Road network trajectory generation method and system based on pre-training, and medium

Abstract

The invention provides a road network trajectory generation method and system based on pre-training, and a medium. The method comprises the following steps: S1, carrying out pre-training based on simulation data to obtain a pre-trained autoregression Transform model; step S2, based on real data, carrying out fine adjustment on the auto-regression Transform model after pre-training; and S3, generating a road edge trajectory through autoregressive reasoning by the finely adjusted model. According to the invention, a large amount of data generated by a simulator is used for training and generating a model, and then the model is aligned with observation results of the real world. The generative model provides a higher fidelity than a simulator. Different from a model which only depends on real world data, the method is more effectively popularized to the road network of the whole city range. Besides, as long as a simulator running on the urban road network can be accessed in the pre-training step, a trajectory generation model can be pre-trained for the city, and the trajectory generation model is used as a basic model further aligned with a real world observation result.

Description

Technical Field

[0001] The present invention belongs to the field of computer software and algorithms, and relates to the problem of traffic trajectory prediction and generation algorithm in the field of urban transportation. Specifically, it relates to a road network trajectory generation method, system and medium based on pre-training. Background Art

[0002] Traffic research relies heavily on traffic simulation systems, which are essential for traffic policy making because they can efficiently conduct traffic policy research and testing. However, existing simulations are very different from real-world road traffic, which limits the applicability of simulator-based research results in reality. In contrast, data-driven research uses actual traffic data for tasks such as predictive modeling, but because the data comes from specific scenarios, the research results can usually only focus on specific real-world data sets.

[0003] Traffic simulation systems can be used to test various hypotheses and are crucial for decision makers in traffic policy research. However, existing traffic simulators face the challenge of lack of realism, which limits the applicability of the research results obtained from simulations in the real world. Taking the research on reinforcement learning in traffic signal control as an example, the bottleneck lies in the deviation between the simulator and the actual conditions, making it difficult for the learned decision model to be effectively applied in real-world scenarios.

[0004] A lot of research in the field of artificial intelligence focuses on real traffic data. Some studies explore generating trajectories by learning vehicle movements in actual urban environments, such as using generative adversarial networks (GANs) and diffusion methods to generate realistic trajectories. However, the key limitation is that it can only learn from observed data, which leads to the OOD Out of Distribution problem.

[0005] In existing research work, generating effective traffic flow from traffic networks has become a research hotspot in the field of transportation. Researchers use methods based on classic traffic models to generate traffic flow, build comprehensive traffic simulation tools, and calibrate traffic models or simulations with real data, but these physical models and simulations may be far from real-world traffic flow because the ideal assumptions do not match reality.

[0006] In terms of traffic flow generation models, some studies in recent years have used data-driven methods to generate urban traffic flow, using various methods and neural network structures to achieve traffic generation. However, there are some problems with current research, such as ignoring road network constraints, which may generate unrealistic trajectories, GAN-based methods that generate "routes" rather than "trajectories with timestamps", and diffusion models based on GPS locations that may generate non-continuous trajectories.

[0007] In terms of city-based models, they have become increasingly popular in the last two years since the emergence of GPT. There are many studies on different issues, such as using OpenStreetMap or satellite images to represent urban areas or points, using a single model for various spatial-temporal predictions, and using the capabilities of LLM to assist in traffic simulation processes.

[0008] Existing patent CN202210427356.8 discloses a road network generation method, device, readable storage medium and electronic device, the method comprising: obtaining multiple target trajectory data of the target area; predicting the target road grid map of the target area through a road prediction model based on the multiple target trajectory data, the road prediction model is trained based on multiple sample trajectory data of the first area and sample road network data of the second area, and the target area, the first area and the second area are different from each other; vectorizing the target road grid map to obtain the target road network data of the target area. However, this patent relies on sample trajectory data and sample road network data training models to generate road networks, which requires labeled data, is costly and inefficient. Moreover, the physical constraints of real-world trajectories are not considered, and the fidelity and scope of application are insufficient.

[0009] In summary, in order to solve the above shortcomings, it is urgent to study a road network trajectory generation method and system based on pre-training. Summary of the invention

[0010] In view of the defects in the prior art, the purpose of the present invention is to provide a method, system and medium for generating road network trajectories based on pre-training.

[0011] A method for generating a road network trajectory based on pre-training provided by the present invention comprises the following steps:

[0012] Step S1, pre-training based on simulated data to obtain a pre-trained autoregressive Transformer model;

[0013] Step S2, based on real data, fine-tuning the pre-trained autoregressive Transformer model to obtain a fine-tuned model;

[0014] Step S3: The fine-tuned model generates road edge trajectories through autoregressive reasoning.

[0015] Preferably, step S1 includes the following sub-steps:

[0016] Step S1.1, simulation data generation: determining multiple OD pairs of starting points and ending points from real data, and generating corresponding simulation data using a microscopic traffic simulator;

[0017] Step S1.2, pre-training stage: use simulated data to pre-train the autoregressive Transformer model to obtain a pre-trained autoregressive Transformer model.

[0018] Preferably, in step S1.1, possible OD pairs are enumerated from real data, and then the micro traffic simulator SUMO is used to simulate traffic conditions according to the OD pairs to generate trajectory data of the vehicle from the starting point to the end point as simulation data.

[0019] Preferably, in step S1.2, the autoregressive Transformer model adopts a Decoder-only Transformer architecture, including an encoder head, a Transformer backbone and a decoder head, wherein the encoder head independently processes each Token in the sequence, combining the embedding information of the historical Token and the target Token, the Transformer backbone consists of L Transformer layers, each layer contains a self-attention module and a feedforward network, plus residual connections and LayerNorm, a causal mask is applied to the attention weights to prevent positions from paying attention to subsequent positions, and the activation function uses SiLU. The decoder head is after the final transformer block, and the hidden vector in the sequence is decoded into a prediction of the next token through a standard linear decoder. The result at each position represents the log-likelihood value of the probability distribution of the Token at the next position, and the probability of generating a trajectory is expressed as the joint probability of generating each tag.

[0020] Preferably, during the pre-training process of step S1.2, each simulated data trajectory is decomposed into blocks of size B, and slid with a sliding step size of 1, and the trajectory of length T is converted into T-B+1 blocks, and pre-training is performed with the goal of minimizing the cross entropy loss.

[0021] Preferably, in step S2, in the fine-tuning stage, a KL regularization term is introduced into the loss function, and the pre-trained autoregressive Transformer model is further trained using real data.

[0022] Preferably, in step S2, the KL regularization term λ is introduced KL Into the loss function, fine-tune the loss L FT It is expressed as:

[0023] L FT =L PT +λ KL L KL

[0024]

[0025]

[0026] Where P represents the trajectory distribution of the real data, θ pt represents fixed pre-trained model parameters, θ ft Indicates the model parameters being fine-tuned. The dot “·” indicates a probability distribution rather than a single probability value.

[0027] Preferably, in step S3, the starting point o and the end point d are preset, o=e0 is set as the first marker of the sequence, and d is set as the target-oriented condition of the fine-tuned model. Then, the fine-tuned model is used to iteratively sample the next marker from the predicted distribution, and the sampled marker is appended to the end of the sequence until the preset end marker or the maximum length is reached.

[0028] The present invention also provides a road network trajectory generation system based on pre-training, comprising:

[0029] Module M1, pre-training based on simulated data to obtain a pre-trained autoregressive Transformer model;

[0030] Module M2, based on real data, fine-tunes the pre-trained autoregressive Transformer model to obtain a fine-tuned model;

[0031] Module M3, the fine-tuned model generates road edge trajectories through autoregressive reasoning.

[0032] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a method for generating a road network trajectory based on pre-training according to any one of claims 1 to 8.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. This paper proposes a method for solving the problem of learning to generate edge trajectories on a traffic network. This method first uses simulated data to pre-train the model and then fine-tunes it with real data. The performance is better than simulation methods and existing data-driven trajectory generation methods. The results show that using a large amount of simulated data in the pre-training step is beneficial, which suggests that the model can be used as a base model for further trajectory learning tasks and other downstream tasks.

[0035] 2. The present invention proposes to use a large amount of data generated by the simulator to train a pre-trained generative model, and then further align the model with real-world observations. Compared with the simulator, the generative model has higher fidelity. In addition, unlike models that rely only on real-world data, this method can be more effectively generalized to the entire city-wide road network. In the present invention, the pre-training step is applicable to any city because the only requirement is to have access to a simulator running on the city's road network. This means that a trajectory generation model can be pre-trained for any city and then used as a base model for further alignment with real-world observations.

[0036] 3. The model proposed in this paper is an autoregressive generative model based on a decoder-only Transformer architecture, which is particularly suitable for the generation of trajectory data. On a road network, a trajectory is essentially a sequence of edges that a vehicle crosses through the road network. To optimize the configuration of the generative model, a recently released real city-scale vehicle trajectory dataset from traffic camera video information is used. The dataset comes from Jinan and Shenzhen, China, and covers a road network containing more than 20,000 road edges.

[0037] 4. Experimental results show the effectiveness of combining simulated data and real data. Although the simulator can be calibrated using real data, the back-propagation learning mechanism in the model enables us to align with real data more effectively than the brute force search method commonly used for calibration. In addition, experiments have shown that even when observations in certain areas of the city are occluded, generative models based only on real data often produce trajectories that are far from the actual distribution, while the model proposed in this invention can still generate realistic trajectories in these occluded areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0039] Figure 1 A schematic diagram showing a continuous vehicle trajectory as a set of edges in an embodiment of the present invention;

[0040] Figure 2 Schematic diagram of the algorithm flow in the embodiment of the present invention;

[0041] Figure 3 This is a diagram of the overall architecture of the model in an embodiment of the present invention;

[0042] Figure 4 Schematic diagram of a large-scale trajectory data experiment conducted on a road network within a city in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0044] In the present invention, the road network is modeled as a graph G = (V, E), where the road intersections and road sections correspond to the vertex set V and edge set E of the graph respectively. The movement of vehicles in the city can be described by a trajectory consisting of a series of GPS position points. Since the vehicle must travel on the edge of the road network, the GPS trajectory of the vehicle can be mapped to the road network graph and converted into the trajectory of the road network edge on the graph.

[0045] A trajectory is expressed as s = {e0, e1, e2, ..., e i , ..., e T},

[0046] where e i ∈V represents the edge that the vehicle runs on at the i-th time step, and T is the total number of time steps for this trajectory.

[0047] Further: o = e0 and d = e T Indicates the start and end point of this trajectory.

[0048] Then: A set of trajectories can be expressed as: S = {s (1) ,s (2) , ..., s (j) , ..., s (N)}, where N represents the length of the trajectory.

[0049] Their starting and ending points are represented as:

[0050] OD={(o (1) , d (1) ), (o (2) , d (2) ),...,(o (j) , d (j) ),...,(o (N) , d (N) )}.

[0051] Figure 1 A schematic diagram showing a continuous vehicle trajectory expressed as a set of edges in an embodiment of the present invention.

[0052] like Figure 1 Shown are examples of vehicle trajectories.

[0053] Regarding the learning problem of road network edge trajectories, a road network G = (V, E) and a set of OD information containing the starting point and the end point are preset. The goal is to learn a mapping function q θ (s|o, d) to generate a corresponding set of trajectories S. This maximizes the probability of generating the real vehicle trajectory:

[0054]

[0055] The probability of each vehicle trajectory generation is expressed as:

[0056]

[0057] This problem is significantly different from previous work on GPS trajectory learning for the following reasons: (1) The road-side trajectory representation naturally takes into account the connectivity of the road network, avoiding the generation of unrealistic trajectories that do not conform to the constraints of real-world roads. (2) This algorithm enables the trajectory generator to learn more general trajectory patterns based on real road conditions, allowing the generator to learn how vehicles choose routes and adjust their speed on the road.

[0058] The present invention designs a training process combining simulation and reality, aiming to achieve two goals:

[0059] (1) Establish a link between generated data and real data.

[0060] (2) Ensure that the generated data complies with the basic physical constraints of real-world trajectories, such as continuity along the road network and reasonable driving speed.

[0061] Figure 2 Schematic diagram of the algorithm flow in an embodiment of the present invention.

[0062] like Figure 2 As shown, the present invention proposes a process to allow the model of the present invention to learn basic physical constraints from simulated data during pre-training, and then through a subsequent fine-tuning process with real data, the pre-trained basic model is matched with real-world driving behavior. This combination of real data and simulated data helps to solve the challenges of limited data in the real world and beyond the distribution range of real data.

[0063] For the neural network representation and learning of road network trajectories, the present invention designs a model structure based on autoregressive Transformer.

[0064] Embodiment 1:

[0065] This embodiment provides a method for generating a road network trajectory based on pre-training, comprising the following steps:

[0066] Step S1, pre-training is performed based on simulated data to obtain a pre-trained autoregressive Transformer model.

[0067] Specifically, step S1 includes the following sub-steps:

[0068] Step S1.1, simulation data generation; specifically, multiple OD pairs of starting points and end points are determined from real data, and corresponding simulation data are generated using a microscopic traffic simulator.

[0069] Specifically, possible OD pairs are enumerated from real data, and then the microscopic traffic simulator SUMO simulates traffic conditions according to the OD pairs to generate the trajectory data of the vehicle from the starting point to the end point as the simulation data. The purpose of pre-training is to enable the model to explore scenarios on different OD data distributions as much as possible and enhance the adaptability of the model. It is expected that the simulation data will teach the pre-trained basic model to grasp the basic physical properties when generating trajectories, such as the model can select continuous road network edges and choose relatively short paths in the network.

[0070] Step S1.2, pre-training stage; specifically, using simulated data to pre-train the autoregressive Transformer model to obtain a pre-trained autoregressive Transformer model.

[0071] In this example, each road edge is encoded as a label. To mimic the short-term dependencies observed in real trajectories, the autoregressive Transformer model is designed to consider at most the first B steps. The model accepts a label sequence block and the target tag d as input, and outputs a prediction of the probability of the next tag

[0072] Figure 3 This is a diagram of the overall architecture of the model in an embodiment of the present invention.

[0073] like Figure 3 As shown in the figure, the autoregressive Transformer model is designed to effectively represent trajectories. It adopts a Decoder-only Transformer architecture, including the encoder head, the Transformer backbone, and the decoder head. The encoder head independently processes each token in the sequence, combining the embedding information of the historical token and the target token. This process is expressed as:

[0074] h d =NormEmbedding(d),

[0075] h=NormEmbedding(s),

[0076] h0=MLP(Concat([h,h d ]))+E pos ,

[0077] Where: h d ∈R 1×C is the destination code, h∈R B×C is the encoding of the historical token, and C is the hidden layer size. Here Concat connects h d and h.E pos It is an absolute position code, and the final output code header is h0∈R B×C .

[0078] The Transformer backbone consists of L Transformer layers, each of which contains a self-attention module and a feed-forward network, all of which are added with residual connections and LayerNorm, causal masks are applied to the attention weights to prevent positions from paying attention to subsequent positions, and the activation function uses SiLU. This process is expressed as:

[0079] k takes values ​​from 1 to L:

[0080]

[0081]

[0082] in is an intermediate variable.

[0083] The decoder head, after the final transformer block, decodes the hidden vector in the sequence into a prediction of the next token through a standard linear decoder, where the result at each position represents the log-likelihood value of the probability distribution of the token at the next position.

[0084] y=Linear(LN(h L )),

[0085]

[0086] Where i = 1, 2, ..., B, y∈R B×|E| is the final output, y i ∈R 1×|E| is the i-th value of y.

[0087] Finally, the probability of generating a trajectory is expressed as the joint probability of generating each label:

[0088]

[0089] During the pre-training process of step S1.2, each simulated data trajectory is decomposed into blocks of size B and slid with a sliding step size of 1. The trajectory of length T is converted into T-B+1 blocks and pre-trained with the goal of minimizing the cross entropy loss.

[0090] Specifically, we use the previously generated trajectory data to pre-train a θ To avoid confusion of symbols, we define an abbreviation for the autoregressive Transformer model represented by The goal is to minimize the following cross entropy loss as the pre-training loss L PT ,

[0091]

[0092] in represents the simulated data trajectory distribution, and t represents the tail position of the block.

[0093] Step S2: fine-tune the pre-trained autoregressive Transformer model based on real data to obtain a fine-tuned model.

[0094] In the fine-tuning stage, the KL regularization term is introduced into the loss function, and the pre-trained autoregressive Transformer model is further trained using real data.

[0095] Specifically, in order to maintain the balance between pre-trained knowledge and fine-tuned knowledge, a KL regularization term λ is introduced KL to the loss function to prevent significant deviation from the pre-trained model and fine-tune the loss L FT It is expressed as:

[0096] L FT =L PT +λ KL L KL

[0097]

[0098]

[0099] Where P represents the trajectory distribution of the real data, q is a representation of the probability distribution, and θ pt represents fixed pre-trained model parameters, θ ft Indicates the model parameters being fine-tuned. The dot “·” indicates a probability distribution rather than a single probability value.

[0100] Step S3: The fine-tuned model generates road edge trajectories through autoregressive reasoning.

[0101] In this embodiment, the starting point o and the end point d are preset, o=e0 is set as the first marker of the sequence, and d is set as the target-oriented condition of the fine-tuned model. Then, the fine-tuned model is used to iteratively sample the next marker from the predicted distribution, and the sampled marker is appended to the end of the sequence until the preset end marker or the maximum length is reached.

[0102] Embodiment 2:

[0103] The present embodiment provides a road network trajectory generation system based on pre-training. The road network trajectory generation system based on pre-training can be implemented by executing the process steps of the road network trajectory generation method based on pre-training, that is, those skilled in the art can understand the road network trajectory generation method based on pre-training as a preferred implementation of the road network trajectory generation system based on pre-training.

[0104] The road network trajectory generation system includes:

[0105] Module M1 is pre-trained based on simulated data to obtain a pre-trained autoregressive Transformer model.

[0106] Module M2 fine-tunes the pre-trained autoregressive Transformer model based on real data to obtain a fine-tuned model.

[0107] Module M3, the fine-tuned model generates road edge trajectories through autoregressive reasoning.

[0108] Embodiment 3:

[0109] This embodiment provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned method for generating a road network trajectory based on pre-training are implemented.

[0110] The present invention combines the advantages of simulation and data-driven methods and introduces a generative vehicle trajectory model. The model is first trained using simulated traffic data based on real road network information and then aligned with real-world traffic observations.

[0111] Large-scale trajectory data experiments were conducted on a road network within a city, and the results fully demonstrated the effectiveness of the method of the present invention.

[0112] Figure 4 Schematic diagram of a large-scale trajectory data experiment conducted on a road network within a city in an embodiment of the present invention, wherein the upper left corner picture is a real trajectory, the upper right corner picture is a trajectory generated by the present invention, and the pictures in the second row are trajectories generated by other methods.

[0113] like Figure 4As shown, the trajectory comparison chart generated by different methods clearly shows that the trajectory generated by the present invention is closer to the real trajectory and is significantly better than the trajectory generated by other methods. At the same time, comparing the model effect tables under different indicators, taking Jinan and Shenzhen as examples, under the JSD (Jensen Shannon Divergence) and WD (Jensen Shannon Divergence) indicators, the performance of the present invention is significantly better than methods such as DiffTraj and TS-TrajGen. Specifically, the JSD and WD indicators of the present invention in Jinan reached 0.047 and 38.482 respectively, and the JSD and WD indicators in Shenzhen were 0.081 and 216.291 respectively, and the improvement was obvious, which fully proved the excellent performance of the present invention in road network trajectory generation. The lower the JSD and WD indicators, the better.

[0114] Table 1: Model performance under different indicators

[0115]

[0116] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0117] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for generating a road network trajectory based on pre-training, characterized in that: The steps include: Step S1, pre-training based on simulated data to obtain a pre-trained autoregressive Transformer model; Step S2, fine-tuning the pre-trained autoregressive Transformer model based on real data to obtain a fine-tuned model; Step S3: the fine-tuned model generates road edge trajectories through autoregressive reasoning.

2. The method for generating a road network trajectory based on pre-training according to claim 1, characterized in that: The step S1 includes the following sub-steps: Step S1.1, simulation data generation: determining multiple OD pairs of starting points and ending points from real data, and generating corresponding simulation data using a microscopic traffic simulator; Step S1.2, pre-training stage: using the simulated data to pre-train the autoregressive Transformer model to obtain a pre-trained autoregressive Transformer model.

3. The method for generating a road network trajectory based on pre-training according to claim 2, characterized in that: In the step S1.1, possible OD pairs are enumerated from real data, and then the micro traffic simulator SUMO simulates the traffic situation according to the OD pairs to generate the trajectory data of the vehicle from the starting point to the end point as simulation data.

4. The method for generating a road network trajectory based on pre-training according to claim 2, characterized in that: In step S1.2, the autoregressive Transformer model adopts a Decoder-only Transformer architecture, including an encoder head, a Transformer backbone and a decoder head, wherein the encoder head independently processes each Token in the sequence, combining the embedding information of historical Tokens and target Tokens, the Transformer backbone consists of L Transformer layers, each layer contains a self-attention module and a feedforward network, plus residual connections and LayerNorm, a causal mask is applied to the attention weights to prevent positions from paying attention to subsequent positions, and the activation function uses SiLU. The decoder head is after the final transformer block, and decodes the hidden vector in the sequence into a prediction of the next token through a standard linear decoder. The result at each position represents the log-likelihood value of the probability distribution of the Token at the next position, and the probability of generating a trajectory is expressed as the joint probability of generating each tag.

5. The method for generating a road network trajectory based on pre-training according to claim 4, characterized in that: During the pre-training process of step S1.2, each simulated data trajectory is decomposed into blocks of size B and slid with a sliding step size of 1. The trajectory of length T is converted into T-B+1 blocks, and pre-training is performed with the goal of minimizing the cross entropy loss.

6. The method for generating a road network trajectory based on pre-training according to claim 1, characterized in that: In step S2, in the fine-tuning stage, a KL regularization term is introduced into the loss function, and the pre-trained autoregressive Transformer model is further trained using the real data.

7. The method for generating a road network trajectory based on pre-training according to claim 1, characterized in that: In step S2, the KL regularization term λ is introduced KL Into the loss function, fine-tune the loss L FT It is expressed as: L FT =L PT +λ KL L KL Where P represents the trajectory distribution of the real data, θ pt represents fixed pre-trained model parameters, θ ft Indicates the model parameters being fine-tuned, and the dot "·" represents a probability distribution rather than a single probability value.

8. The method for generating a road network trajectory based on pre-training according to claim 1, characterized in that: In step S3, the starting point o and the end point d are preset, o=e0 is set as the first marker of the sequence, and d is set as the target-oriented condition of the fine-tuned model. Then, the fine-tuned model is used to iteratively sample the next marker from the predicted distribution, and the sampled marker is appended to the end of the sequence until the preset end marker or the maximum length is reached.

9. A road network trajectory generation system based on pre-training, characterized in that: include: Module M1, pre-training based on simulated data to obtain a pre-trained autoregressive Transformer model; Module M2, based on real data, fine-tuning the pre-trained autoregressive Transformer model to obtain a fine-tuned model; Module M3, the fine-tuned model generates road edge trajectories through autoregressive reasoning.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of a method for generating a road network trajectory based on pre-training described in any one of claims 1 to 8 are implemented.

Images