Method and device for driver intention prediction

Abstract

A method performed by a driver intention prediction device (10) for predicting a driver's intention, the method comprising: receiving (S600) driving environment information including lane information and driving trajectory information, receiving (S600) driving information, outputting (S610) a lane feature vector by inputting the lane information to a first encoder, outputting (S610) a trajectory feature vector by inputting the driving trajectory information to a second encoder, and predicting (S640) the driver's intention by inputting a plurality of feature vectors to a driver intention prediction model.

Description

Technical area

[0001] The present disclosure and invention relates to a method and apparatus for driver intention prediction. background

[0002] The description in this section merely provides background information relating to embodiments of the present disclosure and does not necessarily represent prior art.

[0003] An autonomous vehicle aims to navigate to the vehicle's destination without driver intervention. Level 4 and lower autonomous vehicles require driver intervention, and in situations where a driver takes over the wheel, driver intent must be predicted for driving safety control purposes.

[0004] Conventional methods have limited ability to detect driver intentions, such as turning, driving straight ahead, changing lanes, and staying in a vehicle's lane (also called "travel lane") based on information such as vehicle speed, steering angle, and yaw rate. They have difficulty detecting complex driver intentions, such as sudden lane changes, lane entry, and one-way driving within a lane.

[0005] Driver intention is causally related to the lane network, such as intersections, traffic lights, and previous driving trajectories. However, no conventional method takes these factors into account. Therefore, to accurately predict driver intention, a method for determining driver intention that considers the lane network, traffic signals, and previous driving trajectories is required.

[0006] Furthermore, conventional methods for detecting driver intention using artificial intelligence have proposed methods for predicting the vehicle's destination coordinates to predict the future route, but the prediction results may be unstable due to inputs not considered during the training of the artificial intelligence. Therefore, there is a need for a method for detecting driver intention using a stable model. Short explanation

[0007] The present disclosure or invention (hereinafter referred to as "disclosure") relates to a method and apparatus for predicting driver intention. Particular embodiments relate to a method and apparatus for predicting driver intention using a grid map.

[0008] Embodiments of the present disclosure provide a reliable driver intention prediction model and training method that can accurately predict complex driver intentions.

[0009] The embodiments of the present disclosure are not limited to those mentioned above, and other embodiments not mentioned herein will be clearly understood by those skilled in the art from the following description.

[0010] At least one embodiment of the present disclosure provides a method performed by a driver intention prediction device for predicting a driver's intention, the method comprising: receiving driving environment information including lane information and driving trajectory information, and receiving driving information, outputting a lane feature vector by inputting the lane information to a first encoder, outputting a trajectory feature vector by inputting the driving trajectory information to a second encoder, and predicting the driver's intention by inputting a plurality of feature vectors to a driver intention prediction model.

[0011] Another embodiment of the present disclosure provides a driver intention prediction device comprising: an input unit configured to receive driving environment information including lane information and driving trajectory information and to receive driving information; a feature extraction unit configured to output a lane feature vector upon receiving the lane information and output a lane feature vector upon receiving the driving trajectory information; and a driver intention prediction unit configured to predict a driver intention by inputting a plurality of feature vectors into a driver intention prediction model.

[0012] Embodiments of the aforementioned method and device may have one or more of the following features.

[0013] Embodiments may include training the driver intent prediction model using the driver intent and the trajectory information.

[0014] Outputting the lane feature vector comprises: normalizing the lane information by removing disturbances (e.g., noise) from the lane information and displaying the de-noised (e.g., denoised) lane information as a graph (e.g., a curve), and inputting the normalized lane information to the first encoder.

[0015] Outputting the trajectory feature vector comprises: normalizing the trajectory information by removing disturbances (e.g., noise) from the trajectory information and displaying the de-noised (e.g., denoised) trajectory information as a graph (e.g., a curve), and inputting the normalized trajectory information to the second encoder.

[0016] The first encoder and the second encoder can each have a graph convolutional network (GCN).

[0017] Embodiments may further include generating a grid map (e.g., a map having a grid division) using the driving information and the lane information, wherein the driver intent is represented as a destination on the grid map.

[0018] According to embodiments of the present disclosure, it may be possible to predict a complex driver intention by predicting the destination location of a vehicle using a grid map.

[0019] According to embodiments of the present disclosure, it may be possible to accurately predict driver intention by using lane information, trajectory information, and traffic signal information.

[0020] According to embodiments of the present disclosure, it may be possible to provide a method for training a reliable driver intention prediction model by predicting driver intention within a limited grid map.

[0021] The effects of the embodiments of the present disclosure are not limited to the above-mentioned effects, and other effects not mentioned will be clearly understood by those skilled in the art from the following description. Short description of the drawings Fig. 1 is a schematic block diagram of a driver intention prediction device according to at least one embodiment of the present disclosure. Fig. 2 is a schematic block diagram of a feature extraction unit according to at least one embodiment of the present disclosure. Fig. 3 is a diagram illustrating lane information and trajectory information according to at least one embodiment of the present disclosure. Fig. 4A and Fig. 4B illustrate exemplary grid maps generated by a grid map generation unit according to at least one embodiment of the present disclosure. Fig. 5 is a diagram illustrating driver intention predicted by a driver intention prediction model according to at least one embodiment of the present disclosure on a grid map. Fig. 6 is a flowchart of a driver intent prediction method according to at least one embodiment of the present disclosure. Fig. 7 is a flowchart of a method for training a driver intent prediction model according to at least one embodiment of the present disclosure. Detailed description of exemplary embodiments

[0022] Some embodiments of the present disclosure will now be described in detail with reference to the accompanying illustrative drawings. In the following description, like reference numerals preferably refer to like elements, even though the elements are shown in different drawings. Furthermore, in the following description of some embodiments, a detailed description of related known components and functions that might obscure the subject matter of the embodiments of the present disclosure is omitted for the sake of clarity and brevity.

[0023] Various ordinal numbers or alpha codes such as first, second, i), ii), a), b), etc. are placed at the front only to distinguish the individual components, but not to imply or suggest the subject matter, order, or sequence of the components. Whenever in this description a part "comprises" or "includes" a component, this means that the part also includes other components and does not exclude them unless expressly stated otherwise. The terms "unit," "module," and the like refer to one or more units for processing at least one function or operation, which may be implemented by hardware, software, or a combination thereof.

[0024] The description of embodiments of the present disclosure presented below in conjunction with the accompanying drawings is intended to describe exemplary embodiments of the present disclosure and is not intended to represent the only embodiments in which the technical idea of ​​the embodiments of the present disclosure can be implemented.

[0025] Fig. 1 is a schematic block diagram of a driver intention prediction device (e.g., a driver intention prediction device) according to at least one embodiment of the present disclosure. Fig. 1 represent functionally distinguishable elements, and at least one of the components may be implemented so that it is integrated with the others in a real physical environment.

[0026] With reference to Fig. 1, a driver intention prediction device 10 includes all or part of an input unit 100, a feature extraction unit 110, a grid map generation unit 120, and a driver intention prediction unit (e.g., a driver intention prediction unit) 130.

[0027] The driver intention prediction device 10 according to at least one embodiment of the present disclosure may use a coordinate system to represent the positions of objects and elements in the environment. For example, the driver intention prediction device 10 may represent the positions of objects and elements in the environment using a coordinate system with the current position of the vehicle as the origin. The driver intention prediction device 10 may represent the positions of objects and elements in the environment using a coordinate system with the current traveling direction of the vehicle as the x-direction and the width direction of the vehicle as the y-direction.

[0028] The input unit 100 receives the driving information of the vehicle. The driving information of the vehicle includes at least one of an acceleration of the vehicle, a yaw rate of the vehicle, a steering angle of the vehicle, a speed of the vehicle, input values ​​of an accelerator pedal and a brake pedal, or any combination thereof. In one example, the input unit 100 may receive driving information from a sensor unit (not shown). In another example, the input unit 100 may obtain the driving information of the vehicle from a storage device.

[0029] The input unit 100 receives the driving environment information of the vehicle. The driving environment information comprises at least one of a high-resolution map, trajectory information, traffic signal information (e.g., traffic light information), the current position of the vehicle, or any combination thereof. A high-resolution map may include lane information, information about intersections, and the like. The lane information may include information about the lane center point, the lane centerline, and the width of a lane (also called a "lane"). In one example, the input unit 100 may receive the driving environment information from a server. In another example, the input unit 100 may receive the driving environment information from the storage device.

[0030] The feature extraction unit 110 performs pre-processing of the driving information and driving environment information received by the input unit 100 and encodes the pre-processed information to extract feature vectors.

[0031] In the following, the feature extraction unit 110 is described with reference to the Fig. 2 and Fig. 3 described.

[0032] Fig. 2 is a schematic block diagram of the feature extraction unit 110 according to at least one embodiment of the present disclosure. Fig. 2 represent functionally distinguishable elements, and at least one of the components may be implemented so that it is integrated with the others in a real physical environment.

[0033] Fig. 3 is a diagram illustrating lane information and trajectory information according to at least one embodiment of the present disclosure.

[0034] With reference to the Fig. 2 and Fig. 3, the feature extraction unit 110 includes all or part of a lane information unit 200, a trajectory information unit 210, and a driving information unit 220.

[0035] The lane information unit 200 receives lane information from the input unit 100. The lane information is information about a lane included in the high-resolution map. For example, the lane information may include information about a lane center point and a lane center line.

[0036] The lane information unit 200 preprocesses the received lane information (e.g., performs preprocessing on the received lane information). The lane information unit 200 can remove noise from the lane information using a moving average filter. The lane information unit 200 can convert the coordinates of each point of the lane information into relative coordinates based on the current position of the vehicle and store them as a graph.

[0037] The lane information unit 200 may convert the coordinates of each point of the denoised (e.g., denoised) lane information into relative coordinates. The relative coordinates may be those in a coordinate system with the current position of the vehicle as the origin. The lane information may be represented by a set of points. For example, if a lane in which the vehicle is currently located is a straight lane, the lane may be represented by a series of coordinates such as (1, 0), (2, 0), and (3, 0). As another example, a left-turn lane may be represented by a set of coordinates such as (1, 0, 2), (2, 1), and (2, 2, 1, 5). These coordinates are for illustrative purposes only and are not intended to have significant meaning or represent limiting factors in the present disclosure.

[0038] The lane information unit 200 stores the denoised lane information in the form of a graph (e.g., a curve). The graph represents a lane. For example, a straight lane may be represented by a graph connecting the coordinates (1, 0), (2, 0), and (3, 0). In another example, a left-turn lane may be represented by a graph connecting the coordinates of (1, 0, 2), (2, 1), and (2, 2, 1, 5).

[0039] The lane information unit 200 inputs the preprocessed lane information to a first encoder. The first encoder may include a pre-trained artificial neural network. The pre-trained artificial neural network may be trained using machine learning or deep learning. The pre-trained artificial neural network may be, for example, a graph convolutional network (GCN). The lane information unit 200 extracts a feature vector from the lane information using the first encoder and outputs the feature vector. For example, the lane information unit 200 may extract a one-dimensional feature vector from the lane information and output it. The feature vector output by the lane information unit 200 is referred to as a lane feature vector.

[0040] The lane information unit 200 may use only lane information within a certain distance from the current position of the vehicle. For example, the lane information unit 200 may use only lane information from three lanes, including the current lane of the vehicle (e.g., the current lane of the vehicle and two other lanes), to output a one-dimensional feature vector.

[0041] The trajectory information unit 210 receives trajectory information from the input unit 100. The trajectory information is, for example, information about the previous driving trajectory of the vehicle.

[0042] The trajectory information unit 210 preprocesses the received trajectory information (e.g., performs preprocessing on the received trajectory information). The trajectory information unit 210 can remove noise (e.g., noise) from the trajectory information using a moving average filter. The trajectory information unit 210 can convert the coordinates of each point of the trajectory information into relative coordinates based on the current position of the vehicle and store them as a graph.

[0043] The trajectory information unit 210 can convert the coordinates of each point of the denoised (e.g., denoised) trajectory information into relative coordinates. The relative coordinates can be those in a coordinate system with the current position of the vehicle as the origin. The trajectory information can be represented by a set of points. For example, if the vehicle has continuously traveled straight ahead, a trajectory can be represented by the set of (-3, 0), (-2, 0), and (-1, 0). If the vehicle has made a right turn to reach the current position, a trajectory can be represented by the set of (-2, -1.5), (-1.5, -0.5), and (-1, -0.2). The aforementioned coordinates are for illustrative purposes only and do not have any significant meaning or limitation of the disclosure.

[0044] The trajectory information unit 210 stores the denoised (e.g., denoised) trajectory information as a graph (e.g., a curve). A graph represents a trajectory. For example, the trajectory that the vehicle has continuously traveled straight ahead may be represented by a graph connecting the coordinates (-3, 0), (-2, 0), and (-1, 0). In another example, when the vehicle has made a right turn to reach its current position, the trajectory may be represented by a graph connecting the coordinates (-2, -1.5), (-1.5, -0.5), and (-1, -0.2). Normally, a vehicle moves along a lane centerline. Therefore, according to at least one embodiment of the present disclosure, a graph may be a graph that coincides with the lane centerline.

[0045] The trajectory information unit 210 inputs the preprocessed trajectory information to a second encoder. The second encoder may include a pre-trained artificial neural network. The pre-trained artificial neural network may be trained by machine learning or deep learning. The pre-trained neural network may be, for example, a graph convolutional neural network (GCN). The trajectory information unit 210 extracts a feature vector from the trajectory information using the second encoder and outputs the feature vector. For example, the trajectory information unit 210 may extract a one-dimensional feature vector from the trajectory information and output it. The feature vector output by the trajectory information unit 210 is referred to as a trajectory feature vector.

[0046] The trajectory information unit 210 may only use trajectory information within a certain distance from the current position of the vehicle. For example, the trajectory information unit 210 may only use one kilometer of trajectory information to output a one-dimensional feature vector.

[0047] The driving information unit 220 receives the vehicle's driving information from the input unit 100. The driving information unit 220 preprocesses the vehicle's driving information (e.g., performs preprocessing on the received driving information). Preprocessing is a process for removing interference (e.g., noise). The driving information unit 220 may preprocess the vehicle's driving information using a low-pass filter (LPF).

[0048] The driving information unit 220 inputs the preprocessed driving information to a third encoder. The third encoder may include a pre-trained artificial neural network. The pre-trained artificial neural network may be trained by machine learning or deep learning. The driving information unit 220 extracts a feature vector from the driving information using the third encoder and outputs the feature vector. For example, the driving information unit 220 may extract a one-dimensional feature vector from the driving information and output it. The feature vector output by the driving information unit 220 is referred to as a driving feature vector (also, e.g., driving information feature vector).

[0049] The grid map generation unit 120 generates a grid map (e.g., a map with a grid or raster subdivision). The grid map generation unit 120 will be described below with reference to Fig. 4A and Fig. 4B.

[0050] Fig. 4A and Fig. 4B illustrate exemplary grid maps generated by a grid map generation unit according to at least one embodiment of the present disclosure.

[0051] The grid map generation unit 120 generates a grid map for an area (e.g., an area, region) in which driving is possible, based on driving information and driving environment information. The grid map generation unit 120 determines an x-direction length of the grid map using the vehicle speed and acceleration. The grid map generation unit 120 determines the y-direction length of the grid map using the width of a lane and one or more adjacent lanes. The grid map generation unit 120 generates the grid map by dividing the area into a plurality of grid elements (e.g., in the form of pixels, boxes, grid squares, and the like). In this case, the grid map generation unit 120 generates a grid map only for the direction of travel of the vehicle. In other words, the grid map is generated only for the front area of ​​the vehicle (e.g.,an area ahead of the vehicle).

[0052] According to at least one embodiment of the present disclosure, the size of the individual grid elements in the grid map may be the same. With reference to Fig. 4A, all grid elements in the grid map can have the same length in the x-direction and the same length in the y-direction.

[0053] According to another embodiment of the present disclosure, the size of the individual grid elements in the grid map may be different. With reference to Fig. 4B, the size of each grid element in the lane in which the vehicle is currently located may be smaller than the size of each grid element in another lane. For example, the size of each grid element in the vehicle's current lane may be 0.5 times the size of the grid element in another lane. In other words, the y-direction length of each grid element in the vehicle's current lane may be 0.5 times the y-direction length of the grid element in the other lane.

[0054] Alternatively, each grid element of the grid map can be designed differently than described above.

[0055] The driver intention prediction unit 130 predicts a driver's intention. The driver intention prediction unit 130 uses a driver intention prediction model to predict the driver's intention. The driver intention prediction model may be pre-trained using machine learning or deep learning. For example, the driver intention prediction model may be implemented based on a deep neural network (DNN).

[0056] The driver intention prediction unit 130 receives the feature vectors from the feature extraction unit 110. The driver intention prediction unit 130 receives traffic signal information from the input unit 100. The feature vectors include at least one of the driving feature vector, the lane feature vector, the trajectory feature vector, or any combination thereof. The driver intention prediction unit 130 may use a plurality of feature vectors and the traffic signal information to predict the driver intention and train the driver intention prediction model.

[0057] The driver intent prediction model receives the feature vectors and the traffic signal information as inputs and predicts the driver intent. According to at least one embodiment of the present disclosure, the driver intent predicted by the driver intent prediction model may be represented as a grid area on the grid map.

[0058] Fig. 5 is a diagram depicting a grid map illustrating driver intentions predicted by the driver intention prediction model according to at least one embodiment of the present disclosure.

[0059] Referring to Fig. 5, when the predicted driver intention is to travel to the left (e.g., a left turn or a left turn), the driver intent prediction model may predict the fourth (4th) grid area of ​​the grid map as the destination, as shown by reference numeral 500. When the predicted driver intention is to make a sudden lane change, the driver intent prediction model may predict the fourteenth (14th) grid area of ​​the grid map as the destination, as shown by reference numeral 510. When the predicted driver intention is to make a slow lane change, the driver intent prediction model may predict the fifteenth (15th) grid area of ​​the grid map as the destination, as shown by reference numeral 520. If the predicted driver intention is one-way lane driving (e.g., driving off-center and more on one side of the lane), the driver intention prediction model can use the twelfth (12th) lane.) grid area of ​​the grid map as the destination, as shown by reference numeral 530. If the predicted driver intent is a U-turn, then the first (1st) grid area of ​​the grid map may be predicted as the destination, as shown by reference numeral 540.

[0060] Although the Fig. 4A and Fig. While the grid maps shown in Figure 4B have twelve or sixteen grid elements, these are for illustrative purposes only and are not intended to limit the present disclosure. The grid map may have more than sixteen grid elements or fewer than twelve grid elements.

[0061] The driver intention prediction unit 130 may use the driver intention predicted by the driver intention prediction model to predict a driving route of the vehicle. For example, the driver intention prediction unit 130 may predict the driving route of the vehicle using the driver intention and a path equation. The driver intention prediction unit 130 may predict the driving route of the vehicle based on a destination on a grid map and a path equation. The driver intention prediction model may be pre-trained using training data generated from driving history data collected from a plurality of vehicles. The driver intention prediction unit 130 may label (e.g., mark, designate, name) the actual driving route of the vehicle and train the driver intention prediction model using the labeled actual driving route as training data.The actual driving route of the vehicle can be obtained from trajectory information. For example, the actual driving route may be the same as the vehicle's trajectory contained in the trajectory information.

[0062] Fig. 6 is a flowchart of a driver intent prediction method according to at least one embodiment of the present disclosure.

[0063] Referring to Fig. 6, the input unit 100 receives driving information and driving environment information (S600). The driving environment information includes at least one of lane information, trajectory information, traffic signal information, or any combination thereof.

[0064] The feature extraction unit 110 preprocesses the lane information, the trajectory information, and the driving information and extracts their corresponding feature vectors (S610). For example, the feature extraction unit 110 may represent the lane information and the trajectory information from which disturbances (e.g., noise) have been removed as graphs (e.g., curves). For example, the feature extraction unit 110 may remove disturbances (e.g., noise) from the driving information. The feature extraction unit 110 may input the preprocessed lane information to the first encoder to extract a lane feature vector. The feature extraction unit 110 may input the trajectory information to the second encoder to extract a trajectory feature vector.The feature extraction unit 110 may input the preprocessed driving information to the third encoder to extract a driving feature vector.

[0065] The grid map generation unit 120 uses the driving information and the driving environment information to generate a grid map (S620). The grid map generation unit 120 may determine an area (e.g., area, region) in which the grid map is to be generated using the lane width, the vehicle speed, and the vehicle acceleration, and may generate the grid map by dividing the area in which the grid map is to be generated.

[0066] The driver intention prediction unit 130 concatenates (e.g., concatenates) the feature vectors. The feature vectors are those extracted from the lane information, the trajectory information, and the driving information. The driver intention prediction unit 130 inputs the concatenated feature vectors into the driver intention prediction model (S630).

[0067] The driver intention prediction model predicts driver intention (S640). The driver intention prediction model can predict driver intention by identifying a specific grid area (e.g., which belongs to a specific grid element) on the grid map as the destination.

[0068] According to at least one embodiment of the present disclosure, the driver intention prediction model may be a model obtained by a learning method described with reference to Fig. 7 described.

[0069] Fig. 7 is a flowchart of a method for training a driver intent prediction model according to at least one embodiment of the present disclosure.

[0070] With reference to Fig. 7, the driver intention prediction model predicts a driver's intention (S700). The driver intention prediction model can predict the driver's intention by identifying a specific grid area on the grid map as a destination.

[0071] The driver intention prediction model may use a path equation to generate a driving route to the specific grid area determined as the destination (S710). The driver intention prediction model may predict a driving route to the destination using, for example, a cubic path equation.

[0072] The driver intention prediction model can receive an actual driving route. The actual driving route can be obtained, for example, from the trajectory information. The driver intention prediction model can mark (e.g., mark, designate, name), also called "label" (S720), the actual driving route or the trajectory information.

[0073] The driver intention prediction model can be trained using the labeled actual driving route or the labeled trajectory information (S730). For example, the driver intention prediction model can be trained by comparing the labeled actual driving route and the predicted driving route, and using the comparison results to modify the weights of the plurality of layers.

[0074] The apparatus or method according to embodiments of the present disclosure may include the respective components implemented as hardware or software, or as a combination of hardware and software. Furthermore, each component may be functionally implemented by software, and a microprocessor may perform the function by software for each component, if implemented.

[0075] Numerous example implementations of the systems and methods described herein may be realized by digital electronic circuits, integrated circuits, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These numerous implementations may also include those realized in one or more computer programs executable on a programmable system. The programmable system includes at least one programmable processor coupled to receive and send data and instructions from and to a memory system, at least one input device, and at least one output device, where the programmable processor may be a special-purpose processor or a general-purpose processor.The computer programs (which are also referred to as programs, software, software applications or code) contain instructions for a programmable processor and are stored on a “computer-readable recording medium”.

[0076] The computer-readable recording medium includes any type of recording device on which data readable by a computer system can be recorded. Examples of computer-readable recording media include non-transitory or non-transitory media, such as a ROM, a CD-ROM, a magnetic tape, a floppy disk, a memory card, a hard disk, an optical / magnetic disk, storage devices, and the like. The computer-readable recording media may also include transient or transitory media, such as a data transmission medium. Furthermore, the computer-readable recording medium may be distributed among computer systems connected via a network, wherein the computer-readable codes may be stored and executed in a distributed manner.

[0077] Although the steps in the respective flowcharts / flowcharts are described as being executed sequentially, they merely illustrate the technical idea of ​​some embodiments of the present disclosure. Therefore, a person skilled in the art can perform the steps by changing the sequences described in the respective flowcharts / flowcharts or by executing two or more of the steps in parallel, and therefore, the steps in the respective flowcharts / flowcharts are not limited to the illustrated chronological sequences.

[0078] Although exemplary embodiments of the present disclosure have been described for illustrative purposes, those skilled in the art will understand that numerous modifications, additions, and substitutions are possible without departing from the spirit and scope of the claimed invention. Therefore, exemplary embodiments of the present disclosure have been described for brevity and clarity. The scope of the technical idea of ​​the embodiments of the present disclosure is not limited by these illustrations. Accordingly, one skilled in the art will understand that the scope of the claimed invention is not intended to be limited by the embodiments explicitly described above, but rather by the claims and their equivalents.

Claims

[1] A method performed by a driver intention prediction device (10) for predicting a driver's intention, the method comprising: Receiving (S600) driving environment information, which includes lane information and driving trajectory information, Receiving (S600) driving information, Outputting (S610) a lane feature vector by inputting the lane information into a first encoder, Outputting (S610) a trajectory feature vector by inputting the travel trajectory information into a second encoder, and Predicting (S640) the driver intention by inputting a plurality of feature vectors into a driver intention prediction model. [2] The method of claim 1, further comprising: training the driver intention prediction model using the driver intention and the trajectory information. [3] The method of claim 2, wherein training the driver intention prediction model comprises: Marking the trajectory information, Generating a predicted route using driver intent, Comparing the marked trajectory information with the predicted route, and Changing a weight of the driver intention prediction model based on a result of the comparison. [4] The method according to any one of claims 1 to 3, further comprising: outputting a driving feature vector by inputting the driving information to a third encoder. [5] The method of claim 4, wherein outputting the driving feature vector comprises: Preprocessing the driving information by removing interference, such as noise, from the driving information, and Input the preprocessed driving information into the third encoder. [6] The method according to any one of claims 1 to 5, further comprising: generating a grid map using the driving information and the lane information, wherein the driver intention is represented as a destination on the grid map. [7] The method of claim 6, wherein generating the grid map comprises: Determining a size of an area for which the grid map is to be generated using the driving information and the lane information, and Creating the grid map by dividing the area into a plurality of grid elements. [8] The method of claim 7, wherein generating the grid map by dividing the area into the plurality of grid elements comprises: dividing the area such that each grid element of the plurality of grid elements in a lane of a vehicle has a different size than other grid elements of the plurality of grid elements on other lanes. [9] The method according to any one of claims 1 to 8, wherein predicting (S640) the driver intention comprises: concatenating and inputting the plurality of feature vectors into the driver intention prediction model. [10] The method according to any one of claims 1 to 9, wherein predicting (S640) the driver's intention comprises: predicting the driver's intention by inputting the plurality of feature vectors and traffic signal information included in the driving environment information into the driver's intention prediction model. [11] A method according to any one of claims 1 to 10, wherein: outputting the lane feature vector: Normalizing the lane information by removing disturbances, e.g. noise, from the lane information and displaying the normalized lane information as a first graph, and Entering the normalized lane information into the first encoder, and outputting the trajectory feature vector: Normalizing the trajectory information by removing disturbances, e.g. noise, from the trajectory information and displaying the normalized trajectory information as a second graph, and Input the normalized trajectory information into the second encoder. [12] Method according to one of claims 1 to 11, wherein the first encoder comprises a graph convolutional network, GCN for short. [13] Driver intention prediction device (10), the device comprising: one or more processors, and a memory device storing a program executable by the one or more processors, the program comprising instructions for: Receiving driving environment information, which includes lane information and trajectory information, Receiving driving information, Outputting a lane feature vector upon receipt of the lane information, Outputting a trajectory feature vector upon receipt of the trajectory information, and Predicting a driver intention by inputting a plurality of feature vectors into a driver intention prediction model. [14] The apparatus (10) of claim 13, wherein the program further comprises instructions for training the driver intent prediction model using the driver intent and the trajectory information. [15] The apparatus (10) of claim 14, wherein the program further comprises instructions for: Generating a predicted route using driver intent, Marking the trajectory information, Comparing the marked trajectory information with the predicted route, and Changing a weight of the driver intention prediction model based on a result of the comparison. [16] The apparatus (10) of any one of claims 13 to 15, wherein the program further comprises instructions for outputting a driving feature vector by inputting the driving information to a third encoder upon receipt of the driving information. [17] Apparatus (10) according to any one of claims 13 to 16, wherein the program further comprises instructions for: Generating a grid map using the driving information and the lane information, Determining a size of an area for which the grid map is to be generated using the driving information and the lane information, Generating the grid map by dividing the area into a plurality of grid elements, and. Representing the driver's intention as a destination on the grid map. [18] Apparatus (10) according to any one of claims 13 to 17, wherein the program further comprises instructions for: Generating normalized lane information and normalized trajectory information by displaying the lane information and the trajectory information as a respective graph, Outputting the lane information from a first encoder upon receipt of the normalized lane information, and Outputting the trajectory information from a second encoder upon receipt of the normalized trajectory information [19] Device (10) according to claim 18, wherein the first encoder and the second encoder each comprise a graph convolutional network, GCN for short. [20] The apparatus (10) according to any one of claims 13 to 19, wherein the program further comprises instructions for predicting the driver's intention by inputting the plurality of feature vectors and traffic signal information included in the driving environment information into the driver's intention prediction model.

Images