Decision planning method, decision planning model training method and related device

Through the decision planning model, real-time input information is encoded and decoded to generate intelligent driving decisions, solving the problem of poor flexibility in the rule matching method in the existing technology, and achieving a safe, comfortable and efficient driving experience.

CN119065361APending Publication Date: 2024-12-03MOMENTA (SUZHOU) TECHNOLOGY CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202310614772.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-05-29
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

In existing intelligent driving technology, decision-making and planning is carried out through rule matching methods, which requires a lot of manpower and material resources to collect rules, which are poor in flexibility, often create a sense of jerk during the vehicle driving, affecting the driving experience and safety.

Method used

Using the decision planning model, the real-time input bicycle, other vehicles, static landmarks, lane-level navigation and traffic light information is encoded and decoded through the encoding module and the decoding module to generate planning trajectories and decisions, and use pre-trained AI models. Automatic decision planning, combined with real-time environment learning, reduce dependence on rules.

Benefits of technology

It realizes safer, more comfortable and efficient driving decisions without relying on a large amount of manpower and material resources to collect rules, avoiding the feeling of jerk caused by mismatch between the rules and the actual environment, and improving driving experience and safety.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119065361A_ABST
    Figure CN119065361A_ABST
Patent Text Reader

Abstract

The invention discloses a decision-making planning method, a training method of a decision-making planning model and a related device, the method is realized through the decision-making planning model, the method comprises the following steps: obtaining first input information, the first input information comprising first own vehicle information, first other vehicle information, first static landmark information, first lane level navigation information and first traffic light information; performing attention coding on the first input information based on a coding module to obtain a first target vector; the first target vector is decoded based on a decoding module and the first query vector, first output information is obtained, and the first output information comprises a first planning track of the first vehicle, a score of the first planning track and a first planning decision corresponding to the first planning track; and / or the first output information comprises a second planning track of the first other vehicle, a score of the second planning track and a second planning decision corresponding to the second planning track. The AI model is used for decision planning, and the safe, comfortable and efficient driving effect can be achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of intelligent driving technology, and more specifically, to a decision planning method, a training method for a decision planning model, and related devices. Background Art

[0002] Perceptual intelligence and cognitive intelligence are the two core technologies in the field of intelligent driving. Cognitive intelligence includes decision-making and planning, which can make decisions and plans for the vehicle's driving behavior in advance during the intelligent driving process, such as lane changes, congestion start-stop, detour avoidance, etc. Related technologies use rule matching to make decisions and plans, but the vehicle driving scenarios are diverse and complex. This decision-making and planning method not only requires a lot of manpower and material resources to collect and set rules, but also has poor flexibility, often causing frustration during vehicle driving, seriously affecting the driving experience and driving safety. Summary of the invention

[0003] The present application provides a decision planning method, a training method for a decision planning model, and related devices, which can solve the technical problems caused by decision planning using rule matching: (1) a large amount of manpower and material resources are required to collect and set rules; (2) poor flexibility often causes frustration during vehicle driving, seriously affecting driving experience and driving safety.

[0004] The specific technical solutions are as follows:

[0005] In a first aspect, an embodiment of the present application provides a decision planning method, the method is implemented by a decision planning model, the decision planning model includes an encoding module and a decoding module, the method includes:

[0006] Acquire first input information, wherein the first input information includes first self-vehicle information, first other-vehicle information, first static landmark information, first lane-level navigation information, and first traffic light information, the first other-vehicle information includes information about a first other-vehicle around the first self-vehicle, the first static landmark information includes static landmark information around the first self-vehicle, the first lane-level navigation information includes lane-level navigation information for navigating the first self-vehicle, and the first traffic light information includes information about a first traffic light in front of the first self-vehicle;

[0007] Performing attention encoding on the first input information based on the encoding module to obtain a first target vector;

[0008] The first target vector is decoded based on the decoding module and the first query vector to obtain first output information, wherein:

[0009] The first output information includes a first planned trajectory of the first self-vehicle, a score of the first planned trajectory, and a first planning decision corresponding to the first planned trajectory; and / or, the first output information includes a second planned trajectory of the first other vehicle, a score of the second planned trajectory, and a second planning decision corresponding to the second planned trajectory;

[0010] The first query vector includes a first target query vector and / or a first object decision query vector, wherein the first target query vector includes a vector of a first target lane that the first self-vehicle is to plan to enter and / or a vector of a second target lane that the first other vehicle is to plan to enter; the first object decision query vector includes a vector of a decision corresponding to when the first self-vehicle enters the first target lane and / or a vector of a decision corresponding to when the first other vehicle enters the second target lane.

[0011] It can be seen from the above scheme that the embodiment of the present application can input the first input information including the first self-vehicle information, the first other-vehicle information, the first static landmark information, the first lane-level navigation information and the first traffic light information into the decision planning model, and the encoding module of the decision planning model first encodes the first input information to obtain the first target vector, and then the decoding module of the decision planning model decodes the first target vector by the first query vector to obtain the first output information including the first planned trajectory of the first self-vehicle, the score of the first planned trajectory, and the first planning decision corresponding to the first planned trajectory, and / or, including the second planned trajectory of the first other vehicle, the score of the second planned trajectory, and the first output information of the second planning decision corresponding to the second planned trajectory, so as to subsequently determine the final first planned trajectory of the first self-vehicle and the corresponding first planning decision according to the first output information. It can be seen from this that the embodiments of the present application can use a pre-trained AI (Artificial Intelligence) model to automatically make decisions and plans, and the first input information is real-time data automatically collected during the driving of the first vehicle, so the AI ​​model will automatically learn in combination with the real-time actual driving environment to make planning decisions with safer, more comfortable and efficient driving effects, without relying on a large amount of manpower and material resources to collect and set rules, and there will be no sense of frustration caused by deviations between the rules and the actual driving environment.

[0012] In a first possible implementation manner of the first aspect, performing attention encoding on the first input information based on the encoding module to obtain a first target vector includes:

[0013] Based on the encoding module, each information in the first input information is respectively subjected to instance-by-instance attention encoding to obtain an instance-by-instance vector;

[0014] Based on the encoding module, inter-instance attention encoding is performed on each of the instance intra-vectors to obtain the first target vector.

[0015] It can be seen from the above scheme that the embodiment of the present application adopts the attention mechanism to perform intra-instance encoding and inter-instance encoding on the first input information, so as to obtain not only the information of a single input information itself, but also the correlation information between multiple input information, which provides a solid foundation for the decoding module to make decision-making plans that are more in line with the driving effect of experts.

[0016] In a second possible implementation manner of the first aspect, a method for acquiring the first static landmark information and the first lane-level navigation information includes:

[0017] Acquire the first static landmark information and the first lane-level navigation information from lane-level navigation map information; or,

[0018] Acquire input data of a landmark perception model, process the input data of the landmark perception model based on the landmark perception model, obtain the first static landmark information contained in the road environment image at a target moment, and determine the first lane-level navigation information according to an extended navigation path, wherein the input data of the landmark perception model includes the road environment image at the target moment and a first local posture at the target moment, the first local posture includes an offset of the global posture of the first ego vehicle when collecting the road environment image relative to the global posture at the target starting point, the target moment is any moment when the first ego vehicle collects the road environment image on the extended navigation path, and the extended navigation path is an extended path of the on-board navigation path.

[0019] It can be seen from the above scheme that when there is lane-level navigation map information, the embodiment of the present application can directly obtain the first static landmark information and the first lane-level navigation information from the lane-level navigation map information. When there is no lane-level navigation map information or it is not dependent on the lane-level navigation map information, the embodiment of the present application can use the landmark perception model to automatically perceive the first static landmark information, and determine the first lane-level navigation information based on the extended path of the vehicle-mounted navigation path, thereby achieving the effect of emphasizing perception and neglecting mapping.

[0020] In a third possible implementation manner of the first aspect, the method for generating the landmark perception model includes:

[0021] Acquire a landmark training sample set, wherein each landmark training sample in the landmark training sample set includes: a road environment sample image group, a second local pose corresponding to each frame of the road environment sample image in the road environment sample image group, and a landmark true value corresponding to each frame of the road environment sample image, and the road environment image group includes multiple consecutive frames of road environment sample images;

[0022] The landmark perception model is obtained by training using the landmark training sample set.

[0023] It can be seen from the above scheme that since the landmark perception model is trained based on multiple road environment sample image groups, during the machine learning process, when perceiving landmark information for each frame of the road environment sample image, the previous and next adjacent frames in the road environment sample image group (which can be one adjacent frame or multiple adjacent frames) will be referenced. Therefore, when performing landmark perception on a single frame of the road environment image based on the landmark perception model, the invisible static landmark information in the road environment image can be perceived, including static landmark information that is not included in the current image or cannot be clearly displayed, such as static landmark information outside the sensor detection range, obscured static landmark information, and static landmark information that cannot be clearly displayed due to poor image quality (night, glare scenes, etc.), thereby achieving an invisible but obtainable perception effect.

[0024] In a fourth possible implementation manner of the first aspect, the training method of the decision planning model includes:

[0025] Acquire a decision planning training sample set, wherein each decision planning training sample in the decision planning training sample set includes second self-vehicle information, second other-vehicle information, second static landmark information, second lane-level navigation information, and second traffic light information, wherein the second other-vehicle information includes information about a second other-vehicle around the second self-vehicle, the second static landmark information includes static landmark information around the second self-vehicle, the second lane-level navigation information includes lane-level navigation information for navigating the second self-vehicle, and the second traffic light information includes information about a second traffic light in front of the second self-vehicle;

[0026] Performing attention encoding on the decision planning training sample set based on the encoding module in the initial decision planning model to obtain a second target vector corresponding to each decision planning training sample;

[0027] For each of the decision planning training samples, the second target vector is decoded based on the decoding module and the second query vector in the initial decision planning model to obtain second output information corresponding to each of the decision planning training samples, wherein:

[0028] The second output information includes a third planned trajectory of the second self-vehicle, a score of the third planned trajectory, and a third planning decision corresponding to the third planned trajectory; and / or, the second output information includes a fourth planned trajectory of the second other vehicle, a score of the fourth planned trajectory, and a fourth planning decision corresponding to the fourth planned trajectory;

[0029] The second query vector includes a second target query vector and / or a second object decision query vector, wherein:

[0030] The second target query vector includes a vector of a third target lane that the second self-vehicle is to be planned to enter and / or a vector of a fourth target lane that the second other vehicle is to be planned to enter; the second object decision query vector includes a vector of a decision corresponding to when the second self-vehicle enters the third target lane and / or a vector of a decision corresponding to when the second other vehicle enters the fourth target lane;

[0031] Calculate the loss value based on the third planned trajectory with the highest score, the trajectory truth value of the second self-vehicle, the fourth planned trajectory with the highest score, and the trajectory truth value of the second other-vehicle;

[0032] When the loss value is greater than a preset loss threshold, the model parameters of the initial decision planning model are adjusted, and the decision planning model with adjusted model parameters is continuously trained based on the decision planning training sample set until the loss value is less than or equal to the preset loss threshold, thereby obtaining the final decision planning model.

[0033] In a fifth possible implementation manner of the first aspect, the target self-vehicle information includes the first self-vehicle information and / or the second self-vehicle information; when the target self-vehicle information is the first self-vehicle information, the target self-vehicle is the first self-vehicle, the target other-vehicle information is the first other-vehicle information, the target other-vehicle is the first other-vehicle, the target traffic light information is the first traffic light information, and the target traffic light is the first traffic light; when the target self-vehicle information is the second self-vehicle information, the target self-vehicle is the second self-vehicle, the target other-vehicle information is the second other-vehicle information, the target other-vehicle is the second other-vehicle, the target traffic light information is the second traffic light information, and the target traffic light is the second traffic light, and the method includes:

[0034] The target vehicle information includes at least one of the position of the target vehicle relative to the target vehicle at the current moment within the preset historical time window, the global heading angle of the target vehicle within the preset historical time window, the actual speed of the target vehicle within the preset historical time window, the body size type of the target vehicle, the following gear position of the target vehicle at the current moment, and the target speed of the target vehicle; and / or,

[0035] The target other vehicle information includes at least one of the position of the target other vehicle relative to the target self-vehicle at the current moment within the preset historical time window, the global heading angle of the target other vehicle within the preset historical time window, the actual speed of the target other vehicle within the preset historical time window, the heading angle of the target other vehicle relative to the target self-vehicle at the current moment within the preset historical time window, and the body size type of the target other vehicle; and / or,

[0036] The target traffic light information includes the type and / or lighting status of the target traffic light within the preset historical time window.

[0037] In a second aspect, an embodiment of the present application provides a method for training a decision planning model, the method comprising:

[0038] Acquire a decision planning training sample set, wherein each decision planning training sample in the decision planning training sample set includes second self-vehicle information, second other-vehicle information, second static landmark information, second lane-level navigation information, and second traffic light information, wherein the second other-vehicle information includes information about a second other-vehicle around the second self-vehicle, the second static landmark information includes static landmark information around the second self-vehicle, the second lane-level navigation information includes lane-level navigation information for navigating the second self-vehicle, and the second traffic light information includes information about a second traffic light in front of the second self-vehicle;

[0039] Performing attention encoding on the decision planning training sample set based on the encoding module in the initial decision planning model to obtain a second target vector corresponding to each decision planning training sample;

[0040] For each of the decision planning training samples, the second target vector is decoded based on the decoding module in the initial decision planning model and the second query vector to obtain second output information corresponding to each of the decision planning training samples, wherein:

[0041] The second output information includes a third planned trajectory of the second self-vehicle, a score of the third planned trajectory, and a third planning decision corresponding to the third planned trajectory; and / or, the second output information includes a fourth planned trajectory of the second other vehicle, a score of the fourth planned trajectory, and a fourth planning decision corresponding to the fourth planned trajectory;

[0042] The second query vector includes a second target query vector and / or a second object decision query vector, wherein:

[0043] The second target query vector includes a vector of a third target lane that the second self-vehicle is to be planned to enter and / or a vector of a fourth target lane that the second other vehicle is to be planned to enter; the second object decision query vector includes a vector of a decision corresponding to when the second self-vehicle enters the third target lane and / or a vector of a decision corresponding to when the second other vehicle enters the fourth target lane;

[0044] Calculate the loss value based on the third planned trajectory with the highest score, the trajectory truth value of the second self-vehicle, the fourth planned trajectory with the highest score, and the trajectory truth value of the second other-vehicle;

[0045] When the loss value is greater than a preset loss threshold, the model parameters of the initial decision planning model are adjusted, and the decision planning model with adjusted model parameters is continuously trained based on the decision planning training sample set until the loss value is less than or equal to the preset loss threshold, thereby obtaining the final decision planning model.

[0046] It can be seen from the above scheme that the embodiment of the present application can train the second self-vehicle information, the second other vehicle information, the second static landmark information, the second lane-level navigation information and the second traffic light information during the expert driving process, and continuously adjust the decision planning model parameters through the expert's actual driving data (i.e., the true value) during the training process to obtain a decision planning model that can automatically make expert-level decision planning. That is, the trained decision planning model can be used to automatically learn in combination with the real-time actual driving environment to make planning decisions with safer, more comfortable and efficient driving effects, without relying on a large amount of manpower and material resources to collect and set rules, and there will be no sense of frustration caused by deviations between the rules and the actual driving environment.

[0047] In a third aspect, an embodiment of the present application provides a decision planning device, the device is implemented by a decision planning model, the decision planning model includes an encoding module and a decoding module, and the device includes:

[0048] an acquisition unit, configured to acquire first input information, wherein the first input information includes first self-vehicle information, first other-vehicle information, first static landmark information, first lane-level navigation information, and first traffic light information, wherein the first other-vehicle information includes information about a first other-vehicle around the first self-vehicle, the first static landmark information includes information about static landmarks around the first self-vehicle, the first lane-level navigation information includes lane-level navigation information for navigating the first self-vehicle, and the first traffic light information includes information about a first traffic light in front of the first self-vehicle;

[0049] an encoding unit, configured to perform attention encoding on the first input information based on the encoding module to obtain a first target vector;

[0050] A decoding unit, configured to decode the first target vector based on the decoding module and the first query vector to obtain first output information, wherein:

[0051] The first output information includes a first planned trajectory of the first self-vehicle, a score of the first planned trajectory, and a first planning decision corresponding to the first planned trajectory; and / or, the first output information includes a second planned trajectory of the first other vehicle, a score of the second planned trajectory, and a second planning decision corresponding to the second planned trajectory;

[0052] The first query vector includes a first target query vector and / or a first object decision query vector, wherein the first target query vector includes a vector of a first target lane that the first self-vehicle is to plan to enter and / or a vector of a second target lane that the first other vehicle is to plan to enter; the first object decision query vector includes a vector of a decision corresponding to when the first self-vehicle enters the first target lane and / or a vector of a decision corresponding to when the first other vehicle enters the second target lane.

[0053] In a first possible implementation manner of the third aspect, the encoding unit includes:

[0054] A first encoding module, configured to perform instance-instance attention encoding on each piece of information in the first input information based on the encoding module to obtain an instance-instance vector;

[0055] The second encoding module is used to perform inter-instance attention encoding on each of the instance intra-vectors based on the encoding module to obtain the first target vector.

[0056] In a second possible implementation manner of the third aspect, the acquisition unit includes: a first acquisition module or a second acquisition module;

[0057] The first acquisition module is used to acquire the first static landmark information and the first lane-level navigation information from the lane-level navigation map information;

[0058] The second acquisition module is used to obtain input data of the landmark perception model, process the input data of the landmark perception model based on the landmark perception model, obtain the first static landmark information contained in the road environment image at the target time, and determine the first lane-level navigation information according to the extended navigation path, wherein the input data of the landmark perception model includes the road environment image at the target time and the first local posture at the target time, the first local posture includes the offset of the global posture of the first ego vehicle when collecting the road environment image relative to the global posture at the target starting point, the target time is any time when the first ego vehicle collects the road environment image on the extended navigation path, and the extended navigation path is an extended path of the vehicle-mounted navigation path.

[0059] In a third possible implementation of the third aspect, the second acquisition module is further used to obtain a landmark training sample set before obtaining the input data of the landmark perception model, wherein each landmark training sample in the landmark training sample set includes: a road environment sample image group, a second local pose corresponding to each frame of the road environment sample image in the road environment sample image group, and a landmark true value corresponding to each frame of the road environment sample image, and the road environment image group includes multiple consecutive frames of road environment sample images; the landmark perception model is obtained by training using the landmark training sample set.

[0060] In a fourth possible implementation manner of the third aspect, the device further includes:

[0061] A training unit, used for training the decision planning model;

[0062] The training unit comprises:

[0063] a third acquisition module, configured to acquire a decision planning training sample set, wherein each decision planning training sample in the decision planning training sample set includes second self-vehicle information, second other-vehicle information, second static landmark information, second lane-level navigation information, and second traffic light information, wherein the second other-vehicle information includes information about a second other-vehicle around the second self-vehicle, the second static landmark information includes static landmark information around the second self-vehicle, the second lane-level navigation information includes lane-level navigation information for navigating the second self-vehicle, and the second traffic light information includes information about a second traffic light in front of the second self-vehicle;

[0064] An encoding module, used for performing attention encoding on the decision planning training sample set based on the encoding module in the initial decision planning model to obtain a second target vector corresponding to each decision planning training sample;

[0065] A decoding module is used to decode the second target vector for each decision planning training sample based on the decoding module and the second query vector in the initial decision planning model to obtain second output information corresponding to each decision planning training sample, wherein:

[0066] The second output information includes a third planned trajectory of the second self-vehicle, a score of the third planned trajectory, and a third planning decision corresponding to the third planned trajectory; and / or, the second output information includes a fourth planned trajectory of the second other vehicle, a score of the fourth planned trajectory, and a fourth planning decision corresponding to the fourth planned trajectory;

[0067] The second query vector includes a second target query vector and / or a second object decision query vector, wherein:

[0068] The second target query vector includes a vector of a third target lane that the second self-vehicle is to be planned to enter and / or a vector of a fourth target lane that the second other vehicle is to be planned to enter; the second object decision query vector includes a vector of a decision corresponding to when the second self-vehicle enters the third target lane and / or a vector of a decision corresponding to when the second other vehicle enters the fourth target lane;

[0069] a calculation module, configured to calculate a loss value based on the third planned trajectory with the highest score, the trajectory truth value of the second self-vehicle, the fourth planned trajectory with the highest score, and the trajectory truth value of the second other-vehicle;

[0070] The adjustment training module is used to adjust the model parameters of the initial decision planning model when the loss value is greater than a preset loss threshold, and continue to train the decision planning model after adjusting the model parameters based on the decision planning training sample set until the loss value is less than or equal to the preset loss threshold, thereby obtaining the final decision planning model.

[0071] In a fifth possible implementation manner of the third aspect, the target self-vehicle information includes the first self-vehicle information and / or the second self-vehicle information; when the target self-vehicle information is the first self-vehicle information, the target self-vehicle is the first self-vehicle, the target other-vehicle information is the first other-vehicle information, the target other-vehicle is the first other-vehicle, the target traffic light information is the first traffic light information, and the target traffic light is the first traffic light; when the target self-vehicle information is the second self-vehicle information, the target self-vehicle is the second self-vehicle, the target other-vehicle information is the second other-vehicle information, the target other-vehicle is the second other-vehicle, the target traffic light information is the second traffic light information, and the target traffic light is the second traffic light, and the method includes:

[0072] The target vehicle information includes at least one of the position of the target vehicle relative to the target vehicle at the current moment within the preset historical time window, the global heading angle of the target vehicle within the preset historical time window, the actual speed of the target vehicle within the preset historical time window, the body size type of the target vehicle, the following gear position of the target vehicle at the current moment, and the target speed of the target vehicle; and / or,

[0073] The target other vehicle information includes at least one of the position of the target other vehicle relative to the target self-vehicle at the current moment within the preset historical time window, the global heading angle of the target other vehicle within the preset historical time window, the actual speed of the target other vehicle within the preset historical time window, the heading angle of the target other vehicle relative to the target self-vehicle at the current moment within the preset historical time window, and the body size type of the target other vehicle; and / or,

[0074] The target traffic light information includes the type and / or lighting status of the target traffic light within the preset historical time window.

[0075] It can be seen from the above scheme that the embodiment of the present application can input the first input information including the first self-vehicle information, the first other-vehicle information, the first static landmark information, the first lane-level navigation information and the first traffic light information into the decision planning model, and the encoding module of the decision planning model first encodes the first input information to obtain the first target vector, and then the decoding module of the decision planning model decodes the first target vector by the first query vector to obtain the first output information including the first planned trajectory of the first self-vehicle, the score of the first planned trajectory, and the first planning decision corresponding to the first planned trajectory, and / or, including the second planned trajectory of the first other vehicle, the score of the second planned trajectory, and the first output information of the second planning decision corresponding to the second planned trajectory, so as to subsequently determine the final first planned trajectory of the first self-vehicle and the corresponding first planning decision according to the first output information. It can be seen from this that the embodiments of the present application can use a pre-trained AI model to automatically make decisions and plans, and the first input information is real-time data automatically collected during the driving of the first vehicle, so the AI ​​model will automatically learn in combination with the real-time actual driving environment to make planning decisions with safer, more comfortable and efficient driving effects, without relying on a large amount of manpower and material resources to collect and set rules, and there will be no sense of frustration caused by deviations between the rules and the actual driving environment.

[0076] In a fourth aspect, an embodiment of the present application provides a training device for a decision planning model, the device comprising:

[0077] an acquisition unit, configured to acquire a decision planning training sample set, wherein each decision planning training sample in the decision planning training sample set includes second self-vehicle information, second other-vehicle information, second static landmark information, second lane-level navigation information, and second traffic light information, wherein the second other-vehicle information includes information about a second other-vehicle around the second self-vehicle, the second static landmark information includes static landmark information around the second self-vehicle, the second lane-level navigation information includes lane-level navigation information for navigating the second self-vehicle, and the second traffic light information includes information about a second traffic light in front of the second self-vehicle;

[0078] An encoding unit, configured to perform attention encoding on the decision planning training sample set based on an encoding module in an initial decision planning model, to obtain a second target vector corresponding to each of the decision planning training samples;

[0079] A decoding unit is used to decode the second target vector for each of the decision planning training samples based on the decoding module in the initial decision planning model and the second query vector to obtain second output information corresponding to each of the decision planning training samples, wherein:

[0080] The second output information includes a third planned trajectory of the second self-vehicle, a score of the third planned trajectory, and a third planning decision corresponding to the third planned trajectory; and / or, the second output information includes a fourth planned trajectory of the second other vehicle, a score of the fourth planned trajectory, and a fourth planning decision corresponding to the fourth planned trajectory;

[0081] The second query vector includes a second target query vector and / or a second object decision query vector, wherein:

[0082] The second target query vector includes a vector of a third target lane that the second self-vehicle is to be planned to enter and / or a vector of a fourth target lane that the second other vehicle is to be planned to enter; the second object decision query vector includes a vector of a decision corresponding to when the second self-vehicle enters the third target lane and / or a vector of a decision corresponding to when the second other vehicle enters the fourth target lane;

[0083] a calculation unit, configured to calculate a loss value based on the third planned trajectory with the highest score, the trajectory truth value of the second self-vehicle, the fourth planned trajectory with the highest score, and the trajectory truth value of the second other-vehicle;

[0084] The adjustment training unit is used to adjust the model parameters of the initial decision planning model when the loss value is greater than a preset loss threshold, and continue to train the decision planning model after adjusting the model parameters based on the decision planning training sample set until the loss value is less than or equal to the preset loss threshold, thereby obtaining the final decision planning model.

[0085] It can be seen from the above scheme that the embodiment of the present application can train the second self-vehicle information, the second other vehicle information, the second static landmark information, the second lane-level navigation information and the second traffic light information during the expert driving process, and continuously adjust the decision planning model parameters through the expert's actual driving data (i.e., the true value) during the training process to obtain a decision planning model that can automatically make expert-level decision planning. That is, the trained decision planning model can be used to automatically learn in combination with the real-time actual driving environment to make planning decisions with safer, more comfortable and efficient driving effects, without relying on a large amount of manpower and material resources to collect and set rules, and there will be no sense of frustration caused by deviations between the rules and the actual driving environment.

[0086] In a fifth aspect, an embodiment of the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any possible implementation of the first aspect or any possible implementation of the second aspect.

[0087] In a sixth aspect, an embodiment of the present application provides an electronic device, the electronic device comprising:

[0088] one or more processors;

[0089] The processor is coupled to a storage device, and the storage device is used to store one or more programs;

[0090] When one or more programs are executed by one or more processors, the electronic device implements the method described in any possible implementation manner of the first aspect or any possible implementation manner of the second aspect.

[0091] In a seventh aspect, an embodiment of the present application provides a vehicle, the vehicle comprising an apparatus as described in any possible implementation of the third aspect or any possible implementation of the fourth aspect, or comprising an electronic device as described in the sixth aspect.

[0092] In an eighth aspect, an embodiment of the present application provides a computer program product, which includes instructions. When the instructions are executed on a computer or a processor, the computer or the processor executes the method described in any possible implementation of the first aspect or any possible implementation of the second aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art description are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative work.

[0094] Figure 1 A flowchart of a decision-making planning method provided in an embodiment of the present application;

[0095] Figure 2 An example diagram of a decision planning framework provided in an embodiment of the present application;

[0096] Figure 3 A flowchart of a training method for a decision planning model provided in an embodiment of the present application;

[0097] Figure 4 A block diagram of a decision-making planning device provided in an embodiment of the present application;

[0098] Figure 5 A block diagram of a training device for a decision planning model provided in an embodiment of the present application;

[0099] Figure 6 A schematic diagram of the structure of an electronic device or computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0100] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0101] It should be noted that, in the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The terms "including" and "having" in the embodiments of the present application and the accompanying drawings and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device comprising a series of steps or units is not limited to the listed steps or units, but optionally also includes steps or units that are not listed, or optionally also includes other steps or units inherent to these processes, methods, products or devices.

[0102] Figure 1 The flowchart of a decision planning method is shown in FIG. 1 . The method can be applied to electronic devices or computer devices, and can be applied to vehicles or servers. Specifically, the method is implemented by a decision planning model, and the decision planning model includes an encoding module and a decoding module. The method may include the following steps:

[0103] S110: Obtain first input information.

[0104] The first input information includes first vehicle information, first other vehicle information, first static landmark information, first lane-level navigation information, and first traffic light information. The first other vehicle information includes information about the first other vehicle around the first vehicle, the first static landmark information includes static landmark information around the first vehicle, the first lane-level navigation information includes lane-level navigation information for navigating the first vehicle, and the first traffic light information includes information about the first traffic light in front of the first vehicle.

[0105] The first vehicle information includes at least one of the following: the position of the first vehicle relative to the first vehicle at the current moment within the preset historical time window, the global heading angle of the first vehicle within the preset historical time window, the actual speed of the first vehicle within the preset historical time window, the body size type of the first vehicle, the following gear position of the first vehicle at the current moment, and the target speed of the first vehicle; and / or,

[0106] The first other vehicle information includes at least one of the following: the position of the first other vehicle relative to the first vehicle at the current moment within the preset historical time window, the global heading angle of the first other vehicle within the preset historical time window, the actual speed of the first other vehicle within the preset historical time window, the heading angle of the first other vehicle relative to the first vehicle at the current moment within the preset historical time window, and the body size type of the first other vehicle; and / or,

[0107] The first traffic light information includes the type and / or lighting status of the first traffic light within a preset historical time window.

[0108] The first self-vehicle information may also include the current position of the first self-vehicle, the global orientation angle of the first self-vehicle, and the actual speed of the first self-vehicle at the current moment; the first other-vehicle information may also include the current position of the first other-vehicle relative to the first self-vehicle, the global orientation angle of the first other-vehicle at the current moment, the actual speed of the first other-vehicle at the current moment, and the orientation angle of the first other-vehicle relative to the first self-vehicle at the current moment. The first traffic light information may also include the traffic light status of the first traffic light at the current moment. The types of the first traffic lights include motor vehicle traffic lights, non-motor vehicle traffic lights, lane lights, flashing warning lights, pedestrian crossing lights, direction indicator lights, road and railway level crossing lights, etc. The lighting status of the first traffic light includes which color light is in the lighting state, if there is a digital light countdown, how long is left, etc. The specific value of the preset historical time window can be determined according to actual experience, for example, it can be 20s.

[0109] The position of the first self-vehicle at each time contained in the first self-vehicle information can be obtained by positioning the first self-vehicle's positioning system, the actual speed and global heading angle of the first self-vehicle at each time contained in the first self-vehicle information can be measured by the inertial navigation system of the first self-vehicle, and the target speed of the first self-vehicle is the expected speed set by the user. The body size type of the first self-vehicle can be directly read from the system of the first self-vehicle. Similarly, the position, global heading angle, and actual speed of the first other vehicle can be measured by the relevant measurement device of the first other vehicle.

[0110] When the embodiment of the present application is applied to the first self-vehicle, the position and orientation angle of the first other vehicle relative to the first self-vehicle can be obtained by comprehensive calculation through the image collector, laser radar and other related devices of the first self-vehicle, and the global orientation angle of the first other vehicle can be transmitted from the first other vehicle to the first self-vehicle through V2X (Vehicle-to-Everything). The body size type of the first other vehicle can be obtained by comprehensive calculation through the image collector, laser radar and other related devices of the first self-vehicle, and can also be transmitted from the first other vehicle to the first self-vehicle through V2V. When the embodiment of the present application is applied to the server, the first self-vehicle information, the first other vehicle information and the first traffic light information can all be obtained by the vehicle measurement and uploaded to the server.

[0111] The first static landmark information includes static landmark information on the road surface within a preset distance range around the first self-vehicle, such as a lane centerline, a road centerline, a road boundary line, a static obstacle, etc.

[0112] The first lane-level navigation information includes a lane-level navigation path, a real-time position of the first ego vehicle on the lane-level navigation path, and the like.

[0113] Methods for obtaining the first static landmark information and the first lane-level navigation information include but are not limited to the following two:

[0114] The first method is to obtain the first static landmark information and the first lane-level navigation information from the lane-level navigation map information when the lane-level navigation map information is available.

[0115] When the lane-level navigation map information can be obtained, the first static landmark information and the first lane-level navigation information can be directly obtained from the lane-level navigation map information. The lane-level navigation map information includes a high-precision map.

[0116] The second method: in the absence of lane-level navigation map information or without relying on lane-level navigation map information, obtaining input data of the landmark perception model, processing the input data of the landmark perception model based on the landmark perception model, obtaining the first static landmark information contained in the road environment image at the target moment, and determining the first lane-level navigation information according to the extended navigation path.

[0117] The input data of the landmark perception model include the road environment image at the target time and the first local pose at the target time. The first local pose includes the offset of the global pose of the first ego vehicle when collecting the road environment image relative to the global pose at the target starting point. The target time is any time when the first ego vehicle collects the road environment image on the extended navigation path. The extended navigation path is the extended path of the vehicle navigation path. The vehicle navigation path includes basic information such as road geometry, road type, and road name.

[0118] The coordinate system where the local posture is located can be a boot coordinate system. In the boot coordinate system, the position point where the vehicle is powered on is the origin of the coordinate system, so the coordinate system is also called the boot coordinate system. In addition, the local posture can also be located based on the navigation and positioning system without the need for high-precision positioning. The target starting point can be any specified position, for example, it can be the position when the first ego vehicle is powered on, or it can be any specified position during the driving process of the first ego vehicle.

[0119] When there is a difference between the acquisition cycle of the image acquisition device for acquiring the road environment image and the positioning cycle of the navigation and positioning system, for example, if the road environment image is acquired at the target time but positioning is not performed at the target time, then the local pose at the most recent positioning time can be taken as the local pose at the target time. However, since the orders of magnitude of the acquisition cycle and the positioning cycle are relatively small, when there is a time difference, the error generated when taking the road environment image and the local pose for the same target time is negligible.

[0120] The pose includes the vehicle position and attitude. The global pose includes the vehicle position and attitude based on global positioning. When obtaining the global pose, high-precision positioning such as RTK (Real-time kinematic, carrier phase differential technology) is not required. Only road-level positioning is required, such as navigation positioning. Global positioning can use the earth as a reference system for positioning, that is, the coordinate system of global positioning can be the geodetic coordinate system.

[0121] When the user enters the starting point (or the default current location) and the end point in the vehicle navigation software, the vehicle navigation software will generate a vehicle navigation path based on the starting point and the end point, and the navigation positioning system will locate the global position of the first vehicle in real time. The vehicle navigation software can be software on the vehicle computer, or software on a mobile terminal that communicates with the vehicle computer. Therefore, the vehicle navigation path can be a path generated by the vehicle navigation software on the vehicle computer, or a path generated by the vehicle navigation software on the mobile terminal that communicates with the vehicle computer. The embodiment of the present application does not limit the source of the vehicle navigation path.

[0122] The method for obtaining an extended navigation path includes: first obtaining a vehicle navigation path and a global posture of a first vehicle when traveling on the vehicle navigation path; then extracting POI (Point of Interest) information corresponding to the global posture from target data; and finally adding the POI information to the vehicle navigation path to obtain the extended navigation path.

[0123] Among them, the target data includes navigation events and / or vehicle navigation maps. Navigation events refer to events broadcast in the vehicle navigation system, for example, a ramp is about to be entered 1 km ahead, there is a merging point 1 km ahead, there is a fork in the road ahead, please take the left fork, and a tunnel is about to be entered. POI information includes information such as the number of lanes, merging points, speed limit information, long solid lines, etc. The POI information corresponding to the global posture is the POI information within a preset distance range ahead of the global posture on the vehicle navigation path. The preset distance range is greater than or equal to the path length contained in the road environment image collected by the first self-vehicle at the global posture.

[0124] After obtaining the extended navigation path, first lane-level navigation information can be determined based on the extended navigation path and the real-time position of the first ego vehicle in the extended navigation path. For example, navigation information including the extended navigation path and the real-time position of the first ego vehicle in the extended navigation path can be determined as the first lane-level navigation information. Information obtained by integrating the first static landmark information, the extended navigation path and the real-time position of the first ego vehicle in the extended navigation path can also be used as the first lane-level navigation information.

[0125] In one embodiment, a method for generating a landmark perception model includes: obtaining a landmark training sample set, wherein each landmark training sample in the landmark training sample set includes: a road environment sample image group, a second local pose corresponding to each frame of the road environment sample image in the road environment sample image group, and a landmark true value corresponding to each frame of the road environment sample image, and the road environment image group includes multiple consecutive frames of road environment sample images; and obtaining a landmark perception model by training using the landmark training sample set.

[0126] The landmark perception model can be a generative model or other neural network model. The embodiment of the present application does not limit the model algorithm used by the landmark perception model. The road environment image group includes a continuous multi-frame road environment sample image. The "continuous" here includes not only the literal continuous meaning, but also the total number of road environment sample image frames collected within a predetermined time period exceeds a preset value.

[0127] It can be seen from the above scheme that since the landmark perception model is trained based on multiple road environment sample image groups, during the machine learning process, when perceiving landmark information for each frame of the road environment sample image, the previous and next adjacent frames in the road environment sample image group (which can be one adjacent frame or multiple adjacent frames) will be referenced. Therefore, when performing landmark perception on a single frame of the road environment image based on the landmark perception model, the invisible static landmark information in the road environment image can be perceived, including static landmark information that is not included in the current image or cannot be clearly displayed, such as static landmark information outside the sensor detection range, obscured static landmark information, and static landmark information that cannot be clearly displayed due to poor image quality (night, glare scenes, etc.), thereby achieving an invisible but obtainable perception effect.

[0128] In one embodiment, the landmark true value can be manually annotated or automatically annotated. The automatic annotation method includes: for each road environment sample image group to be processed, according to the third local pose of each frame of the road environment sample image in the road environment sample image group, obtaining the road test time series data corresponding to the road environment sample image group; constructing the target local map based on the road test time series data; and obtaining the landmark true value of each frame of the road environment sample image by projecting the target local map onto each frame of the road environment sample image in the road environment sample image group.

[0129] The road test time series data includes the vehicle driving trajectory and the road environment images collected during the vehicle driving. The road test time series data can be data recorded by a driving recorder, data collected by vehicle-mounted sensors, data recorded when the server interacts with the vehicle, or data recorded by other devices, as long as it is real time series data.

[0130] When obtaining the road test timing data corresponding to the road environment sample image group based on the third local posture of each frame of the road environment sample image in the road environment sample image group, the third local posture of each frame of the road environment sample image in the road environment sample image group can be obtained first, and then the road test timing data containing these third local postures can be searched, and the road test timing data containing these third local postures can be used as the road test timing data corresponding to the road environment sample image group.

[0131] A method for projecting a target local map onto a road environment sample image to obtain a landmark truth value of the road environment sample image includes: converting the target local map from a map coordinate system to an image coordinate system, and using static landmark information of the target local map converted to the image coordinate system in the road environment sample image as the landmark truth value of the road environment sample image.

[0132] S120: Perform attention encoding on the first input information based on the encoding module to obtain a first target vector.

[0133] Among them, attention encoding refers to encoding using the attention mechanism. The core goal of the attention mechanism is to select the more critical information for the current task goal from a large amount of information and focus on it.

[0134] Specifically, each information in the first input information can be firstly subjected to instance-instance attention encoding based on the encoding module to obtain each instance-instance vector; and each instance-instance vector can be subjected to inter-instance attention encoding based on the encoding module to obtain the first target vector. In order to distinguish the method for obtaining the second target vector, the instance-instance vector in this step can be referred to as the first instance-instance vector. Both the first target vector and the second target vector include at least one multidimensional vector.

[0135] Intra-instance attention encoding refers to encoding based on intra-instance attention, and inter-instance attention encoding refers to encoding based on inter-instance attention. Intra-instance attention refers to focusing attention within an instance, rather than between different instances, while inter-instance attention refers to focusing attention between different instances, rather than focusing on a single instance.

[0136] In the embodiment of the present application, each of the first vehicle information, the first other vehicle information, the first static landmark information, the first lane-level navigation information and the first traffic light information can be represented as a whole by an instance, or each specific information contained in each information can be represented by an instance. For example, the first static landmark information can be represented as a whole as an instance, or each lane centerline, road centerline and other information in the first static landmark information can be represented by an instance.

[0137] like Figure 2 As shown in the figure, each row of boxes in the encoding module can abstractly represent an instance. For example, the first row represents the center line of a lane, so each box in this row can represent each point on the center line of the lane. For another example, the second row represents the information of the first self-vehicle, so each box in this row can represent the position and actual speed of the first self-vehicle at each moment. After the input information is input into the encoding module, each type of input information can be Figure 2 As shown, each input information is converted into rows of vectors (each box represents a vector), and then horizontal and vertical information fusion is performed, where horizontal fusion is equivalent to intra-instance attention encoding, and vertical fusion is equivalent to inter-instance attention encoding.

[0138] The embodiment of the present application adopts the attention mechanism to perform intra-instance encoding and inter-instance encoding on the first input information, so as to obtain not only the information of a single input information itself, but also the correlation information between multiple input information, thereby providing a solid foundation for the decoding module to make decision-making plans that are more in line with the driving effect of experts.

[0139] S130: Decode the first target vector based on the decoding module and the first query vector to obtain first output information.

[0140] like Figure 2 As shown, decoding the first target vector includes decoding the first target vector according to time attention and agent attention, thereby obtaining the position information changes of different agents (including the self-car and other cars) in time (i.e., planning trajectories). Time attention refers to focusing on the time dimension, i.e., the changes of different agents in the time dimension, and agent attention refers to focusing on different agents respectively.

[0141] It should be noted that Figure 2 The input information in can be the first input information or the decision planning training sample set. Similarly, the output information can be the first output information or the second output information.

[0142] The first output information includes a first planned trajectory of the first self-vehicle, a score of the first planned trajectory, and a first planning decision corresponding to the first planned trajectory; and / or, the first output information includes a second planned trajectory of the first other vehicle, a score of the second planned trajectory, and a second planning decision corresponding to the second planned trajectory.

[0143] The first query vector includes a first target query vector and / or a first object decision query vector, wherein the first target query vector includes a vector of a first target lane that the first self-vehicle is to plan to enter and / or a vector of a second target lane that the first other vehicle is to plan to enter; the first object decision query vector includes a vector of a decision corresponding to when the first self-vehicle enters the first target lane and / or a vector of a decision corresponding to when the first other vehicle enters the second target lane.

[0144] The vector of the decision corresponding to when the first ego vehicle enters the first target lane can be: a vector generated according to the decision automatically calculated by the upstream intelligent driving system, or: a vector corresponding to all possible decisions, such as changing lanes to the left, changing lanes to the right, overtaking from the left, overtaking from the right, following the vehicle start / stop, etc. Correspondingly, the vector of the first target lane to be planned to be entered by the first ego vehicle includes the vector corresponding to the left lane adjacent to the current lane, the vector corresponding to the right lane adjacent to the current lane, and the vector corresponding to the current lane.

[0145] The vector of the decision corresponding to when the first other vehicle enters the second target lane may include: traversing vectors corresponding to all possible decisions, such as changing lanes to the left, changing lanes to the right, overtaking from the left, overtaking from the right, following the vehicle start / stop, etc. Correspondingly, the vector of the second target lane to be planned to be entered by the first other vehicle includes a vector corresponding to the left lane adjacent to the current lane, a vector corresponding to the right lane adjacent to the current lane, and a vector corresponding to the current lane.

[0146] In one implementation, target query and object decision query related information may be pre-stored, and then a multimodal query method may be used to query the targets and object decisions of the own vehicle and other vehicles, and the queried information may be converted into a vector to participate in the decoding calculation.

[0147] In this step, after obtaining the first query vector and the first target vector, the first target vector can be decoded based on the decoding module and the first query vector to obtain the first planning trajectory and score of each target and decision of the first self-vehicle represented by the first query vector, and the second planning trajectory and score of each target and decision of the first other vehicle represented by the first query vector, so as to determine the final first planning trajectory of the first self-vehicle according to the score, or according to the score and other auxiliary information (such as traffic rules). For example, the first planning trajectory with the highest score is selected as the final planning trajectory of the first self-vehicle.

[0148] It should be added that the first planning decision corresponds to: the planning decision represented by the vector of the decision corresponding to the first self-vehicle entering the first target lane included in the first object decision query vector, and the second planning decision corresponds to: the planning decision represented by the vector of the decision corresponding to the first other vehicle entering the second target lane included in the second object decision query vector.

[0149] The decision planning method provided in the embodiment of the present application can input first input information including first self-vehicle information, first other-vehicle information, first static landmark information, first lane-level navigation information and first traffic light information into a decision planning model, and firstly encode the first input information by an encoding module of the decision planning model to obtain a first target vector, and then decode the first target vector by a first query vector of a decoding module of the decision planning model to obtain first output information including a first planned trajectory of the first self-vehicle, a score of the first planned trajectory, and a first planning decision corresponding to the first planned trajectory, and / or first output information including a second planned trajectory of the first other-vehicle, a score of the second planned trajectory, and a second planning decision corresponding to the second planned trajectory, so as to subsequently determine the final first planned trajectory of the first self-vehicle and the corresponding first planning decision according to the first output information. It can be seen from this that the embodiments of the present application can use a pre-trained AI model to automatically make decisions and plans, and the first input information is real-time data automatically collected during the driving of the first vehicle, so the AI ​​model will automatically learn in combination with the real-time actual driving environment to make planning decisions with safer, more comfortable and efficient driving effects, without relying on a large amount of manpower and material resources to collect and set rules, and there will be no sense of frustration caused by deviations between the rules and the actual driving environment.

[0150] Corresponding to the above method embodiment, another embodiment of the present application provides a training method for a decision planning model, such as Figure 3 As shown, the method includes:

[0151] S210: Obtain a decision planning training sample set.

[0152] Among them, each decision planning training sample in the decision planning training sample set includes second self-vehicle information, second other-vehicle information, second static landmark information, second lane-level navigation information, and second traffic light information. The second other-vehicle information includes information about the second other-vehicle around the second self-vehicle, the second static landmark information includes static landmark information around the second self-vehicle, the second lane-level navigation information includes lane-level navigation information for navigating the second self-vehicle, and the second traffic light information includes information about the second traffic light in front of the second self-vehicle. The method for obtaining each information in the decision planning training sample set can refer to the method for obtaining the corresponding input information in S110, which will not be repeated here.

[0153] The second self-vehicle information includes at least one of the following: the position of the second self-vehicle relative to the second self-vehicle at the current moment within the preset historical time window, the global heading angle of the second self-vehicle within the preset historical time window, the actual speed of the second self-vehicle within the preset historical time window, the body size type of the second self-vehicle, the following gear position of the second self-vehicle at the current moment, and the target speed of the second self-vehicle; and / or,

[0154] The second other vehicle information includes at least one of the position of the second other vehicle relative to the second own vehicle at the current moment within the preset historical time window, the global heading angle of the second other vehicle within the preset historical time window, the actual speed of the second other vehicle within the preset historical time window, the heading angle of the second other vehicle relative to the second own vehicle at the current moment within the preset historical time window, and the body size type of the second other vehicle; and / or,

[0155] The second traffic light information includes the type and / or lighting status of the second traffic light within a preset historical time window.

[0156] The second self-vehicle information may also include the current position of the second self-vehicle, the global heading angle of the second self-vehicle, and the actual speed of the second self-vehicle at the current moment; the second other-vehicle information may also include the current position of the second other-vehicle relative to the second self-vehicle, the global heading angle of the second other-vehicle at the current moment, the actual speed of the second other-vehicle at the current moment, and the heading angle of the second other-vehicle relative to the second self-vehicle at the current moment. The second traffic light information may also include the traffic light status of the second traffic light at the current moment. The types of second traffic lights include motor vehicle traffic lights, non-motor vehicle traffic lights, lane lights, flashing warning lights, pedestrian crossing lights, direction indicator lights, road and railway level crossing lights, etc. The lighting status of the second traffic light includes which color light is in the lighting state, if there is a digital light countdown, how much time is left, etc.

[0157] The position and global heading angle of the second self-vehicle at each moment contained in the second self-vehicle information can be located by the positioning system of the second self-vehicle, and the actual speed of the second self-vehicle at each moment contained in the second self-vehicle information can be measured by the speed sensor of the second self-vehicle, and the target speed of the second self-vehicle is the expected speed set by the user. The body size type of the second self-vehicle can be directly read from the system of the second self-vehicle. Similarly, the position, global heading angle, actual speed, and body size type of the second other vehicle can be measured by the relevant measuring device of the second other vehicle. The position and heading angle of the second other vehicle relative to the second self-vehicle can be comprehensively calculated by the image collector, laser radar and other relevant devices of the second self-vehicle.

[0158] The second static landmark information includes static landmark information on the road surface within a preset distance range around the second ego vehicle, such as lane centerline, road centerline, road boundary line, static obstacles, etc. The second lane-level navigation information includes a lane-level navigation path, a real-time position of the second ego vehicle on the lane-level navigation path, etc.

[0159] The method for obtaining the second static landmark information and the second lane-level navigation information is the same as the method for obtaining the first static landmark information and the first lane-level navigation information, and will not be repeated here.

[0160] The second self-car is the self-car corresponding to the sample data collected when the decision planning model is trained. The first self-car is the self-car corresponding to the decision planning model applied after the decision planning model is trained. In theory, the first self-car may also participate in the decision planning model, that is, the second self-car includes the first self-car.

[0161] If the first ego vehicle information and the second ego vehicle information are collectively referred to as target ego vehicle information, the related information can be described as follows:

[0162] The target self-vehicle information includes first self-vehicle information and / or second self-vehicle information; when the target self-vehicle information is the first self-vehicle information, the target self-vehicle is the first self-vehicle, the target other-vehicle information is the first other-vehicle information, the target other-vehicle is the first other-vehicle, the target traffic light information is the first traffic light information, and the target traffic light is the first traffic light; when the target self-vehicle information is the second self-vehicle information, the target self-vehicle is the second self-vehicle, the target other-vehicle information is the second other-vehicle information, the target other-vehicle is the second other-vehicle, the target traffic light information is the second traffic light information, and the target traffic light is the second traffic light, the method includes:

[0163] The target vehicle information includes at least one of the position of the target vehicle relative to the target vehicle at the current moment within the preset historical time window, the global heading angle of the target vehicle within the preset historical time window, the actual speed of the target vehicle within the preset historical time window, the body size type of the target vehicle, the following gear position of the target vehicle at the current moment, and the target speed of the target vehicle; and / or,

[0164] The target other vehicle information includes at least one of the position of the target other vehicle relative to the target own vehicle at the current moment within the preset historical time window, the global heading angle of the target other vehicle within the preset historical time window, the actual speed of the target other vehicle within the preset historical time window, the heading angle of the target other vehicle relative to the target own vehicle at the current moment within the preset historical time window, and the body size type of the target other vehicle; and / or,

[0165] The target traffic light information includes the type and / or lighting status of the target traffic light within a preset historical time window.

[0166] S220: Performing attention encoding on the decision planning training sample set based on the encoding module in the initial decision planning model to obtain a second target vector corresponding to each decision planning training sample.

[0167] Based on the encoding module, each information in the decision planning training sample set is respectively encoded with intra-instance attention to obtain each second instance internal vector; based on the encoding module, each second instance internal vector is encoded with inter-instance attention to obtain a second target vector.

[0168] S230: For each decision planning training sample, decode the second target vector based on the decoding module and the second query vector in the initial decision planning model to obtain second output information corresponding to each decision planning training sample.

[0169] like Figure 2 As shown, decoding the second target vector includes decoding the second target vector according to time attention and agent attention, thereby obtaining the temporal position information changes (i.e., planning trajectories) of different agents (including the self-car and other cars).

[0170] Among them, the second output information includes the third planned trajectory of the second self-vehicle, the score of the third planned trajectory, and the third planning decision corresponding to the third planned trajectory; and / or, the second output information includes the fourth planned trajectory of the second other vehicle, the score of the fourth planned trajectory, and the fourth planning decision corresponding to the fourth planned trajectory.

[0171] The second query vector includes a second target query vector and / or a second object decision query vector, wherein the second target query vector includes a vector of a third target lane that the second vehicle is to be planned to enter and / or a vector of a fourth target lane that the second other vehicle is to be planned to enter; the second object decision query vector includes a vector of a decision corresponding to when the second vehicle enters the third target lane and / or a vector of a decision corresponding to when the second other vehicle enters the fourth target lane.

[0172] S240: Calculate the loss value based on the third planned trajectory with the highest score, the true value of the trajectory of the second self-vehicle, the fourth planned trajectory with the highest score, and the true value of the trajectory of the second other vehicle.

[0173] The true value of the trajectory of the second self-car includes the actual trajectory of the expert when driving the second self-car, and the true value of the trajectory of the second other car includes the actual trajectory of the expert when driving the second other car. The expert here may refer to a driver with mature driving skills. The information included in the decision planning training sample set corresponds to the true value of the trajectory of the second self-car and the true value of the trajectory of the second other car, and is the information collected by the expert when driving the second self-car or the second other car.

[0174] In the embodiment of the present application, the first loss value can be calculated based on the third planned trajectory with the highest score and the true value of the second self-vehicle trajectory, and the second loss value can be calculated based on the fourth planned trajectory with the highest score and the true value of the second other-vehicle trajectory, and then the average of the first loss value and the second loss value can be used as the final loss value; after obtaining the first loss value and the second loss value, corresponding preset loss thresholds can be set for the first loss value and the second loss value respectively, and the corresponding preset loss thresholds may be different. When the first loss value and the second loss value are respectively less than or equal to the corresponding preset loss thresholds, the final decision planning model is obtained. As long as one loss value is greater than the corresponding preset loss threshold, it is necessary to adjust the model parameters of the decision planning model, and continue to train the decision planning model after adjusting the model parameters based on the decision planning training sample set.

[0175] S250: When the loss value is greater than the preset loss threshold, adjust the model parameters of the initial decision planning model, and continue to train the decision planning model with adjusted model parameters based on the decision planning training sample set until the loss value is less than or equal to the preset loss threshold, and obtain the final decision planning model.

[0176] When the loss value is greater than the preset loss threshold, the model parameters of the initial decision planning model are adjusted, and the decision planning model after the model parameters are adjusted is continuously trained based on the decision planning training sample set, that is, returning to step S210, and continuing to execute steps S210-S240. When the loss value is less than or equal to the preset loss threshold, the training can be stopped, and the decision planning model trained at this time is used as the final decision planning model.

[0177] In supervised training, when calculating losses, supervised training methods such as generative pre-training, imitation learning, and reinforcement learning are included but not limited to.

[0178] The training method of the decision planning model provided in the embodiment of the present application can be achieved by training the second self-vehicle information, the second other-vehicle information, the second static landmark information, the second lane-level navigation information and the second traffic light information during the expert driving process, and continuously adjusting the decision planning model parameters through the expert's actual driving data (i.e., the true value) during the training process to obtain a decision planning model that can automatically make expert-level decision planning. That is, the trained decision planning model can be used to automatically learn in combination with the real-time actual driving environment to make planning decisions with safer, more comfortable and efficient driving effects, without relying on a large amount of manpower and material resources to collect and set rules, and there will be no sense of frustration caused by the deviation between the rules and the actual driving environment.

[0179] Based on the above method embodiment, another embodiment of the present application provides a decision planning device, such as Figure 4As shown, the device is implemented by a decision planning model, the decision planning model includes an encoding module and a decoding module, and the device includes:

[0180] an acquisition unit 310, configured to acquire first input information, wherein the first input information includes first self-vehicle information, first other-vehicle information, first static landmark information, first lane-level navigation information, and first traffic light information, wherein the first other-vehicle information includes information about a first other-vehicle around the first self-vehicle, the first static landmark information includes information about static landmarks around the first self-vehicle, the first lane-level navigation information includes lane-level navigation information for navigating the first self-vehicle, and the first traffic light information includes information about a first traffic light in front of the first self-vehicle;

[0181] An encoding unit 320, configured to perform attention encoding on the first input information based on the encoding module to obtain a first target vector;

[0182] The decoding unit 330 is configured to decode the first target vector based on the decoding module and the first query vector to obtain first output information, wherein:

[0183] The first output information includes a first planned trajectory of the first self-vehicle, a score of the first planned trajectory, and a first planning decision corresponding to the first planned trajectory; and / or, the first output information includes a second planned trajectory of the first other vehicle, a score of the second planned trajectory, and a second planning decision corresponding to the second planned trajectory;

[0184] The first query vector includes a first target query vector and / or a first object decision query vector, wherein the first target query vector includes a vector of a first target lane that the first self-vehicle is to plan to enter and / or a vector of a second target lane that the first other vehicle is to plan to enter; the first object decision query vector includes a vector of a decision corresponding to when the first self-vehicle enters the first target lane and / or a vector of a decision corresponding to when the first other vehicle enters the second target lane.

[0185] In a possible implementation, the encoding unit 320 includes:

[0186] A first encoding module, configured to perform instance-instance attention encoding on each piece of information in the first input information based on the encoding module to obtain an instance-instance vector;

[0187] The second encoding module is used to perform inter-instance attention encoding on each of the instance intra-vectors based on the encoding module to obtain the first target vector.

[0188] In a possible implementation, the acquisition unit 310 includes: a first acquisition module or a second acquisition module;

[0189] The first acquisition module is used to acquire the first static landmark information and the first lane-level navigation information from the lane-level navigation map information;

[0190] The second acquisition module is used to obtain input data of the landmark perception model, process the input data of the landmark perception model based on the landmark perception model, obtain the first static landmark information contained in the road environment image at the target time, and determine the first lane-level navigation information according to the extended navigation path, wherein the input data of the landmark perception model includes the road environment image at the target time and the first local posture at the target time, the first local posture includes the offset of the global posture of the first ego vehicle when collecting the road environment image relative to the global posture at the target starting point, the target time is any time when the first ego vehicle collects the road environment image on the extended navigation path, and the extended navigation path is an extended path of the vehicle-mounted navigation path.

[0191] In a possible implementation, the second acquisition module is also used to obtain a landmark training sample set before obtaining the input data of the landmark perception model, wherein each landmark training sample in the landmark training sample set includes: a road environment sample image group, a second local pose corresponding to each frame of the road environment sample image in the road environment sample image group, and a landmark true value corresponding to each frame of the road environment sample image, and the road environment image group includes multiple consecutive frames of road environment sample images; the landmark perception model is obtained by training using the landmark training sample set.

[0192] In a possible implementation manner, the device further includes:

[0193] A training unit, used for training the decision planning model;

[0194] The training unit comprises:

[0195] a third acquisition module, configured to acquire a decision planning training sample set, wherein each decision planning training sample in the decision planning training sample set includes second self-vehicle information, second other-vehicle information, second static landmark information, second lane-level navigation information, and second traffic light information, wherein the second other-vehicle information includes information about a second other-vehicle around the second self-vehicle, the second static landmark information includes static landmark information around the second self-vehicle, the second lane-level navigation information includes lane-level navigation information for navigating the second self-vehicle, and the second traffic light information includes information about a second traffic light in front of the second self-vehicle;

[0196] An encoding module, used for performing attention encoding on the decision planning training sample set based on the encoding module in the initial decision planning model to obtain a second target vector corresponding to each decision planning training sample;

[0197] A decoding module is used to decode the second target vector for each decision planning training sample based on the decoding module and the second query vector in the initial decision planning model to obtain second output information corresponding to each decision planning training sample, wherein:

[0198] The second output information includes a third planned trajectory of the second self-vehicle, a score of the third planned trajectory, and a third planning decision corresponding to the third planned trajectory; and / or, the second output information includes a fourth planned trajectory of the second other vehicle, a score of the fourth planned trajectory, and a fourth planning decision corresponding to the fourth planned trajectory;

[0199] The second query vector includes a second target query vector and / or a second object decision query vector, wherein:

[0200] The second target query vector includes a vector of a third target lane that the second self-vehicle is to be planned to enter and / or a vector of a fourth target lane that the second other vehicle is to be planned to enter; the second object decision query vector includes a vector of a decision corresponding to when the second self-vehicle enters the third target lane and / or a vector of a decision corresponding to when the second other vehicle enters the fourth target lane;

[0201] a calculation module, configured to calculate a loss value based on the third planned trajectory with the highest score, the trajectory truth value of the second self-vehicle, the fourth planned trajectory with the highest score, and the trajectory truth value of the second other-vehicle;

[0202] The adjustment training module is used to adjust the model parameters of the initial decision planning model when the loss value is greater than a preset loss threshold, and continue to train the decision planning model after adjusting the model parameters based on the decision planning training sample set until the loss value is less than or equal to the preset loss threshold, thereby obtaining the final decision planning model.

[0203] In a possible implementation, the target self-vehicle information includes the first self-vehicle information and / or the second self-vehicle information; when the target self-vehicle information is the first self-vehicle information, the target self-vehicle is the first self-vehicle, the target other-vehicle information is the first other-vehicle information, the target other-vehicle is the first other-vehicle, the target traffic light information is the first traffic light information, and the target traffic light is the first traffic light; when the target self-vehicle information is the second self-vehicle information, the target self-vehicle is the second self-vehicle, the target other-vehicle information is the second other-vehicle information, the target other-vehicle is the second other-vehicle, the target traffic light information is the second traffic light information, and the target traffic light is the second traffic light, the method includes:

[0204] The target vehicle information includes at least one of the position of the target vehicle relative to the target vehicle at the current moment within the preset historical time window, the global heading angle of the target vehicle within the preset historical time window, the actual speed of the target vehicle within the preset historical time window, the body size type of the target vehicle, the following gear position of the target vehicle at the current moment, and the target speed of the target vehicle; and / or,

[0205] The target other vehicle information includes at least one of the position of the target other vehicle relative to the target self-vehicle at the current moment within the preset historical time window, the global heading angle of the target other vehicle within the preset historical time window, the actual speed of the target other vehicle within the preset historical time window, the heading angle of the target other vehicle relative to the target self-vehicle at the current moment within the preset historical time window, and the body size type of the target other vehicle; and / or,

[0206] The target traffic light information includes the type and / or lighting status of the target traffic light within the preset historical time window.

[0207] The decision planning device provided in the embodiment of the present application can input first input information including first self-vehicle information, first other-vehicle information, first static landmark information, first lane-level navigation information and first traffic light information into the decision planning model, and firstly encode the first input information by the encoding module of the decision planning model to obtain a first target vector, and then decode the first target vector by the first query vector of the decoding module of the decision planning model to obtain first output information including a first planned trajectory of the first self-vehicle, a score of the first planned trajectory, and a first planning decision corresponding to the first planned trajectory, and / or first output information including a second planned trajectory of the first other-vehicle, a score of the second planned trajectory, and a second planning decision corresponding to the second planned trajectory, so as to subsequently determine the final first planned trajectory of the first self-vehicle and the corresponding first planning decision according to the first output information. It can be seen from this that the embodiments of the present application can use a pre-trained AI model to automatically make decisions and plans, and the first input information is real-time data automatically collected during the driving of the first vehicle, so the AI ​​model will automatically learn in combination with the real-time actual driving environment to make planning decisions with safer, more comfortable and efficient driving effects, without relying on a large amount of manpower and material resources to collect and set rules, and there will be no sense of frustration caused by deviations between the rules and the actual driving environment.

[0208] Based on the above method embodiment, another embodiment of the present application provides a training device for a decision planning model, such as Figure 5 As shown, the device comprises:

[0209] an acquisition unit 410, configured to acquire a decision planning training sample set, wherein each decision planning training sample in the decision planning training sample set includes second self-vehicle information, second other-vehicle information, second static landmark information, second lane-level navigation information, and second traffic light information, wherein the second other-vehicle information includes information about a second other-vehicle around the second self-vehicle, the second static landmark information includes static landmark information around the second self-vehicle, the second lane-level navigation information includes lane-level navigation information for navigating the second self-vehicle, and the second traffic light information includes information about a second traffic light in front of the second self-vehicle;

[0210] The encoding unit 420 is used to perform attention encoding on the decision planning training sample set based on the encoding module in the initial decision planning model to obtain a second target vector corresponding to each decision planning training sample;

[0211] The decoding unit 430 is used to decode the second target vector for each of the decision planning training samples based on the decoding module in the initial decision planning model and the second query vector to obtain second output information corresponding to each of the decision planning training samples, wherein:

[0212] The second output information includes a third planned trajectory of the second self-vehicle, a score of the third planned trajectory, and a third planning decision corresponding to the third planned trajectory; and / or, the second output information includes a fourth planned trajectory of the second other vehicle, a score of the fourth planned trajectory, and a fourth planning decision corresponding to the fourth planned trajectory;

[0213] The second query vector includes a second target query vector and / or a second object decision query vector, wherein:

[0214] The second target query vector includes a vector of a third target lane that the second self-vehicle is to be planned to enter and / or a vector of a fourth target lane that the second other vehicle is to be planned to enter; the second object decision query vector includes a vector of a decision corresponding to when the second self-vehicle enters the third target lane and / or a vector of a decision corresponding to when the second other vehicle enters the fourth target lane;

[0215] A calculation unit 440 is used to calculate a loss value based on the third planned trajectory with the highest score, the trajectory truth value of the second self-vehicle, the fourth planned trajectory with the highest score, and the trajectory truth value of the second other vehicle;

[0216] The adjustment training unit 450 is used to adjust the model parameters of the initial decision planning model when the loss value is greater than a preset loss threshold, and continue to train the decision planning model after adjusting the model parameters based on the decision planning training sample set until the loss value is less than or equal to the preset loss threshold, thereby obtaining the final decision planning model.

[0217] The training device for the decision planning model provided in the embodiment of the present application can be trained by the second self-vehicle information, the second other vehicle information, the second static landmark information, the second lane-level navigation information and the second traffic light information during the expert driving process, and the decision planning model parameters are continuously adjusted through the expert's actual driving data (i.e., the true value) during the training process to obtain a decision planning model that can automatically make expert-level decision planning. That is, the trained decision planning model can be used to automatically learn in combination with the real-time actual driving environment to make planning decisions with safer, more comfortable and efficient driving effects, without relying on a large amount of manpower and material resources to collect and set rules, and there will be no sense of frustration caused by the deviation between the rules and the actual driving environment.

[0218] Based on the above method embodiments, another embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the method described in any of the above embodiments is implemented.

[0219] Based on the above method embodiment, another embodiment of the present application provides an electronic device or a computer device, such as Figure 6 As shown, including:

[0220] one or more processors 510;

[0221] The processor 510 is coupled to a storage device 520, and the storage device 520 is used to store one or more programs;

[0222] When the one or more programs are executed by the one or more processors 510, the electronic device or computer device implements the method described in any of the above embodiments.

[0223] Based on the above method embodiments, another embodiment of the present application provides a vehicle, which includes the device as described in any of the above embodiments, or includes the electronic device as described above.

[0224] The vehicle includes a CPU (Central Processing Unit), a T-Box (Telematics Box), an image collector, a navigation and positioning device, a V2X device, an inertial navigation system, etc. Among them, the image collector is used to collect road environment images, the navigation and positioning device is used to locate the vehicle, obtain the global position and local position of the vehicle, the inertial navigation system is used to measure the actual speed of the vehicle, the global orientation angle and other information, and the V2X device is used to obtain information transmitted by other vehicles. The CPU is used to obtain first input information including first self-vehicle information, first other-vehicle information, first static landmark information, first lane-level navigation information and first traffic light information based on various information measured in the vehicle, and process the first input information through a decision planning model to obtain first output information including a first planned trajectory of the first self-vehicle, a score of the first planned trajectory, and a first planning decision corresponding to the first planned trajectory, and / or, including a second planned trajectory of the first other-vehicle, a score of the second planned trajectory, and a first output information of a second planning decision corresponding to the second planned trajectory, so as to make a final decision plan according to the first output information. The CPU can also send this input information to the server through the T-Box, and the server will use the decision-making planning model to make the final decision plan and feed the final decision plan back to the vehicle for regulation and control.

[0225] Based on the above embodiments, another embodiment of the present application provides a computer program product, which includes instructions. When the instructions are executed on a computer or a processor, the computer or the processor executes the method described in any of the above embodiments.

[0226] The above device embodiment corresponds to the method embodiment and has the same technical effect as the method embodiment. For specific description, please refer to the method embodiment. The device embodiment is obtained based on the method embodiment. For specific description, please refer to the method embodiment part, which will not be repeated here. It can be understood by those skilled in the art that the accompanying drawings are only schematic diagrams of one embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present application.

[0227] Those skilled in the art can understand that the modules in the device in the embodiment can be distributed in the device in the embodiment according to the description of the embodiment, or can be changed accordingly and located in one or more devices different from the embodiment. The modules in the above embodiment can be combined into one module, or can be further divided into multiple sub-modules.

[0228] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A decision planning method, characterized in that: The method is implemented by a decision planning model, the decision planning model includes an encoding module and a decoding module, and the method includes: Acquire first input information, wherein the first input information includes first self-vehicle information, first other-vehicle information, first static landmark information, first lane-level navigation information, and first traffic light information, the first other-vehicle information includes information of a first other-vehicle around the first self-vehicle, the first static landmark information includes static landmark information around the first self-vehicle, the first lane-level navigation information includes lane-level navigation information for navigating the first self-vehicle, and the first traffic light information includes information of a first traffic light in front of the first self-vehicle; Performing attention encoding on the first input information based on the encoding module to obtain a first target vector; The first target vector is decoded based on the decoding module and the first query vector to obtain first output information, wherein: The first output information includes a first planned trajectory of the first self-vehicle, a score of the first planned trajectory, and a first planning decision corresponding to the first planned trajectory; and / or, the first output information includes a second planned trajectory of the first other vehicle, a score of the second planned trajectory, and a second planning decision corresponding to the second planned trajectory; The first query vector includes a first target query vector and / or a first object decision query vector, wherein the first target query vector includes a vector of a first target lane that the first self-vehicle is to plan to enter and / or a vector of a second target lane that the first other vehicle is to plan to enter; the first object decision query vector includes a vector of a decision corresponding to when the first self-vehicle enters the first target lane and / or a vector of a decision corresponding to when the first other vehicle enters the second target lane.

2. The method according to claim 1, characterized in that The performing attention encoding on the first input information based on the encoding module to obtain a first target vector comprises: Based on the encoding module, each information in the first input information is respectively subjected to instance-by-instance attention encoding to obtain an instance-by-instance vector; Based on the encoding module, inter-instance attention encoding is performed on each of the instance intra-vectors to obtain the first target vector.

3. The method according to claim 1, characterized in that The method for obtaining the first static landmark information and the first lane-level navigation information includes: Acquire the first static landmark information and the first lane-level navigation information from lane-level navigation map information; or, Acquire input data of a landmark perception model, process the input data of the landmark perception model based on the landmark perception model, obtain the first static landmark information contained in the road environment image at a target moment, and determine the first lane-level navigation information according to an extended navigation path, wherein the input data of the landmark perception model includes the road environment image at the target moment and a first local posture at the target moment, the first local posture includes an offset of the global posture of the first ego vehicle when collecting the road environment image relative to the global posture at the target starting point, the target moment is any moment when the first ego vehicle collects the road environment image on the extended navigation path, and the extended navigation path is an extended path of the on-board navigation path.

4. The method according to claim 3, characterized in that: The method for generating the landmark perception model includes: Acquire a landmark training sample set, wherein each landmark training sample in the landmark training sample set includes: a road environment sample image group, a second local pose corresponding to each frame of the road environment sample image in the road environment sample image group, and a landmark true value corresponding to each frame of the road environment sample image, and the road environment image group includes multiple consecutive frames of road environment sample images; The landmark perception model is obtained by training using the landmark training sample set.

5. The method according to any one of claims 1 to 4, characterized in that The training method of the decision planning model includes: Acquire a decision planning training sample set, wherein each decision planning training sample in the decision planning training sample set includes second self-vehicle information, second other-vehicle information, second static landmark information, second lane-level navigation information, and second traffic light information, wherein the second other-vehicle information includes information about a second other-vehicle around the second self-vehicle, the second static landmark information includes static landmark information around the second self-vehicle, the second lane-level navigation information includes lane-level navigation information for navigating the second self-vehicle, and the second traffic light information includes information about a second traffic light in front of the second self-vehicle; Performing attention encoding on the decision planning training sample set based on the encoding module in the initial decision planning model to obtain a second target vector corresponding to each decision planning training sample; For each of the decision planning training samples, the second target vector is decoded based on the decoding module and the second query vector in the initial decision planning model to obtain second output information corresponding to each of the decision planning training samples, wherein: The second output information includes a third planned trajectory of the second self-vehicle, a score of the third planned trajectory, and a third planning decision corresponding to the third planned trajectory; and / or, the second output information includes a fourth planned trajectory of the second other vehicle, a score of the fourth planned trajectory, and a fourth planning decision corresponding to the fourth planned trajectory; The second query vector includes a second target query vector and / or a second object decision query vector, wherein the second target query vector includes a vector of a third target lane to be planned for the second self-vehicle to enter and / or a vector of a fourth target lane to be planned for the second other vehicle to enter; the second object decision query vector includes a vector of a decision corresponding to when the second self-vehicle enters the third target lane and / or a vector of a decision corresponding to when the second other vehicle enters the fourth target lane; Calculate the loss value based on the third planned trajectory with the highest score, the trajectory truth value of the second self-vehicle, the fourth planned trajectory with the highest score, and the trajectory truth value of the second other-vehicle; When the loss value is greater than a preset loss threshold, the model parameters of the initial decision planning model are adjusted, and the decision planning model with adjusted model parameters is continuously trained based on the decision planning training sample set until the loss value is less than or equal to the preset loss threshold, thereby obtaining the final decision planning model.

6. The method according to claim 5, characterized in that The target self-vehicle information includes the first self-vehicle information and / or the second self-vehicle information; when the target self-vehicle information is the first self-vehicle information, the target self-vehicle is the first self-vehicle, the target other-vehicle information is the first other-vehicle information, the target other-vehicle is the first other-vehicle, the target traffic light information is the first traffic light information, and the target traffic light is the first traffic light; when the target self-vehicle information is the second self-vehicle information, the target self-vehicle is the second self-vehicle, the target other-vehicle information is the second other-vehicle information, the target other-vehicle is the second other-vehicle, the target traffic light information is the second traffic light information, and the target traffic light is the second traffic light, the method includes: The target vehicle information includes at least one of the position of the target vehicle relative to the target vehicle at the current moment within the preset historical time window, the global heading angle of the target vehicle within the preset historical time window, the actual speed of the target vehicle within the preset historical time window, the body size type of the target vehicle, the following gear position of the target vehicle at the current moment, and the target speed of the target vehicle; and / or, The target other vehicle information includes at least one of the position of the target other vehicle relative to the target self-vehicle at the current moment within the preset historical time window, the global heading angle of the target other vehicle within the preset historical time window, the actual speed of the target other vehicle within the preset historical time window, the heading angle of the target other vehicle relative to the target self-vehicle at the current moment within the preset historical time window, and the body size type of the target other vehicle; and / or, The target traffic light information includes the type and / or lighting status of the target traffic light within the preset historical time window.

7. A training method for a decision planning model, characterized in that: The method comprises: Acquire a decision planning training sample set, wherein each decision planning training sample in the decision planning training sample set includes second self-vehicle information, second other-vehicle information, second static landmark information, second lane-level navigation information, and second traffic light information, wherein the second other-vehicle information includes information about a second other-vehicle around the second self-vehicle, the second static landmark information includes static landmark information around the second self-vehicle, the second lane-level navigation information includes lane-level navigation information for navigating the second self-vehicle, and the second traffic light information includes information about a second traffic light in front of the second self-vehicle; Performing attention encoding on the decision planning training sample set based on the encoding module in the initial decision planning model to obtain a second target vector corresponding to each decision planning training sample; For each of the decision planning training samples, the second target vector is decoded based on the decoding module in the initial decision planning model and the second query vector to obtain second output information corresponding to each of the decision planning training samples, wherein: The second output information includes a third planned trajectory of the second self-vehicle, a score of the third planned trajectory, and a third planning decision corresponding to the third planned trajectory; and / or, the second output information includes a fourth planned trajectory of the second other vehicle, a score of the fourth planned trajectory, and a fourth planning decision corresponding to the fourth planned trajectory; The second query vector includes a second target query vector and / or a second object decision query vector, wherein the second target query vector includes a vector of a third target lane to be planned for the second self-vehicle to enter and / or a vector of a fourth target lane to be planned for the second other vehicle to enter; the second object decision query vector includes a vector of a decision corresponding to when the second self-vehicle enters the third target lane and / or a vector of a decision corresponding to when the second other vehicle enters the fourth target lane; Calculate the loss value based on the third planned trajectory with the highest score, the trajectory truth value of the second self-vehicle, the fourth planned trajectory with the highest score, and the trajectory truth value of the second other-vehicle; When the loss value is greater than a preset loss threshold, the model parameters of the initial decision planning model are adjusted, and the decision planning model with adjusted model parameters is continuously trained based on the decision planning training sample set until the loss value is less than or equal to the preset loss threshold, thereby obtaining the final decision planning model.

8. A decision-making planning device, characterized in that: The device is implemented by a decision planning model, the decision planning model includes an encoding module and a decoding module, and the device includes: an acquisition unit, configured to acquire first input information, wherein the first input information includes first self-vehicle information, first other-vehicle information, first static landmark information, first lane-level navigation information, and first traffic light information, wherein the first other-vehicle information includes information about a first other-vehicle around the first self-vehicle, the first static landmark information includes information about static landmarks around the first self-vehicle, the first lane-level navigation information includes lane-level navigation information for navigating the first self-vehicle, and the first traffic light information includes information about a first traffic light in front of the first self-vehicle; an encoding unit, configured to perform attention encoding on the first input information based on the encoding module to obtain a first target vector; A decoding unit, configured to decode the first target vector based on the decoding module and the first query vector to obtain first output information, wherein: The first output information includes a first planned trajectory of the first self-vehicle, a score of the first planned trajectory, and a first planning decision corresponding to the first planned trajectory; and / or, the first output information includes a second planned trajectory of the first other vehicle, a score of the second planned trajectory, and a second planning decision corresponding to the second planned trajectory; The first query vector includes a first target query vector and / or a first object decision query vector, wherein the first target query vector includes a vector of a first target lane that the first self-vehicle is to plan to enter and / or a vector of a second target lane that the first other vehicle is to plan to enter; the first object decision query vector includes a vector of a decision corresponding to when the first self-vehicle enters the first target lane and / or a vector of a decision corresponding to when the first other vehicle enters the second target lane.

9. The device according to claim 8, characterized in that The encoding unit comprises: A first encoding module, configured to perform instance-instance attention encoding on each piece of information in the first input information based on the encoding module to obtain an instance-instance vector; The second encoding module is used to perform inter-instance attention encoding on each of the instance intra-vectors based on the encoding module to obtain the first target vector.

10. The device according to claim 8, characterized in that The acquisition unit includes: a first acquisition module or a second acquisition module; The first acquisition module is used to acquire the first static landmark information and the first lane-level navigation information from the lane-level navigation map information; The second acquisition module is used to obtain input data of the landmark perception model, process the input data of the landmark perception model based on the landmark perception model, obtain the first static landmark information contained in the road environment image at the target time, and determine the first lane-level navigation information according to the extended navigation path, wherein the input data of the landmark perception model includes the road environment image at the target time and the first local posture at the target time, the first local posture includes the offset of the global posture of the first ego vehicle when collecting the road environment image relative to the global posture at the target starting point, the target time is any time when the first ego vehicle collects the road environment image on the extended navigation path, and the extended navigation path is an extended path of the vehicle-mounted navigation path.

11. The device according to claim 10, characterized in that The second acquisition module is also used to obtain a landmark training sample set before obtaining the input data of the landmark perception model, wherein each landmark training sample in the landmark training sample set includes: a road environment sample image group, a second local pose corresponding to each frame of the road environment sample image in the road environment sample image group, and a landmark true value corresponding to each frame of the road environment sample image, and the road environment image group includes multiple consecutive frames of road environment sample images; the landmark perception model is obtained by training using the landmark training sample set.

12. The device according to any one of claims 8 to 11, characterized in that The device also includes: A training unit, used for training the decision planning model; The training unit comprises: a third acquisition module, configured to acquire a decision planning training sample set, wherein each decision planning training sample in the decision planning training sample set includes second self-vehicle information, second other-vehicle information, second static landmark information, second lane-level navigation information, and second traffic light information, wherein the second other-vehicle information includes information about a second other-vehicle around the second self-vehicle, the second static landmark information includes static landmark information around the second self-vehicle, the second lane-level navigation information includes lane-level navigation information for navigating the second self-vehicle, and the second traffic light information includes information about a second traffic light in front of the second self-vehicle; An encoding module, configured to perform attention encoding on the decision planning training sample set based on the initial decision planning model, and obtain a second target vector corresponding to each decision planning training sample; A decoding module is used to decode the second target vector for each decision planning training sample based on the decoding module and the second query vector in the initial decision planning model to obtain second output information corresponding to each decision planning training sample, wherein: The second output information includes a third planned trajectory of the second self-vehicle, a score of the third planned trajectory, and a third planning decision corresponding to the third planned trajectory; and / or, the second output information includes a fourth planned trajectory of the second other vehicle, a score of the fourth planned trajectory, and a fourth planning decision corresponding to the fourth planned trajectory; The second query vector includes a second target query vector and / or a second object decision query vector, wherein the second target query vector includes a vector of a third target lane to be planned for the second self-vehicle to enter and / or a vector of a fourth target lane to be planned for the second other vehicle to enter; the second object decision query vector includes a vector of a decision corresponding to when the second self-vehicle enters the third target lane and / or a vector of a decision corresponding to when the second other vehicle enters the fourth target lane; a calculation module, configured to calculate a loss value based on the third planned trajectory with the highest score, the trajectory truth value of the second self-vehicle, the fourth planned trajectory with the highest score, and the trajectory truth value of the second other-vehicle; The adjustment training module is used to adjust the model parameters of the initial decision planning model when the loss value is greater than a preset loss threshold, and continue to train the decision planning model after adjusting the model parameters based on the decision planning training sample set until the loss value is less than or equal to the preset loss threshold, thereby obtaining the final decision planning model.

13. The device according to claim 12, characterized in that The target self-vehicle information includes the first self-vehicle information and / or the second self-vehicle information; when the target self-vehicle information is the first self-vehicle information, the target self-vehicle is the first self-vehicle, the target other-vehicle information is the first other-vehicle information, the target other-vehicle is the first other-vehicle, the target traffic light information is the first traffic light information, and the target traffic light is the first traffic light; when the target self-vehicle information is the second self-vehicle information, the target self-vehicle is the second self-vehicle, the target other-vehicle information is the second other-vehicle information, the target other-vehicle is the second other-vehicle, the target traffic light information is the second traffic light information, and the target traffic light is the second traffic light, the method includes: The target vehicle information includes at least one of the position of the target vehicle relative to the target vehicle at the current moment within the preset historical time window, the global heading angle of the target vehicle within the preset historical time window, the actual speed of the target vehicle within the preset historical time window, the body size type of the target vehicle, the following gear position of the target vehicle at the current moment, and the target speed of the target vehicle; and / or, The target other vehicle information includes at least one of the position of the target other vehicle relative to the target self-vehicle at the current moment within the preset historical time window, the global heading angle of the target other vehicle within the preset historical time window, the actual speed of the target other vehicle within the preset historical time window, the heading angle of the target other vehicle relative to the target self-vehicle at the current moment within the preset historical time window, and the body size type of the target other vehicle; and / or, The target traffic light information includes the type and / or lighting status of the target traffic light within the preset historical time window.

14. A training device for a decision planning model, characterized in that: The device comprises: an acquisition unit, configured to acquire a decision planning training sample set, wherein each decision planning training sample in the decision planning training sample set includes second self-vehicle information, second other-vehicle information, second static landmark information, second lane-level navigation information, and second traffic light information, wherein the second other-vehicle information includes information about a second other-vehicle around the second self-vehicle, the second static landmark information includes static landmark information around the second self-vehicle, the second lane-level navigation information includes lane-level navigation information for navigating the second self-vehicle, and the second traffic light information includes information about a second traffic light in front of the second self-vehicle; An encoding unit, configured to perform attention encoding on the decision planning training sample set based on an encoding module in an initial decision planning model, to obtain a second target vector corresponding to each decision planning training sample; A decoding unit is used to decode the second target vector for each of the decision planning training samples based on the decoding module in the initial decision planning model and the second query vector to obtain second output information corresponding to each of the decision planning training samples, wherein: The second output information includes a third planned trajectory of the second self-vehicle, a score of the third planned trajectory, and a third planning decision corresponding to the third planned trajectory; and / or, the second output information includes a fourth planned trajectory of the second other vehicle, a score of the fourth planned trajectory, and a fourth planning decision corresponding to the fourth planned trajectory; The second query vector includes a second target query vector and / or a second object decision query vector, wherein the second target query vector includes a vector of a third target lane to be planned for the second self-vehicle to enter and / or a vector of a fourth target lane to be planned for the second other vehicle to enter; the second object decision query vector includes a vector of a decision corresponding to when the second self-vehicle enters the third target lane and / or a vector of a decision corresponding to when the second other vehicle enters the fourth target lane; a calculation unit, configured to calculate a loss value based on the third planned trajectory with the highest score, the trajectory truth value of the second self-vehicle, the fourth planned trajectory with the highest score, and the trajectory truth value of the second other-vehicle; The adjustment training unit is used to adjust the model parameters of the initial decision planning model when the loss value is greater than a preset loss threshold, and continue to train the decision planning model after adjusting the model parameters based on the decision planning training sample set until the loss value is less than or equal to the preset loss threshold, thereby obtaining the final decision planning model.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 6 or claim 7 is implemented.

16. An electronic device, characterized in that: The electronic device comprises: one or more processors; The processor is coupled to a storage device, and the storage device is used to store one or more programs; When the one or more programs are executed by the one or more processors, the electronic device implements the method according to any one of claims 1 to 6 or claim 7.

17. A vehicle, characterized in that: The vehicle comprises the apparatus according to any one of claims 8 to 13 or claim 14, or comprises the electronic device according to claim 16.

Citation Information

Cited By

  • Prediction information acquisition method, model training method, and device

    WO2026174451A1

  • Method for acquiring regulation and control information, and model training method and related device

    WO2026193843A1