally, the camera can capture images of the rear or sides of the vehicle.

The captured image is converted into a bird’s eye view, which is a view from above. The converted image is divided into drivable images and non-drivable images through a supervised learning process. The classified images are converted into an Occupancy Grid Map and used as input for a Driving Policy trained through imitation learning.

The driving policy determines the look-ahead point, which is the point the vehicle must reach. A pure pursuit algorithm can be used to calculate the steering angle and velocity of the vehicle to reach the look-ahead point. The determined steering angle and speed are transmitted to the autonomous vehicle.

Figure 2 is a block diagram of an autonomous driving device for an unstructured driving environment according to an embodiment.

The autonomous driving device 100 for an unstructured driving environment according to an embodiment includes a processor 110 and a memory 120 that stores at least one instruction.

The processor 110 is a type of central processing unit and can execute an autonomous driving method according to an embodiment by executing one or more instructions stored in the memory 120. The processor 110 may include all types of devices capable of processing operations on data.

The processor 110 may mean, for example, a data processing device built into hardware that has a physically structured circuit to perform a function expressed by code or instructions included in a program. Examples of data processing devices built into hardware include a microprocessor, central processing unit (CPU), processor core, multiprocessor, and application-specific integrated (ASIC). circuit), FPGA (field programmable gate array), and graphics processing unit (Graphics Processing Unit (GPU)), but is not limited thereto.

Processor 110 may include one or more processors. For example, processor 110 may include a CPU and GPU. For example, the processor 110 may include a plurality of GPUs. Processor 110 may include at least one core.

The memory 120 may store commands for executing an autonomous driving method according to an embodiment. The memory 120 may store an executable program that generates and executes one or more instructions implementing a method according to an embodiment. The processor 110 may execute an autonomous driving method according to an embodiment based on programs and instructions stored in the memory 120.

Memory 120 may include internal memory and/or external memory, such as volatile memory such as DRAM, SRAM, or SDRAM, one time programmable ROM (OTPROM), PROM, EPROM, EEPROM, mask ROM, flash ROM, and NAND. Non-volatile memory such as flash memory or NOR flash memory, flash drives such as SSD, compact flash (CF) card, SD card, Micro-SD card, Mini-SD card, Xd card, or memory stick, etc. Alternatively, it may include a storage device such as an HDD. The memory 120 may include, but is not limited to, magnetic storage media or flash storage media.

In one example, the autonomous driving device 100 for an unstructured driving environment may be mounted on a vehicle. In one example, the autonomous driving device 100 for an unstructured driving environment is implemented as a server device outside the vehicle, receives an input image, executes a method according to an embodiment, and provides look-ahead point information or look-ahead point information. Vehicle control information (for example, steering angle and speed information) determined according to can be transmitted to the vehicle or a device that controls the vehicle.

At least one command stored in the memory 120, when executed by the processor 110, causes the processor 110 to determine the drivable area of the vehicle from the input image and to determine the drivable area using the driving policy network. Determine a candidate look-ahead point, determine a look-ahead point corresponding to the target arrival point of the vehicle based on the candidate look-ahead point and a reference look-ahead point for the drivable area, and determine the driving policy based on the determined look-ahead point. It can be configured to train a network.

At least one instruction stored in the memory 120, when executed by the processor 110, is configured to cause the processor 110 to execute a drivable area recognition network based on the input image to determine the drivable area. You can.

Here, the drivable area recognition network is an artificial neural network-based network that uses input images as input data and classifies the input images into drivable areas or non-drivable areas. The drivable area recognition network will be described later with reference to FIG. 4.

In one example, the driving policy network outputs the mean and variance of the expected positions of candidate lookahead points.

For example, the driving policy network may include an encoder layer that extracts a feature map of the drivable area and a fully connected layer that outputs the average and variance of the expected position of the lookahead point from the extracted feature map. You can.

At least one instruction stored in the memory 120, when executed by the processor 110, causes the processor 110 to calculate the average of the expected positions of the candidate lookahead points and the average of the reference lookahead points to determine the lookahead point. If the distance between positions is greater than a predetermined threshold distance or the variance of the expected positions of the candidate lookahead points is greater than the predetermined threshold variance, the reference lookahead point is determined as the lookahead point, and the drivable area, reference lookahead point, is determined. and may be configured to store the determined distance as learning data for the drivable area.

At least one instruction stored in the memory 120, when executed by the processor 110, causes the processor 110 to use the determined look-ahead point to determine the look-ahead point for the drivable area in order to train the driving policy network. It may be configured to update the learning data set of the driving policy network based on additional learning data for the driving policy network and adjust the weights of the driving policy network using a predetermined loss function on the updated learning data set.

The learning data of the driving policy network includes a drivable area, a reference lookahead point, and the distance between the reference lookahead point and the candidate lookahead point, and at least one instruction stored in the memory 120 is executed by the processor 110. , the processor 110 may be configured to update the remaining learning data based on the additional learning data, according to the similarity between the drivable area of the additional learning data and the drivable area of the remaining learning data of the learning data set.

At least one instruction stored in the memory 120, when executed by the processor 110, causes the processor 110 to determine a drivable area based on a series of input images, determine a candidate lookahead point, determine a lookahead point, and It can be configured to repeatedly perform driving policy network training.

The processor 110 reads/writes commands/data to the memory 120 and executes an autonomous driving method for an unstructured driving environment according to an embodiment. The operation of the autonomous driving device 100 for an unstructured driving environment according to an embodiment will be described in detail in FIG. 3 and below in the autonomous driving method for an unstructured driving environment according to an embodiment.

3 is a flowchart of an autonomous driving method for an unstructured driving environment according to an embodiment.

An autonomous driving method for an unstructured driving environment according to an embodiment includes determining a drivable area of a vehicle from an input image (S1), and determining a candidate look-ahead point for the drivable area using a driving policy network. Step (S2), determining the look-ahead point corresponding to the target arrival point of the vehicle based on the candidate look-ahead point and the reference look-ahead point for the drivable area (S3), and driving policy network based on the determined look-ahead point. It includes a training step (S4).

In step S1, the processor 110 may determine the drivable area of the vehicle from the input image.

In one example, the input image may be a bird's eye view of an image of the vehicle's driving direction. The input image is a bird's eye view converted from an image captured by a camera mounted on a vehicle. For example, the processor 110 may convert an image captured by a camera into a bird's eye view by geometric transformation (WarpPerspective).

Step S1 may include executing, by the processor 110, a drivable area recognition network based on the input image.

Here, the drivable area recognition network may be composed of an artificial neural network-based network that uses the input image as input data and classifies the input image into a drivable area or a non-drivable area. The drivable area recognition network will be discussed with reference to FIG. 4.

Step S1 may further include converting the input image into an occupancy grid map indicating a drivable area by the processor 110.

An occupancy grid map refers to an image in which the drivable area of the input image is displayed in grid units according to a binary random variable indicating the presence of an obstacle.

In step S2, the processor 110 may determine a candidate lookahead point for the drivable area using the driving policy network.

In one example, the driving policy network includes an artificial neural network configured to output expected location information of a candidate lookahead point using a drivable area as input data.

Here, the expected location information of the candidate lookahead point may include the average and variance of the expected location of the lookahead point. The driving policy network will be described later with reference to FIG. 5.

In step S3, the processor 110 may determine a lookahead point corresponding to the vehicle's target arrival point based on the candidate lookahead point and the reference lookahead point for the drivable area.

The autonomous driving method according to the embodiment determines the driving policy of the autonomous vehicle based on vision. For example, in an embodiment, driving policy decisions may use imitation learning, a method of imitating the behavior of an expert.

To collect training data for imitation learning, experts can specify reference lookahead points on the grid occupancy map. In other words, the reference look-ahead point is the arrival point set by the expert for the drivable area, and becomes the learning data, that is, the target of imitation. The autonomous driving method according to the embodiment ensures that the candidate lookahead point determined for the drivable area in step S2 imitates the reference lookahead point that is learning data as well as possible.

Step S3 is performed by the processor 110 when the distance between the average of the expected positions of the candidate lookahead points and the position of the reference lookahead point is greater than a predetermined threshold distance or the variance of the expected positions of the candidate lookahead points is determined by a predetermined value. If greater than the critical variance, determining the reference lookahead point as the lookahead point, and the average of the expected positions of the drivable area, the reference lookahead point, and the candidate lookahead point determined by the processor 110 in step S1. It may include storing the distance between the location of the reference lookahead point and the reference lookahead point as learning data for the drivable area.

Meanwhile, experts can select the standard lookahead point by referring to the following criteria. i) Exists inside the drivable area, ii) A point where the vehicle can preferentially avoid the obstacle closest to the vehicle (through the trajectory the vehicle will drive, it can be confirmed whether the selected look-ahead point can sufficiently avoid the obstacle), iii) If there are no obstacles around the vehicle, located at the point furthest from the vehicle. A reference lookahead point may be selected at a point that satisfies all three criteria. This ultimately allows the vehicle to drive toward the drivable area while avoiding obstacles as quickly as possible according to the look-ahead point determined by imitation learning.

In step S4, a driving policy network may be trained based on the lookahead point determined by the processor 110.

Step S4 is a step of adding the look-ahead point determined in step S3 by the processor 110 to the learning data set of the driving policy network as additional learning data for determining the look-ahead point for the drivable area and learning data. It may include adjusting the weights of the driving policy network using a predetermined loss function on the set.

Here, the predetermined loss function is a weighted loss function that is weighted based on the distance between the reference lookahead point and the candidate lookahead point.

An autonomous driving method for an unstructured driving environment according to an embodiment is to select the remaining learning data based on the additional learning data according to the similarity between the drivable area of the additional learning data and the drivable area of the remaining learning data in the learning data set. An updating step may be further included.

Here, the learning data of the driving policy network may include a drivable area, the reference lookahead point, and the distance between the reference lookahead point and the candidate lookahead point.

An autonomous driving method for an unstructured driving environment according to an embodiment includes determining a drivable area based on a series of input images (S1), determining a candidate look-ahead point (S2), and determining a look-ahead point. It may further include repeatedly performing the decision step (S3) and the driving policy network training step (S4).

The autonomous driving method for an unstructured driving environment according to an embodiment may further include determining driving control information for controlling the vehicle to reach the look-ahead point determined in step S3.

Here, the driving control information may include steering angle information and speed information of the vehicle.

Figure 4 is an exemplary configuration diagram of a driving environment recognition network according to an embodiment.

The drivable area recognition network is an artificial neural network-based network that uses input images as input data and classifies the input images into drivable areas or non-drivable areas, and uses a deep neural network.

Referring to FIG. 3, in step S1, a drivable area is distinguished from a bird's eye view image obtained by, for example, two cameras. In one example, the driving environment recognition network can be learned using a supervised learning method using the correct answer image showing the drivable area in a bird's eye view obtained with a camera.

The driving environment recognition network can output a grid occupancy map that displays the drivable area in grid units. The grid occupancy map is used as input data for the driving policy network in step S3, referring to Figure 3.

An exemplary driving environment awareness network may include an encoder layer and a decoder layer based on convolution operations. The driving environment recognition network outputs a grid occupancy map showing the drivable area in black and white from the input image (e.g. RGB image).

Figure 5 is an exemplary configuration diagram of a driving policy network according to an embodiment.

Referring to FIG. 3, in step S2, the processor 110 executes a driving policy network and outputs candidate lookahead points.

The driving policy network can output expected location information of the candidate lookahead point. For example, the expected location information of the candidate lookahead point may include the average and variance of the expected location of the candidate lookahead point. Here, the expected position is expressed as coordinates on the drivable area.

The driving policy network will include an encoder layer that extracts a feature map of the drivable area and a fully connected layer that outputs the average and variance of the expected position of the candidate lookahead point from the extracted feature map. You can.

The encoder layer includes a convolutional layer and a max pooling layer, and extracts a feature map of the drivable area expressed as a grid occupancy map.

The output of the encoder layer is flattened and input to the fully connected layer. A fully connected layer may include at least three fully connected layers. To prevent overfitting to specific data, the fully connected layer can perform dropout.

The fully connected layer uses the average coordinates of the lookahead points ( ) and variance ( ) is output.

The driving policy network can be trained through supervised learning based on learning data.

Meanwhile, in step S3 described above with reference to FIG. 3, the look-ahead point for a given drivable area is determined.

The look-ahead point is a point that the vehicle will reach in a given drivable area, and vehicle control information (eg, steering angle and speed) for reaching the look-ahead point can be calculated using a pure pursuit algorithm.

In one example, while a vehicle is controlled by a look-ahead point, an image (for example, an RGB image or a black-and-white image) divided into a drivable area may be provided together as one data. An example of this is shown in Figure 5 (Combined Image).

6 is an example pseudocode of some processes of an autonomous driving method for an unstructured driving environment according to an embodiment.

The data aggregation (DAgger) pseudo code shown in FIG. 6 includes steps S2 to S4 with reference to FIG. 3 .

The driving policy obtained through a single imitation learning process (behavior cloning) is , and the training data set is am. The driving policy of the Data Aggregation algorithm is and here represents the number of imitation learning repetitions.

Training dataset collected by DAgger Is It is initialized as

is a measure that quantitatively evaluates the performance of the driving policy (i.e., driving policy network), represents the performance threshold of the driving policy. is an additional data set, obtained through the Decision Rule function, which will be described later with reference to FIG. 7.

In line 7, the existing training data set is additional training data is combined with, and a driving policy network is trained as in line 8. Here, a weighted loss function is used instead of a general loss function.

For example, a typical loss function is:

[Equation 1]

here is the behavior predicted by the network during the training process, and i represents the index of the combined learning data.

In this case, the weight loss function is:

[Equation 2]

Specifically, DAgger is called in Line 1, and each variable is initialized in Lines 2-4.

At Line 6, a policy decision routine (Decision Rule) that executes steps S2 and S3 is called.

Lines 7-8 correspond to step (S4), and when the reference lookahead point is determined to be the current lookahead point by the policy decision routine in Line 6, additional learning data (D _k ) for this is provided in Line 7 as training data ( Add to D), and train the driving policy network using the loss function according to Equation 2 described above on the learning data (D) at Line 8.

By lines 5 and 9, the ratio of candidate lookahead points determined by the driving policy network as lookahead points in step S3 ( ) is the performance threshold of this driving policy ( ) Steps (S2) to (S4) by Line 6 to Line 8 are repeatedly performed while less than or equal to ).

Line 11 returns the final driving policy, i.e. the learned driving policy network.

7 is an example pseudocode of one step of an autonomous driving method for an unstructured driving environment according to an embodiment.

The pseudo code for the decision rule shown in FIG. 7 includes steps S2 and S3 with reference to FIG. 3.

Line 5-7 is the process of finding the difference between the driving policy, that is, the candidate look-ahead point obtained by the driving policy network, and the reference look-ahead point specified by the expert for the drivable area, and the difference is It is expressed as Driving policy obtained represents the distribution of candidate lookahead points.

As in line 9, the difference between lookahead points is I'm a disperser are the thresholds ( ), it can be said that a problem situation has occurred in which the likelihood of the vehicle colliding with an obstacle increases due to the actions of the driving policy. (See right image in Figure 8). In this case, the vehicle is a professional action. It is controlled by a grid occupancy map ( )and There are additional data sets ( ) is stored in

If a problem situation does not occur, the vehicle will use a candidate lookahead point ( ) is controlled. The rate at which the driving policy was used while the vehicle was under control was It can be expressed as (see line 19). If this value is high, it can be seen that the driving policy closely follows the expert's behavior. In other words, it can be said that imitation learning is being done well.

Looking again, Line 2-3 is the variable initialization process, and Line 5 corresponds to step (S2) ( ) represents the average of the expected positions of candidate lookahead points), Line 6 obtains the reference lookahead point (a _{exp, i} ).

In line 7, the difference (i.e. distance) between the candidate lookahead point and the reference lookahead point ( ) is obtained, and the variance of the expected position of the candidate lookahead point in Line 8 is assigned to the variable.

Lines 9-16 correspond to step (S3).

In the conditional statement in Line 9, the distance between the candidate lookahead point and the reference lookahead point ( ) is greater than or equal to the threshold (τ), or the variance of the x coordinate of the expected position of the candidate lookahead point is greater than or equal to the threshold (χ), or the variance of the y coordinate of the expected position of the candidate lookahead point is greater than or equal to the threshold (χ). , In Line 10, vehicle control information is determined according to the reference look-ahead point, and in Line 11, the corresponding drivable area (s _i ), reference look-ahead point (a _{exp, i} ), and distance ( ) is acquired as additional learning data.

If the conditional statement in Line 9 is not satisfied, vehicle control information is determined based on the candidate lookahead point in Line 14, and the control count (n _net ) by the candidate lookahead point is increased in Line 15.

In Line 15, the total control count (n- _tot ) is increased.

Repeat Lines 5-17 by Line 4 and Line 18.

In Line 19, the ratio of the control number count by the candidate lookahead point to the total control number count is calculated, and in Line 20, the additional learning data (Dk) and ratio ( ) returns.

FIG. 8 is an example diagram illustrating determination of a lock-ahead point by an autonomous driving method for an unstructured driving environment according to an embodiment.

The dotted box on the left and the image below in Figure 8 are examples of control based on the candidate lookahead point determined by the driving policy network. Since the distance between the reference lookahead point and the candidate lookahead point is less than the critical distance indicated by the blue circle, and the variance of the candidate lookahead point indicated by the cross is less than the critical variance indicated by the dotted cross, referring to Figure 7, Lines 14-15 are This is the result of the execution.

The dotted box on the right of FIG. 8 and the two images below it are examples of cases controlled according to the reference lookahead point. Referring to FIG. 7, this is the result of Line 10-11 being performed when the conditions of Line 9 are satisfied.

9 is an example pseudocode of one step of an autonomous driving method for an unstructured driving environment according to an embodiment.

FIG. 9 is a diagram illustrating a step of updating the remaining learning data based on the additional learning data according to the similarity between the drivable area of the additional learning data and the drivable area of the remaining learning data of the learning data set.

In the pseudo code shown in FIG. 9, the additional learning data D _k obtained while performing steps S2 to S4 with reference to the DAgger shown in FIG. 6 Using the existing learning data set (D), Update .

In one example, similarity between occupancy grid maps can be calculated using the Structural Similarity Index (SSIM).

Occupancy grid map of Dk in Line 4 ( ) and occupancy grid maps of D ( ) Similarity between ( ) is the threshold ( ), the update process takes place in Line 5-9. If the distance value of D _k is is the distance value of D If it is greater than go is updated with In the opposite case go is updated with

Figure 10 is an example diagram for explaining learning data update according to an embodiment. According to the pseudocode shown in Figure 9, the distance of the learning data ( ) shows two examples of updating.

According to an embodiment, limitations of autonomous driving in an unstructured environment can be improved by training a deep neural network with driving data acquired in various and complex environments. By using the look-ahead point as an imitation learning action, an imitation learning iterative algorithm can be applied to autonomous driving to accurately imitate the actions of an expert.

In addition, by reflecting the length difference between the look-ahead point obtained by the network and the expert look-ahead point in training, it provides an algorithm that requires fewer repetitions than the existing algorithm that repeats imitation learning. As a result, it is possible to learn about more diverse environments with the same data generation time and effort, thereby improving the safety of autonomous driving.

The method according to an embodiment of the present invention described above can be implemented as computer-readable code on a medium on which a program is recorded. Computer-readable media includes all types of recording devices that store data that can be read by a computer system. Examples of computer-readable media include HDD (Hard Disk Drive), SSD (Solid State Disk), SDD (Silicon Disk Drive), ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc. There is.

The description of the embodiments of the present invention described above is for illustrative purposes, and a person skilled in the art to which the present invention pertains can easily transform it into another specific form without changing the technical idea or essential features of the present invention. You will be able to understand that Therefore, the embodiments described above should be understood in all respects as illustrative and not restrictive. For example, each component described as unitary may be implemented in a distributed manner, and similarly, components described as distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the claims described below rather than the detailed description above, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention.

100: Autonomous driving device for unstructured driving environments
110: processor
120: memory

Claims (20)

In an autonomous driving method for an unstructured driving environment,
Determining a drivable area of a vehicle from an input image;
determining a candidate lookahead point for the drivable area using a driving policy network;
determining a lookahead point corresponding to a target arrival point of the vehicle based on the candidate lookahead point and a reference lookahead point for the drivable area; and
Comprising training the driving policy network based on the determined lookahead point,
The driving policy network includes an artificial neural network configured to output expected location information of the candidate lookahead point using the drivable area as input data,
The expected position information of the candidate lookahead point includes the mean and variance of the expected position of the candidate lookahead point,
The driving policy network is,
an encoder layer that extracts a feature map of the drivable area; and
Comprising a fully connected layer that outputs the average and variance of the expected position of the candidate lookahead point from the feature map,
Autonomous driving method for unstructured driving environments.
According to claim 1,
The input image is a bird's eye view of the driving direction image of the vehicle,
Autonomous driving method for unstructured driving environments.
According to claim 1,
The step of determining the drivable area is,
Executing a drivable area recognition network based on the input image
Including,
The drivable area recognition network is an artificial neural network-based network that uses the input image as input data and classifies the input image into the drivable area or the non-drivable area.
Autonomous driving method for unstructured driving environment.
According to claim 3,
The step of determining the drivable area is,
Converting the input image into an Occupancy Grid Map indicating the drivable area
Containing more,
Autonomous driving method for unstructured driving environment.
delete delete According to claim 1,
The expected location information of the candidate lookahead point includes the average and variance of the expected location of the lookahead point,
The step of determining the look-ahead point is,
If the distance between the average of the expected positions of the candidate lookahead points and the position of the reference lookahead point is greater than a predetermined threshold distance or the variance of the expected positions of the candidate lookahead points is greater than the predetermined threshold variance, the reference determining a lookahead point as the lookahead point; and
Storing the drivable area, the reference look-ahead point, and the distance as learning data for the drivable area.
Including,
Autonomous driving method for unstructured driving environment.
According to claim 1,
The step of training the driving policy network is,
adding the determined look-ahead point to a learning data set of the driving policy network as additional learning data for determining the look-ahead point for the drivable area; and
Adjusting the weights of the driving policy network using a predetermined loss function on the learning data set
Including,
Autonomous driving method for unstructured driving environments.
According to claim 8,
The predetermined loss function is a weighted loss function that is weighted based on the distance between the reference lookahead point and the candidate lookahead point.
Autonomous driving method for unstructured driving environments.
According to claim 8,
The learning data of the driving policy network includes a drivable area, the reference lookahead point, and the distance between the reference lookahead point and the candidate lookahead point,
The autonomous driving method is,
updating the remaining learning data based on the additional learning data, according to the similarity between the drivable area of the additional learning data and the drivable area of the remaining learning data of the learning data set.
Containing more,
Autonomous driving method for unstructured driving environment.
According to claim 1,
repeatedly performing the steps of determining the drivable area based on a series of input images, determining the candidate lookahead point, determining the lookahead point, and training the driving policy network.
Containing more,
Autonomous driving method for unstructured driving environment.
According to claim 1,
determining driving control information for controlling the vehicle to reach the lookahead point.
Containing more,
Autonomous driving method for unstructured driving environment.
In an autonomous driving device for an unstructured driving environment,
processor; and
Memory that stores at least one instruction
Includes, when the at least one instruction is executed by the processor, causes the processor to:
Determine the drivable area of the vehicle from the input image,
Determine a candidate look-ahead point for the drivable area using a driving policy network,
Determine a lookahead point corresponding to the target arrival point of the vehicle based on the candidate lookahead point and a reference lookahead point for the drivable area,
Configured to train the driving policy network based on the determined look-ahead point,
The driving policy network outputs the mean and variance of the expected location of the candidate lookahead point,
The driving policy network is,
an encoder layer that extracts a feature map of the drivable area; and
A fully connected layer that outputs the average and variance of the expected position of the lookahead point from the feature map.
Autonomous driving devices for unstructured driving environments.
According to claim 13,
When the at least one instruction is executed by the processor, it causes the processor to:
In order to determine the drivable area,
Configured to run a drivable area recognition network based on the input image,
The drivable area recognition network is an artificial neural network-based network that uses the input image as input data and classifies the input image into the drivable area or the non-drivable area.
Autonomous driving devices for unstructured driving environments.
delete According to claim 13,
The driving policy network outputs the mean and variance of the expected location of the candidate lookahead point,
When the at least one instruction is executed by the processor, it causes the processor to:
To determine the look-ahead point,
If the distance between the average of the expected positions of the candidate lookahead points and the position of the reference lookahead point is greater than a predetermined threshold distance or the variance of the expected positions of the candidate lookahead points is greater than the predetermined threshold variance, the reference Determine the look-ahead point as the look-ahead point,
Configured to store the drivable area, the reference look-ahead point, and the distance as learning data for the drivable area,
Autonomous driving devices for unstructured driving environments.
According to claim 13,
When the at least one instruction is executed by the processor, it causes the processor to:
To train the driving policy network,
Adding the determined look-ahead point to the learning data set of the driving policy network as additional learning data for determining the look-ahead point for the drivable area,
Configured to adjust the weights of the driving policy network using a predetermined loss function on the learning data set,
Autonomous driving devices for unstructured driving environments.
According to claim 17,
The learning data of the driving policy network includes a drivable area, the reference lookahead point, and the distance between the reference lookahead point and the candidate lookahead point,
When the at least one instruction is executed by the processor, it causes the processor to:
configured to update the remaining learning data based on the additional learning data, according to the similarity between the drivable area of the additional learning data and the drivable area of the remaining learning data of the learning data set,
Autonomous driving devices for unstructured driving environments.
According to claim 13,
When the at least one instruction is executed by the processor, it causes the processor to:
Configured to repeatedly perform the driving area determination, the candidate look-ahead point determination, the look-ahead point determination, and the driving policy network training based on a series of input images,
Autonomous driving devices for unstructured driving environments.
Stored a computer program including at least one instruction configured to execute, by a processor, an autonomous driving method for an unstructured driving environment according to any one of claims 1 to 4 and 7 to 12. A computer-readable, non-transitory recording medium.

Images