A guide line generation method and system based on simulation feedback data

Abstract

The application provides a guide line generation method and system based on simulation feedback data, and the method comprises the following steps: generating a predicted guide line trajectory based on a trajectory planning model according to a plurality of historical passing trajectories; generating a control feasible guide line trajectory through a simulation platform based on the predicted guide line trajectory; calculating the similarity between the predicted guide line trajectory and the control feasible guide line trajectory; if the two guide line trajectories are not similar, the control feasible guide line trajectory is taken as the predicted guide line trajectory; and if the two guide line trajectories are similar, the predicted guide line trajectory is kept. The application combines the planned guide line trajectory and the simulated guide line trajectory, feeds back the structure of the trajectory planning model through the simulation platform, constructs a new data set, and is used for training and optimizing the trajectory planning model, so that the finally generated lane-level guide line is more in line with the driving behavior habit.

Description

Technical Field

[0001] This invention relates to the field of crowdsourced high-precision map production, and more specifically, to a method and system for generating guide lines based on simulation feedback data. Background Technology

[0002] In crowdsourced high-precision map creation, guide lines are a map element of lanes. Lane-level guide lines are safe, comfortable, and drivable tracks that can be used to guide autonomous vehicles through a section of lane.

[0003] In regular and well-structured lanes with clear and precise features, the lane centerline is usually used as the lane guide line. However, at road intersections where there is no clear road structure, planning a guide line for two connected lanes is a very challenging task.

[0004] Guide lines, a type of static guide line in action planning, are typically generated offline and pre-defined in high-precision maps. Current methods for generating intersection guide lines can be broadly categorized into two types. One type is based on mathematical models, which model the intersection as a non-convex mathematical optimization problem. It defines the objective function and adheres to constraints such as traffic regulations and collision avoidance, using curve fitting and numerical optimization algorithms to generate safe and comfortable guide lines. This method heavily relies on the completeness and accuracy of road elements in the map, such as lane edges and center lines, and is unsuitable for free spaces with unclear road structures. Furthermore, the generated trajectories often fail to meet human comfort and vehicle control requirements. The other type is based on data-driven learning, represented by deep neural networks. It constructs large-scale trajectory datasets and trains a trajectory planning model to generate guide line trajectories that better reflect human driving behavior. However, training neural network-based planning models requires a large amount of trajectory data, and the dataset should cover as many traffic scenarios as possible to improve the model's generalization performance. In addition, the results of neural network-based models are unstable, and evaluating them with real vehicles is costly and even difficult to implement. Summary of the Invention

[0005] This invention addresses the technical problems existing in the prior art by providing a method and system for generating guide lines based on simulation feedback data.

[0006] According to a first aspect of the present invention, a method for generating guide lines based on simulation feedback data is provided, comprising:

[0007] Based on multiple historical travel trajectories, a predicted guide line trajectory is generated using a trajectory planning model.

[0008] Based on the predicted guide line trajectory, a controllable guide line trajectory is generated through a simulation platform;

[0009] Calculate the similarity between the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform;

[0010] If the predicted guideline trajectory generated by the trajectory planning model is not similar to the controllable guideline trajectory generated by the simulation platform, then the controllable guideline trajectory generated by the simulation platform is used as the predicted guideline trajectory; if they are similar, the predicted guideline trajectory is retained.

[0011] Based on the above technical solution, the present invention can also be improved as follows.

[0012] Optionally, the trajectory planning model is a conditional variational autoencoder (CVAE), and the basic mathematical expression of the CVAE is:

[0013] T = g(z, V);

[0014] Given z and V, the desired guide trajectory T is generated by function g. T is a sequence of points consisting of latitude and longitude, z is a fixed-dimensional random vector that follows a Gaussian distribution, V is the lane driving behavior semantic vector extracted by the model, and g is the function expression of the trajectory planning model.

[0015] Optionally, the conditional variational autoencoder (CVAE) includes multiple ST blocks, fusion blocks, multi-layer fully connected networks, and a decoding network. The training process of the trajectory planning model is as follows:

[0016] Obtain a training sample set, which includes multiple samples, each of which includes multiple historical passage trajectories and real guide line trajectories;

[0017] Based on ST block, the spatiotemporal feature vectors of the actual guide line trajectory and each historical passage trajectory are extracted;

[0018] Based on Fusion block, the spatiotemporal feature vectors of each historical traffic trajectory are fused to generate lane representation vectors;

[0019] The spatiotemporal feature vector of the actual guideline trajectory and the lane representation vector are combined, and the prior distribution parameters of the prior network are predicted based on the multilayer fully connected network. The prior distribution parameters include the mean μ and the logarithm of the squared variance logσ of the prior multivariate normal distribution. 2 ;

[0020] Sample a vector ∈ of a specified dimension from the standard normal distribution ∈ ~ N(0,1), whose dimension is consistent with the dimension of the prior multivariate normal distribution, and calculate the vector z = ∈ × σ + μ;

[0021] The decoding network combines the vector z and the lane representation vector to generate a predicted guide line.

[0022] Optionally, the ST block includes a fully connected layer (FC), a convolutional network (CONV), and a recurrent neural network (LSTM).

[0023] Based on the fully connected layer FC, the longitude and latitude of the sampling points in the real guide line trajectory and each historical travel trajectory are encoded into vectors of a specified dimension;

[0024] The spatial features of the sampling points in the real guide line trajectory and each historical passage trajectory are extracted using the convolutional network CONV.

[0025] Based on the recurrent neural network LSTM, the spatiotemporal feature vectors of the real guide line trajectory with temporal characteristics and the sampling points in each historical passage trajectory are extracted.

[0026] Optionally, both the actual guide line trajectory and each historical passage trajectory consist of a sequence of trajectory sampling points with temporal characteristics. The step of generating the predicted guide line trajectory based on a trajectory planning model using multiple historical passage trajectories further includes:

[0027] Converting the actual guide line trajectory and the sampling points in each historical travel trajectory from global Earth coordinates to vehicle local coordinates includes:

[0028] According to formula (2), the global coordinates are converted into local coordinates in the same direction:

[0029]

[0030] Where x and y represent longitude and latitude values ​​respectively, superscripts l and g represent local and global coordinates respectively, and subscript i represents the i-th sampling point;

[0031] Alternatively, global coordinates can be converted to local coordinates based on formulas (3) and (4):

[0032] x new =x old cosθ-y old sinθ (3);

[0033] y new =x old sinθ+y old cosθ (4);

[0034] Where, x new ,y new Let x represent the longitude and latitude of the new local coordinates, respectively. old ,y oldThese represent the longitude and latitude of the old local coordinates, respectively. The direction in which the vehicle enters the intersection is the positive y-axis, and the right side of the y-axis is a Cartesian coordinate system with the positive y-axis direction to the right.

[0035] Alternatively, global coordinates can be converted to local coordinates based on formula (5):

[0036]

[0037] The action coordinates of the current sampling point are its coordinate value minus the coordinate value of the previous sampling point.

[0038] Optionally, the step of generating a predicted guideline trajectory based on multiple historical travel trajectories and a trajectory planning model further includes:

[0039] Calculate the loss and gradient values ​​between the predicted guideline trajectory and the actual guideline trajectory, and adjust the network hyperparameters of the trajectory planning model based on the loss and gradient values.

[0040] Optionally, calculating the similarity between the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform includes:

[0041] Based on an LSTM twin network, binary classification is implemented to determine the similarity between the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform.

[0042] According to a second aspect of the present invention, a system based on simulation feedback data is provided, comprising a trajectory planning module, a simulation platform, a calculation module, and an update module;

[0043] The trajectory planning module is used to generate a predicted guide line trajectory based on multiple historical traffic trajectories and a trajectory planning model.

[0044] A simulation platform is used to generate a controllable guideline trajectory based on the predicted guideline trajectory.

[0045] The calculation module is used to calculate the similarity between the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform;

[0046] The update module is used to use the controllable guideline trajectory generated by the simulation platform as the predicted guideline trajectory if the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform are not similar; otherwise, the predicted guideline trajectory is retained.

[0047] According to a third aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the processor is configured to implement the steps of a guideline generation method based on simulation feedback data when executing a computer management program stored in the memory.

[0048] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer management class program is stored, wherein when executed by a processor, the computer management class program implements the steps of a guide line generation method based on simulation feedback data.

[0049] This invention provides a guideline generation method and system based on simulation feedback data. It combines the planned guideline trajectory with the simulated guideline trajectory, and constructs a new dataset through feedback from the simulation platform on the structure of the trajectory planning model. This dataset is used to train and optimize the trajectory planning model, so that the final generated lane-level guidelines are more in line with driving behavior habits. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of a data closed-loop framework based on simulation feedback;

[0051] Figure 2 A flowchart of a guide line generation method based on simulation feedback data provided by the present invention;

[0052] Figure 3 This is a schematic diagram illustrating the training process of a trajectory planning model based on CVAE-generated guide lines.

[0053] Figure 4 This is a schematic diagram of the inference process of the trajectory planning model;

[0054] Figure 5 This is a schematic diagram of an LSTM twin network;

[0055] Figure 6 A schematic diagram of a guide line generation system based on simulation feedback data provided by the present invention;

[0056] Figure 7 A schematic diagram of the hardware structure of a possible electronic device provided by the present invention;

[0057] Figure 8 This is a schematic diagram of the hardware structure of a possible computer-readable storage medium provided by the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined with each other to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0059] Existing data-driven trajectory planning mainly uses neural network models, which have high requirements for datasets and high real-vehicle evaluation costs. This patent proposes a trajectory planning scheme that automatically constructs datasets for new scenarios and has low evaluation costs.

[0060] See Figure 1 This paper presents a data closed-loop framework based on simulation feedback, which mainly consists of three parts: a dataset, a trajectory planning model based on neural networks, and a simulation platform.

[0061] 1) The dataset section is mainly used to manage training data and generate training and test sets for trajectory planning model training.

[0062] 2) Core module of the trajectory planning model based on neural networks: A trajectory planning model based on a Conditional Variational Autoencoder (CVAE) was designed. The basic mathematical form is described as follows:

[0063] T = g(z, V);

[0064] The formula means that, given z and V, the desired guideline trajectory T is generated through the function g. T is a sequence of points composed of latitude and longitude, z is a fixed-dimensional random vector following a Gaussian distribution, V is the semantic vector of lane driving behavior extracted by the model, and g is the generative model, obtained through data learning. This patent designs a spatiotemporal feature extraction model, ST-block, to extract the spatiotemporal features of the trajectory. ST-block consists of fully connected layers, convolutional modules, and recurrent neural networks. To ensure the successful implementation of the model, relevant technical measures need to be adopted.

[0065] Figure 2A flowchart of a guided line generation method based on simulation feedback data provided by the present invention is shown below. Figure 1 As shown, the guide line generation method includes:

[0066] S1 generates a predicted guide line trajectory based on multiple historical travel trajectories and a trajectory planning model.

[0067] Understandably, guide lines need to be generated in real time to guide vehicles during their journey. Therefore, a predicted guide line can be generated using a trajectory planning model based on the generated actual guide line trajectory and multiple historical vehicle trajectories.

[0068] The actual guide line trajectory and historical traffic trajectory are represented by a sequence of coordinate points, each including longitude and latitude. The longitude and latitude in the original trajectory are Earth coordinate values ​​and cannot be used as input values ​​for model calculation. Sampling point coordinates represent the conversion of global Earth coordinates into local coordinate values ​​suitable for model calculation. This invention employs three coordinate transformation methods: 1) coordinate transformation in the same direction as the global coordinate system; 2) coordinate transformation with the vehicle's lane entry heading angle as the y-axis; and 3) coordinate transformation based on the trajectory behavior sequence.

[0069] 1) Local coordinate transformation in the same direction as the global coordinate system, as shown in formula (2), where

[0070]

[0071] Where x and y represent longitude and latitude values ​​respectively, the superscripts l and g represent local and global coordinates respectively, and the subscript i represents the i-th sampling point.

[0072] 2) The direction in which the vehicle enters the intersection is the positive y-axis, and the right-hand side of the y-axis is a Cartesian coordinate system with the positive y-axis direction, as shown in formulas (3) and (4). This method requires rotating the coordinates.

[0073] x new =x old cosθ-y old sinθ formula (3);

[0074] y new =x old sinθ+y old cosθ formula (4);

[0075] Global coordinates are converted to local coordinates using the method described in 1), where x new ,y new Let x represent the longitude and latitude of the new local coordinates, respectively. old ,y old These represent the longitude and latitude of the old local coordinates, respectively.

[0076] 3) Trajectory behavior sequence, that is, the behavior coordinate of the current sampling point is its coordinate value minus the coordinate value of the previous sampling point, as shown in formula (5), where the meaning of the letter symbols is the same as in 1), and the behavior coordinate of the first sampling point is itself minus.

[0077]

[0078] As an example, the Conditional Variational Autoencoder (CVAE) includes multiple ST blocks, a Fusion block, a multilayer fully connected network, and a decoding network. The training process of the trajectory planning model is as follows: A training sample set is acquired, comprising multiple samples, each sample including multiple historical traffic trajectories and real guide line trajectories; spatiotemporal feature vectors of the real guide line trajectories and each historical traffic trajectory are extracted based on the ST blocks; the spatiotemporal feature vectors of each historical traffic trajectory are fused based on the Fusion block to generate a lane representation vector; the spatiotemporal feature vectors of the real guide line trajectories and the lane representation vector are combined, and the prior distribution parameters of the prior network are predicted based on the multilayer fully connected network. The prior distribution parameters include the mean μ and the logarithm of the squared variance logσ of the prior multivariate normal distribution. 2 Sample a vector of a specified dimension from the standard normal distribution ∈ ~ N(0,1), whose dimension is consistent with the dimension of the prior multivariate normal distribution, and calculate the vector z = ∈ × σ + μ; combine the vector z and the lane representation vector through the decoding network to generate a predicted guide line.

[0079] Understandably, see Figure 3 This is a schematic diagram of the structure of a Conditional Variational Autoencoder (CVAE), which mainly includes multiple ST blocks, Fusion blocks, multi-layer fully connected networks, and a decoding network. The functions of the multiple sub-models are as follows:

[0080] Feature encoding of trajectory points involves encoding them into high-dimensional feature vectors. In this invention, trajectory points are two-dimensional independent features of longitude and latitude, requiring encoding into higher-dimensional feature vectors to improve model performance. This patent employs a fully connected (FC) neural network to encode trajectory points into vectors of a specified dimension, with its network parameters trained together with the overall network.

[0081] The Spatial-Temporal Block (ST block) is a sub-model for encoding the spatiotemporal features of a trajectory, extracting its spatiotemporal feature vectors. Since the point sequence of a trajectory possesses both spatial and temporal features, extracting high-level, semantically meaningful spatiotemporal vectors from the original trajectory is crucial to this method. To this end, this invention designs the ST block, composed of a fully connected (FC) layer, a convolutional layer (CONV), and a Long Short-Term Memory (LSTM) artificial neural network.

[0082] The lane feature representation sub-model, Fusion Block, extracts lane representation vectors that represent lane structure and driving behavior. Each lane contains multiple historical traffic trajectories, each providing information to the lane representation vector. The lane representation vector is obtained by fusing the spatiotemporal feature vectors of these multiple trajectories.

[0083] Calculate the prior distribution of the sample, i.e., assume the sample follows a normal distribution N(μ,σ). 2 The mean and variance of a normal distribution are calculated. During the training of CVAE, the output of the encoding phase is the mean and variance of a normal distribution, which are calculated using FC.

[0084] The decoding layer decodes the input feature vector into a sequence of points for the trajectory. This model is an autoregressive model that relies on the previously predicted trajectory point positions and lane representation vectors.

[0085] Figure 3 In this module, the ST-block is the spatiotemporal feature extraction module, mainly composed of FC, CONV, and LSTM, which extracts the spatiotemporal features of the trajectory. FC stands for fully connected layer, CONV stands for convolutional layer, and LSTM stands for recurrent neural network layer. The fusion block merges the spatiotemporal vectors of multiple historical trajectories, resulting in a lane representation vector. X represents the guide line trajectory of the lane, V... X C represents the spacetime vector of the guide line. i The i-th historical traffic trajectory of lane V i V represents the spatiotemporal vector of the i-th historical travel trajectory. s This represents the lane representation vector after fusion through the Fusion block. μ, logσ 2 Let represent the logarithm of the mean and the squared variance of the prior distribution predicted by the network, respectively, and z be the sampling vector of the prior distribution.

[0086] For trajectory planning models based on neural networks, the process includes training and inference. Specifically, the training process is as follows:

[0087] 1. Input the actual guide line trajectory X into ST block, and input C. i This is the historical trajectory.

[0088] 2. Convert the trajectory points in global coordinates to local coordinates, with the origin of the local coordinates being the center point of the lane trajectory.

[0089] 3. The ST block encodes the actual guide line trajectory X to obtain its spatiotemporal feature vector V. X .

[0090] 4. For each historical travel trajectory C i Encoding, extracting historical travel trajectories C via ST-block i The spatiotemporal characteristics of C are obtained. i spatiotemporal eigenvector V i .

[0091] 5. The spatiotemporal feature vectors V of multiple historical travel trajectories i Lane representation vector V is generated through Fusion-block fusion. s ;

[0092] 6. Combine the spatiotemporal feature vector of the actual guiding line trajectory X with V X Lane representation vector V s The method jointly predicts the mean and logarithm of squared variance of the prior network through a multilayer fully connected network (including multiple FC layers), i.e., the prior distribution parameters, including the mean μ and logarithm of squared variance logσ of the prior multivariate normal distribution. 2 .

[0093] 7. Sample a vector of a specified dimension from the standard normal distribution ∈ ~ N(0,1), with the dimension being consistent with the dimension of the prior multivariate normal distribution, and calculate the vector z = ∈ × σ + μ, which is used as the input to the decoding network.

[0094] 8. Vector z, through the FC layer and lane representation vector V s joint;

[0095] 9. The LSTM network outputs multiple values ​​in a loop, and the FC layer takes the output of the LSTM as input to generate the predicted trajectory points.

[0096] 10. During the training of the trajectory planning model, the loss function and gradient values ​​are calculated between the generated predicted guideline trajectory and the input real guideline trajectory. The network parameters of the trajectory planning model are then updated based on the loss function and gradient values. During inference, the generated sequence of trajectory points constitutes the generated predicted guideline.

[0097] It should be noted that the loss function used in training the trajectory planning model is:

[0098]

[0099] in Represents the recognition network, learning the posterior distribution, p θ For the prior network, learn the prior distribution, Z is the noise vector, and X is the trajectory of the guide line. i Let T be the i-th point of the guiding line trajectory, and T be the trajectory generated by the model. i Generate the i-th point of the trajectory for the model, C = (C1, C2, ..., C n ),C i This represents the i-th trajectory within the lane.

[0100] The training algorithm is a mini-batch-based backpropagation algorithm, and the learning strategy is RMSprop.

[0101] Please refer to the process of generating predictive guide lines based on the trained trajectory planning model. Figure 4 The inference process of the trajectory planning model is as follows:

[0102] 1. Enter C i This is the historical trajectory of the lane.

[0103] 2. Analyze the historical trajectory of lane C separately. i Encoding, obtaining C through ST-block i spatiotemporal eigenvector V i .

[0104] 3. The spatiotemporal feature vectors V of multiple trajectories i Lane representation vector V is generated through Fusion-block fusion. s .

[0105] 4. Sample a vector z of a specified dimension from the standard normal distribution z ~ N(0,1). The vector z is then processed by the lane representation vector V through an FC layer. s joint.

[0106] 5. The LSTM network outputs multiple values ​​in a loop, and the FC layer takes the output of the LSTM as input to generate the predicted guide line.

[0107] It should be noted that a schematic diagram of the internal structure of the ST block can be found in [link to schematic diagram]. Figure 5 The ST block includes fully connected layers (FC), convolutional networks (CONV), and recurrent neural networks (LSTM).

[0108] The longitude and latitude of the sampling points in the real guide line trajectory and each historical passage trajectory are encoded into vectors of specified dimensions based on the fully connected layer FC; the spatial features of the sampling points in the real guide line trajectory and each historical passage trajectory are extracted through the convolutional network CONV; and the spatiotemporal feature vectors of the sampling points in the real guide line trajectory and each historical passage trajectory with temporal features are extracted based on the recurrent neural network LSTM.

[0109] Understandably, the spatiotemporal feature vectors of trajectory sequences are extracted based on the ST block. The ST block comprises a fully connected (FC) layer, a convolutional network (CONV), and a recurrent neural network (LSTM). The main steps for extracting these vectors include: feature encoding of sampling points in the trajectory: FC layers encode the longitude and dimension of the trajectory sampling points into fixed-dimensional semantic vectors. This is a vector representation of the high-level features of the sampling points, encoding the original longitude and dimension features into higher-level features. Spatial feature extraction of the trajectory sequence: Spatial relationships exist between neighboring points in the trajectory sequence. The CONV convolutional network is used to extract these spatial relationships. Feature encoding of the trajectory sequence: Since the trajectory sequence has temporal features, extracting semantic features is crucial. The LSTM network is primarily used to extract the semantic feature vectors of the trajectory, i.e., the spatiotemporal feature vectors.

[0110] S2, Based on the predicted guide line trajectory, a controllable guide line trajectory is generated through a simulation platform.

[0111] Understandably, the simulation platform primarily uses the Carla simulation platform, an open-source simulator for autonomous driving that supports the development, training, and verification of autonomous driving systems. The simulation platform mainly includes an MPC controller, which, based on a trajectory planning model, generates a predicted guideline trajectory to plan a controllably feasible guideline trajectory.

[0112] The Carla platform, through its MPC controller, satisfies the lateral force coefficient and curvature constraints to generate a controllable guideline trajectory, composed of a sequence of points.

[0113]

[0114] S3, calculate the similarity between the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform.

[0115] S4. If the predicted guide trajectory generated by the trajectory planning model is not similar to the controllable guide trajectory generated by the simulation platform, then the controllable guide trajectory generated by the simulation platform is used as the predicted guide trajectory; if they are similar, the predicted guide trajectory is retained.

[0116] As an example, binary classification is implemented based on LSTM twin network to determine the similarity between the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform.

[0117] Understandably, the above steps generate predicted guideline trajectories through a trajectory planning model and controllable guideline trajectories through a simulation platform, respectively. This step uses an LSTM Siamese network to perform binary classification and determine the similarity between T and T'. Since T' is a controllable guideline trajectory, if T and T' are not similar, T' is added to the new dataset as a predicted guideline trajectory; if T and T' are similar, it indicates that the predicted guideline trajectory generated by the trajectory planning model is controllable and feasible, and T is added to the new dataset as a predicted guideline trajectory.

[0118] Figure 6 A structural diagram of a guideline generation system based on simulation feedback data is provided for an embodiment of the present invention, such as... Figure 6 As shown, a guide line generation system based on simulation feedback data includes a trajectory planning module 601, a simulation platform 602, a calculation module 603, and an update module 604, wherein:

[0119] The trajectory planning module 601 is used to generate a predicted guide line trajectory based on the trajectory planning model, according to the actual guide line trajectory and multiple historical traffic trajectories.

[0120] Simulation platform 602 is used to generate a controllable guide line trajectory based on the predicted guide line trajectory;

[0121] Calculation module 603 is used to calculate the similarity between the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform;

[0122] The update module 604 is used to take the controllable guide line trajectory generated by the simulation platform as the predicted guide line trajectory if the predicted guide line trajectory generated by the trajectory planning model and the controllable guide line trajectory generated by the simulation platform are not similar; if they are similar, the predicted guide line trajectory is retained.

[0123] It is understood that the guide line generation system based on simulation feedback data provided by the present invention corresponds to the guide line generation method based on simulation feedback data provided in the foregoing embodiments. The relevant technical features of the guide line generation system based on simulation feedback data can be referred to the relevant technical features of the guide line generation method based on simulation feedback data, and will not be repeated here.

[0124] Please see Figure 7 , Figure 7 This is a schematic diagram illustrating an embodiment of the electronic device provided in this invention. For example... Figure 7 As shown, an embodiment of the present invention provides an electronic device 700, including a memory 710, a processor 720, and a computer program 711 stored in the memory 710 and executable on the processor 720. When the processor 720 executes the computer program 711, it implements the steps of a guide line generation method based on simulation feedback data.

[0125] Please see Figure 8 , Figure 8 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided by the present invention. (See diagram below.) Figure 8 As shown, this embodiment provides a computer-readable storage medium 800, on which a computer program 811 is stored. When the computer program 811 is executed by a processor, it implements the steps of a guide line generation method based on simulation feedback data.

[0126] This invention provides a guideline generation method and system based on simulation feedback data. It combines the planned guideline trajectory with the simulated guideline trajectory, and constructs a new dataset through feedback from the simulation platform on the structure of the trajectory planning model. This dataset is used to train and optimize the trajectory planning model, making the final generated lane-level guidelines more in line with driving behavior habits.

[0127] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0128] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0129] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0130] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0131] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0132] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0133] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for generating guide lines based on simulation feedback data, characterized in that, include: Based on multiple historical travel trajectories, a predicted guide line trajectory is generated using a trajectory planning model. Based on the predicted guide line trajectory, a controllable guide line trajectory is generated through a simulation platform; Calculate the similarity between the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform; If the predicted guideline trajectory generated by the trajectory planning model is not similar to the controllable guideline trajectory generated by the simulation platform, then the controllable guideline trajectory generated by the simulation platform shall be used as the predicted guideline trajectory. If similar, maintain the predicted guideline trajectory; The trajectory planning model is a Conditional Variational Autoencoder (CVAE), which includes multiple ST blocks, Fusion blocks, multi-layer fully connected networks, and a decoding network. The training process of the trajectory planning model is as follows: Obtain a training sample set, which includes multiple samples, each of which includes multiple historical passage trajectories and real guide line trajectories; Based on ST block, the spatiotemporal feature vectors of the actual guide line trajectory and each historical passage trajectory are extracted; Based on Fusion block, the spatiotemporal feature vectors of each historical traffic trajectory are fused to generate lane representation vectors; The spatiotemporal feature vector of the actual guideline trajectory and the lane representation vector are combined, and the prior distribution parameters of the prior network are predicted based on the multilayer fully connected network. The prior distribution parameters include the mean of the prior multivariate normal distribution. Logarithm of the square of the variance ; From the standard normal distribution Mid-sampling of vectors of specified dimensions Its dimension is consistent with the dimension of the prior multivariate normal distribution, and the vector is calculated. ; The decoding network combines the vector z and the lane representation vector to generate a predicted guide line.

2. The guide line generation method according to claim 1, characterized in that, The fundamental mathematical expression for the conditional variational autoencoder (CVAE) is: ； Among them, it is known , In the case of using a function Generate the desired guide line trajectory , A sequence of points consisting of latitude and longitude. For a fixed-dimensional random vector that follows a Gaussian distribution, The lane driving behavior semantic vector extracted by the model, This is the function expression for the trajectory planning model.

3. The guide line generation method according to claim 1, characterized in that, The ST block includes a fully connected layer (FC), a convolutional network (CONV), and a recurrent neural network (LSTM). Based on the fully connected layer FC, the longitude and latitude of the sampling points in the real guide line trajectory and each historical travel trajectory are encoded into vectors of a specified dimension; The spatial features of the sampling points in the real guide line trajectory and each historical passage trajectory are extracted using the convolutional network CONV. Based on the recurrent neural network LSTM, the spatiotemporal feature vectors of the real guide line trajectory with temporal characteristics and the sampling points in each historical passage trajectory are extracted.

4. The guide line generation method according to claim 1, characterized in that, The actual guide line trajectory and each historical passage trajectory are both composed of a sequence of trajectory sampling points with temporal characteristics. The step of generating a predicted guide line trajectory based on a trajectory planning model using multiple historical passage trajectories also includes: Converting the actual guide line trajectory and the sampling points in each historical travel trajectory from global Earth coordinates to vehicle local coordinates includes: According to formula (2), global coordinates are converted into local coordinates in the same direction: in These represent longitude and latitude values ​​respectively, with superscripts. Representing local and global coordinates respectively, subscripts Indicates the first One sampling point; Alternatively, global coordinates can be converted to local coordinates based on formulas (3) and (4): in, These represent the longitude and latitude of the new local coordinates, respectively. These represent the longitude and latitude of the old local coordinates, respectively, and the direction in which the vehicle enters the intersection is... Positive axis direction The right side of the axis is A Cartesian coordinate system along the positive axis; Alternatively, global coordinates can be converted to local coordinates based on formula (5): The behavior coordinates of the current sampling point are its coordinate value minus the coordinate value of the previous sampling point.

5. The guide line generation method according to claim 1, characterized in that, The step of generating a predicted guideline trajectory based on multiple historical travel trajectories and a trajectory planning model also includes: Calculate the loss and gradient values ​​between the predicted guideline trajectory and the actual guideline trajectory, and adjust the network hyperparameters of the trajectory planning model based on the loss and gradient values.

6. The guide line generation method according to claim 1, characterized in that, The calculation of the similarity between the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform includes: Based on an LSTM twin network, binary classification is implemented to determine the similarity between the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform.

7. A guide line generation system based on simulation feedback data, characterized in that, It includes a trajectory planning module, a simulation platform, a calculation module, and an update module; The trajectory planning module is used to generate a predicted guide line trajectory based on multiple historical traffic trajectories and a trajectory planning model. A simulation platform is used to generate a controllable guideline trajectory based on the predicted guideline trajectory. The calculation module is used to calculate the similarity between the predicted guideline trajectory generated by the trajectory planning model and the controllable guideline trajectory generated by the simulation platform; The update module is used to use the controllable guide trajectory generated by the simulation platform as the predicted guide trajectory if the predicted guide trajectory generated by the trajectory planning model and the controllable guide trajectory generated by the simulation platform are not similar. If similar, maintain the predicted guideline trajectory; The trajectory planning model is a Conditional Variational Autoencoder (CVAE), which includes multiple ST blocks, Fusion blocks, multi-layer fully connected networks, and a decoding network. The training process of the trajectory planning model is as follows: Obtain a training sample set, which includes multiple samples, each of which includes multiple historical passage trajectories and real guide line trajectories; Based on ST block, the spatiotemporal feature vectors of the actual guide line trajectory and each historical passage trajectory are extracted; Based on Fusion block, the spatiotemporal feature vectors of each historical traffic trajectory are fused to generate lane representation vectors; The spatiotemporal feature vector of the actual guideline trajectory and the lane representation vector are combined, and the prior distribution parameters of the prior network are predicted based on the multilayer fully connected network. The prior distribution parameters include the mean of the prior multivariate normal distribution. Logarithm of the square of the variance ; From the standard normal distribution Mid-sampling of vectors of specified dimensions Its dimension is consistent with the dimension of the prior multivariate normal distribution, and the vector is calculated. ; The decoding network combines the vector z and the lane representation vector to generate a predicted guide line.

8. An electronic device, characterized in that, It includes a memory and a processor, wherein the processor is used to execute computer management programs stored in the memory to implement the steps of the guide line generation method based on simulation feedback data as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, It stores a computer management program, which, when executed by a processor, implements the steps of the guide line generation method based on simulation feedback data as described in any one of claims 1-6.

Images