Lane-changing trajectory generation method and system based on trajectory start and end point adjustment and vehicle dynamics, terminal and storage medium

Abstract

The present application relates to the technical field of lane-changing trajectory generation and traffic simulation, and discloses a lane-changing trajectory generation method and system based on trajectory start and end point adjustment and vehicle dynamics, a terminal and a storage medium, the method comprising: performing risk assessment on initial lane-changing samples, performing risk promotion processing on unqualified initial lane-changing samples to obtain a target lane-changing sample set; constructing an initial lane-changing model, performing open-loop training on the initial lane-changing model according to the initial lane-changing samples to obtain a pre-training model; performing closed-loop training on the pre-training model using the target lane-changing sample set, and performing lane-changing simulation on a to-be-tested lane-changing vehicle through the trained model to obtain a simulated lane-changing trajectory. The present application simultaneously performs time series coding on the lane-changing vehicle and the historical trajectories of multiple vehicles around the lane-changing vehicle, introduces an attention mechanism to adaptively assign interaction weights, captures the interaction relationship between vehicles changing over time, improves the capture accuracy of the model on complex traffic dynamic characteristics, and makes the output result more accurate.

Description

Technical Field

[0001] This invention relates to the field of trajectory generation and traffic simulation analysis technology, and in particular to a method, system, terminal, and computer-readable storage medium for generating lane-changing trajectories based on trajectory start and end point adjustment and vehicle dynamics. Background Technology

[0002] In the field of traffic simulation, lane-changing behavior modeling and simulation techniques fall into two categories: mechanism-driven methods and data-driven methods.

[0003] Lane-changing behavior in actual driving is influenced by a variety of complex factors, such as driver intent, vehicle performance, road structure, traffic density, and environmental disturbances. There are highly nonlinear and uncertain relationships between various variables of the surrounding vehicles and the longitudinal and lateral accelerations of the lane-changing vehicle. This makes mechanistic methods face certain limitations in characterizing real lane-changing behavior. In this case, mechanistic-driven lane-changing models are too simplistic and lack flexibility, while data-driven methods are limited by imbalanced samples and have difficulty generating high-risk behaviors. Both have the problem of insufficient ability to model high-risk lane-changing behaviors.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a method, system, terminal, and computer-readable storage medium for generating lane-changing trajectories based on trajectory start and end point adjustment and vehicle dynamics. This aims to solve the problem that existing lane-changing models are insufficient in modeling high-risk lane-changing behaviors, resulting in inaccurate simulation results in autonomous driving tests and extreme traffic scenarios.

[0006] To achieve the above objectives, the present invention provides a lane-change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics. The lane-change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics includes the following steps: Multiple initial lane change samples are obtained, a risk assessment is performed on each initial lane change sample, and risk enhancement processing is applied to the trajectory start point and / or trajectory end point of each low-risk lane change sample to finally obtain a target lane change sample set that meets the requirements. An initial lane-changing model is constructed using a long short-term memory encoder, an interactive feature model, and a long short-term memory decoder. The initial lane-changing model is then trained in an open loop based on all the initial lane-changing samples to obtain a pre-trained model. The pre-trained model is trained in a closed loop using the target lane-changing sample set to obtain the target lane-changing model. The target lane-changing model is then used to simulate the lane-changing of the vehicle under test to obtain the simulated lane-changing trajectory of the vehicle under test.

[0007] Optionally, the lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics, wherein obtaining multiple initial lane-changing samples specifically includes: Construct a vehicle lane-changing trajectory generation model and obtain the total lateral displacement, lane-changing speed, and lane-changing time of the lane-changing vehicle; The total lateral displacement, the lane-changing speed, and the lane-changing time are input into the vehicle lane-changing trajectory generation model for simulation, and all initial lane-changing samples of the lane-changing vehicle during the test period are output. Assess the risk of each initial lane change sample, and define all initial lane change samples whose risk assessment value is higher than a preset risk threshold as target lane change samples; For all initial lane change samples whose risk assessment value is not higher than the preset risk threshold, risk enhancement processing is performed on all initial lane change samples according to the trajectory starting point of each initial lane change sample to obtain target lane change samples that meet the requirements. All the target lane change samples are encapsulated to obtain the target lane change sample set.

[0008] Optionally, in the lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics, the initial lane-changing sample includes: lateral displacement and longitudinal displacement. The lateral displacement and the longitudinal displacement are expressed as follows:

[0009] in, Indicates the current time t Lateral displacement, Indicates the current time t longitudinal displacement, Indicates the total lateral displacement. Indicates lane change time. A coefficient representing lane-changing time. This represents the lane-changing characteristic coefficient. Indicates the lane-changing speed.

[0010] Optionally, the lane-change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics, wherein the step of performing risk enhancement processing on all initial lane-change samples whose risk assessment value is not higher than the preset risk threshold, based on the trajectory start point of each initial lane-change sample, to obtain a target lane-change sample that meets the requirements, specifically includes: Obtain the lane change start point, lane change end point, and longitudinal speed of all the initial lane change samples. Randomly select multiple initial lane change samples, pre-set the lane change start point of each initial lane change sample by a distance, and keep the lane change end point and longitudinal speed unchanged to improve the risk level of each initial lane change sample and obtain the corresponding target lane change sample. Randomly select multiple initial lane change samples, pre-set the lane change termination point of each initial lane change sample by a distance, and keep the lane change start point and the longitudinal speed unchanged, so as to increase the risk threshold of each initial lane change sample and obtain the corresponding target lane change sample.

[0011] Optionally, the lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics, wherein the construction of an initial lane-changing model using a long short-term memory encoder, an interactive feature model, and a long short-term memory decoder, and the open-loop training of the initial lane-changing model based on all the initial lane-changing samples to obtain a pre-trained model, specifically includes: A long short-term memory encoder, an interaction feature model, and a long short-term memory decoder are constructed respectively, and an initial lane-changing model is constructed based on the long short-term memory encoder, the interaction feature model, and the long short-term memory decoder; The target lane change sample set is input into the initial lane change model for initial prediction, and the corresponding prediction result is output. A loss function is constructed using the target lane-changing sample set and the corresponding prediction results. The initial lane-changing model is then trained using the loss function to obtain a pre-trained model.

[0012] Optionally, the lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics, wherein the construction of a long short-term memory encoder, an interaction feature model, and a long short-term memory decoder, and the construction of an initial lane-changing model based on the long short-term memory encoder, the interaction feature model, and the long short-term memory decoder, specifically includes: A long short-term memory encoder, an interaction feature model, and a long short-term memory decoder are constructed respectively. Identify the lane-changing vehicle and multiple surrounding vehicles of the lane-changing vehicle, and construct historical temporal features of the lane-changing vehicle and all surrounding vehicles based on the long short-time memory encoder; Based on the interaction feature model, construct the interaction relationship information between the lane-changing vehicle and all the surrounding vehicles; Based on the historical time series features and the interaction relationship information, the predicted coordinate sequence of the lane-changing vehicle is constructed according to the long short-term memory decoder; Based on the predicted coordinate sequence, an initial lane-changing model is constructed between the lane-changing vehicle and all surrounding vehicles: ; ; ; in, Indicates the predicted trajectory. express and The nonlinear relationship between them This indicates the historical trajectory of vehicles changing lanes. Indicating spatial relationships, Indicates the step size of the historical trajectory. Indicates the step size in traffic simulation. Indicates a time step. Indicates vehicles changing lanes at Coordinates at time, Indicates vehicles changing lanes at Coordinate position at time, Indicates vehicles changing lanes at Coordinates at time, Indicates surrounding vehicles The historical trajectory of moments, Indicates surrounding vehicles The historical trajectory of moments, Indicates surrounding vehicles The historical trajectory of a moment.

[0013] Optionally, the lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics, wherein the step of using the target lane-changing sample set to perform closed-loop training on the pre-trained model to obtain the target lane-changing model specifically includes: The target lane change sample set is input into the pre-trained model, wherein the number of samples in the target lane change sample set is greater than the number of samples that cause the pre-trained model to converge. The pre-trained model predicts each target lane change sample and outputs the simulated displacement of the vehicle under test under each target lane change sample. For each simulated displacement, the vehicle under test is controlled to perform a lane-changing simulation based on the simulated displacement to obtain the lane-changing position of the vehicle under test; Determine whether the simulation time of the lane change simulation of the vehicle under test is equal to the actual lane change time. If not, replace the simulation time with the actual lane change time, calculate the mean square error between the lane change simulation trajectory and the actual lane change trajectory, update the neural network parameters of the pre-trained model according to the mean square error, and obtain the target lane change model after the pre-trained model converges.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a lane-changing trajectory generation system based on trajectory start and end point adjustment and vehicle dynamics, wherein the lane-changing trajectory generation system based on trajectory start and end point adjustment and vehicle dynamics includes: The sample construction module is used to acquire multiple initial lane change samples, perform risk assessment on each initial lane change sample, and perform risk enhancement processing on the trajectory start point and / or trajectory end point of each low-risk lane change sample, so as to finally obtain a target lane change sample set that meets the requirements. The model pre-training module is used to construct an initial lane-changing model using a long short-term memory encoder, an interactive feature model, and a long short-term memory decoder, and to perform open-loop training on the initial lane-changing model based on all the initial lane-changing samples to obtain a pre-trained model. The training and simulation module is used to input the target lane-changing sample set into the pre-trained model for closed-loop training, use the trained model to simulate lane-changing for the vehicle under test, and output the simulated lane-changing trajectory of the vehicle under test.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a lane change trajectory generation program based on trajectory start and end point adjustment and vehicle dynamics stored in the memory and executable on the processor. When the lane change trajectory generation program based on trajectory start and end point adjustment and vehicle dynamics is executed by the processor, it implements the steps of the lane change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics as described above.

[0016] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a lane-change trajectory generation program based on trajectory start and end point adjustment and vehicle dynamics, and the lane-change trajectory generation program based on trajectory start and end point adjustment and vehicle dynamics, when executed by a processor, implements the steps of the lane-change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics as described above.

[0017] In this invention, multiple initial lane-change samples are acquired, and a risk assessment is performed on each initial lane-change sample. Risk enhancement processing is applied to the trajectory start and / or trajectory end points of each low-risk lane-change sample to ultimately obtain a target lane-change sample set that meets the requirements. An initial lane-change model is constructed using a Long Short-Term Memory (LSTM) encoder, an interaction feature model, and an LSM decoder. The initial lane-change model is then trained in an open-loop manner using all the initial lane-change samples to obtain a pre-trained model. The target lane-change sample set is input into the pre-trained model for closed-loop training. The trained model is then used to simulate lane changes for the vehicle under test, outputting the simulated lane-change trajectory of the vehicle under test. This invention also performs temporal encoding on the historical trajectories of the lane-changing vehicle and its surrounding multiple vehicles, introducing an attention mechanism to adaptively allocate interaction weights, capturing the time-varying interaction relationships between vehicles, improving the model's accuracy in capturing complex traffic dynamics, and resulting in more accurate output results. Attached Figure Description

[0018] Figure 1 This is a flowchart of a preferred embodiment of the lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics of the present invention; Figure 2 This is a schematic diagram of sample generation of a preferred embodiment of the lane change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics of the present invention. Figure 3 This is a schematic diagram of model transfer of a preferred embodiment of the lane change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics of the present invention. Figure 4 This is a closed-loop training diagram of a preferred embodiment of the lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics of the present invention. Figure 5 This is a comparison diagram of the results of a preferred embodiment of the lane change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics of the present invention; Figure 6 This is a structural diagram of a preferred embodiment of the lane change trajectory generation system based on trajectory start and end point adjustment and vehicle dynamics of the present invention. Figure 7 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] Mechanism-driven lane-changing decision models are mostly based on pre-defined rules, typically determining whether to execute a lane-changing operation through threshold judgments or conditional triggers. This approach struggles to capture the behavioral differences and strategic flexibility exhibited by drivers in real-world driving, resulting in overly conservative and rigid lane-changing strategies. Lane-changing execution models primarily prioritize safety and comfort, employing dynamic programming or optimal control methods to generate smooth and reasonable lane-changing trajectories. These execution strategies tend towards "ideal lane changes," lacking the ability to represent emergency, aggressive, or risky lane-changing maneuvers, thus failing to encompass the intense or sudden lane-changing behaviors commonly encountered in real-world traffic.

[0021] Current data-driven methods widely employ publicly available datasets for model training and validation. However, the vast majority of lane-changing samples in these datasets represent normal or safe lane-changing behaviors, while high-risk lane-changing samples are extremely scarce and suffer from severe imbalance. Without introducing sample augmentation or risk behavior reconstruction mechanisms, directly using such data for training often results in models that can only learn the characteristics of conventional lane-changing behaviors, making it difficult to effectively identify, simulate, and generate lane-changing behaviors with high-risk characteristics.

[0022] The lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics described in the preferred embodiment of the present invention, such as... Figure 1 As shown, the lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics includes the following steps: Step S10 obtains multiple initial lane change samples, performs a risk assessment on each initial lane change sample, and performs risk enhancement processing on the trajectory start point and / or trajectory end point of each low-risk lane change sample, finally obtaining a target lane change sample set that meets the requirements.

[0023] In order to address the limitations of existing mechanism-driven and data-driven methods in modeling high-risk lane-changing behavior, this invention constructs a hybrid-driven modeling method for high-risk lane-changing behavior. By combining high-risk behavior sample augmentation technology with a deep learning model, it can effectively generate lane-changing behaviors with high-risk characteristics.

[0024] Specifically, a vehicle lane-changing trajectory generation model is constructed, and the total lateral displacement, lane-changing speed, and lane-changing time of the lane-changing vehicle are obtained; The total lateral displacement, the lane-changing speed, and the lane-changing time are input into the vehicle lane-changing trajectory generation model for simulation, and all initial lane-changing samples of the lane-changing vehicle during the test period are output. Assess the risk of each initial lane change sample, and define all initial lane change samples whose risk assessment value is higher than a preset risk threshold as target lane change samples; For all initial lane change samples whose risk assessment value is not higher than the preset risk threshold, risk enhancement processing is performed on all initial lane change samples according to the trajectory starting point of each initial lane change sample to obtain target lane change samples that meet the requirements. All the target lane change samples are encapsulated to obtain the target lane change sample set.

[0025] Among them, such as Figure 2 As shown, in order to increase the number of high-risk lane change samples and improve the generalization ability of the model in high-risk scenarios, this invention constructs a lane change trajectory generation model based on trajectory start and end point adjustment and vehicle dynamics to generate high-risk lane change samples (i.e., initial lane change samples). While maintaining the physical rationality of the trajectory, a variety of high-risk lane change samples are generated by reasonably modifying the artificial risk enhancement operation of the original lane change trajectory.

[0026] Furthermore, the initial lane-changing sample includes: lateral displacement and longitudinal displacement; The lateral displacement and the longitudinal displacement are expressed as follows:

[0027] in, Indicates the current time t Lateral displacement, Indicates the current time t longitudinal displacement, Indicates the total lateral displacement. Indicates lane change time. A coefficient representing lane-changing time. This represents the lane-changing characteristic coefficient. Indicates the lane-changing speed.

[0028] In one of the embodiments disclosed in this invention, a lane-changing trajectory generation model based on a linear two-degree-of-freedom vehicle dynamics model is constructed, with the aim of ensuring that the enhanced lane-changing trajectory conforms to the motion characteristics of vehicles in real traffic.

[0029] Specifically, in modeling the lane-changing process of vehicles, the transfer function between wheel rotation angle and vehicle yaw rate and lateral displacement is derived, and physical parameters such as vehicle mass, wheelbase, moment of inertia and tire lateral stiffness are explicitly introduced. The expressions for lateral displacement and longitudinal displacement can directly generate a continuous, differentiable, and physically reasonable lane-changing trajectory given vehicle parameters, lane-changing time and lane-changing speed.

[0030] Specifically, the starting point, ending point and longitudinal speed of all the initial lane change samples are obtained, multiple initial lane change samples are randomly selected, the starting point of each initial lane change sample is preset by a distance, and the ending point and longitudinal speed are kept unchanged, so as to improve the risk level of each initial lane change sample and obtain the corresponding target lane change sample. Randomly select multiple initial lane change samples, pre-set the lane change termination point of each initial lane change sample by a distance, and keep the lane change start point and the longitudinal speed unchanged, so as to improve the risk level of each initial lane change sample and obtain the corresponding target lane change sample.

[0031] The process, based on a vehicle lane-changing trajectory generation model, firstly identifies the lane-changing start and end points of the lane-changing behavior trajectory. For low-risk lane-changing samples, a risk enhancement strategy is randomly selected, and the original trajectory between the modified start and end points is extracted. The lane-changing time and average speed required for the vehicle to complete the lane change are calculated. Based on the lane-changing time and average speed, a new lane-changing trajectory is generated and replaces the original trajectory. It is then determined whether the newly generated trajectory has collided with surrounding vehicles (the collision criterion is that at any given moment, the geometric boundary of the lane-changing vehicle overlaps with any surrounding vehicle). The risk of the newly generated lane-changing behavior sample is assessed, and high-risk lane-changing behavior samples that meet the requirements are collected. The risk assessment indicators may include, but are not limited to, Time to Collision (TTC), Post Encroachment Time (PET), and Deceleration Rate to Avoid Collision (DRAC).

[0032] The first risk enhancement strategy increases lane change risk by adjusting the starting point of the lane change trajectory. While keeping the lane change termination point and longitudinal speed constant, the starting point of the lane change is moved forward, allowing the lane-changing vehicle to enter the target lane earlier, significantly shortening the longitudinal and lateral distance to the vehicle behind in the target lane, reducing the buffer space, and increasing the risk of collision.

[0033] The second risk enhancement strategy increases lane-changing risk by adjusting the termination point of the lane-changing trajectory. While keeping the starting point unchanged, the termination point is moved forward, shortening the lane-changing time and increasing lateral acceleration, thereby reducing the reaction time of vehicles following in the target lane and increasing the probability of conflict. These two methods not only enhance the risk of the trajectory but also alter its shape, thus increasing sample diversity and contributing to the generation of richer high-risk lane-changing samples.

[0034] These two risk enhancement strategies apply to different stages of the lane-changing process: the starting point forward strategy tends to increase the vertical conflict risk in the initial stage of lane changing; the ending point forward strategy tends to increase the horizontal conflict risk in the execution stage of lane changing. In practical applications, the strategy can be selected based on the target risk type (vertical risk or horizontal risk), or both strategies can be used simultaneously to construct different types of lane-changing samples.

[0035] This invention, through trajectory modification and sample augmentation methods, can specifically enhance the risk level of lane-changing behavior while maintaining the physical rationality of the trajectory, significantly increasing the number of high-risk lane-changing samples. This directly improves the model's learning ability for rare or extreme lane-changing behaviors, enabling it to generate high-risk behaviors that are closer to real and complex traffic scenarios in simulations. It also makes up for the problem of insufficient model generalization caused by the scarcity of high-risk behavior samples in existing technologies.

[0036] Step S20: Construct an initial lane-changing model using a long short-term memory encoder, an interactive feature model, and a long short-term memory decoder. Perform open-loop training on the initial lane-changing model based on all the initial lane-changing samples to obtain a pre-trained model.

[0037] In the embodiments disclosed in this invention, a neural network lane-changing behavior modeling method based on the encoder-decoder structure of Long Short-Term Memory (LSTM) combined with the attention mechanism is used. The model is pre-trained using initial lane-changing samples in a teacher-guided manner, so that the model can fully learn the motion and interaction features reflected in real sample data.

[0038] Specifically, a long short-term memory encoder, an interaction feature model, and a long short-term memory decoder are constructed respectively, and an initial lane-changing model is constructed based on the long short-term memory encoder, the interaction feature model, and the long short-term memory decoder; The target lane change sample set is input into the initial lane change model for initial prediction, and the corresponding prediction result is output. A loss function is constructed using the target lane-changing sample set and the corresponding prediction results. The initial lane-changing model is then trained using the loss function to obtain a pre-trained model.

[0039] Among them, the vehicle's lane-changing trajectory is the external manifestation of the change in motion inertia due to the driver's intention and the influence of surrounding vehicles; the long short-term memory encoder can transform the vehicle's motion state over a period of time into a hidden state vector of fixed length, thereby effectively summarizing the historical information of the vehicle's motion and preserving key temporal features.

[0040] Furthermore, in real-world traffic, interactions are complex and change over time. This invention employs an interaction feature model to compress dynamically changing interaction information into a more compact embedding space, making conditional prediction easier for the decoder. During operation, both the generation of a vehicle's lane-changing intention and the execution of its lane-changing behavior are influenced by the movement of surrounding vehicles, but the scope and degree of interaction are unclear. Therefore, this model uses an attention mechanism to learn importance weights when processing the interaction between lane-changing vehicles and surrounding vehicles. Finally, the lane-changing vehicle's encoding vector and the interaction feature vector are concatenated as input to the decoder, which gradually generates a sequence of coordinates for future trajectory points.

[0041] Furthermore, a long short-term memory encoder, an interaction feature model, and a long short-term memory decoder are constructed respectively; Identify the lane-changing vehicle and multiple surrounding vehicles of the lane-changing vehicle, and construct historical temporal features of the lane-changing vehicle and all surrounding vehicles based on the long short-time memory encoder; Based on the interaction feature model, construct the interaction relationship information between the lane-changing vehicle and all the surrounding vehicles; Based on the historical time series features and the interaction relationship information, the predicted coordinate sequence of the lane-changing vehicle is constructed according to the long short-term memory decoder; Based on the predicted coordinate sequence, an initial lane-changing model is constructed between the lane-changing vehicle and all surrounding vehicles: ; ; ; in, Indicates the predicted trajectory. express and The nonlinear relationship between them This indicates the historical trajectory of vehicles changing lanes. Indicating spatial relationships, Indicates the step size of the historical trajectory. Indicates the step size in traffic simulation. Indicates a time step. Indicates vehicles changing lanes at Coordinates at time, Indicates vehicles changing lanes at Coordinates at time, Indicates vehicles changing lanes at Coordinates at time, Indicates surrounding vehicles The historical trajectory of moments, Indicates surrounding vehicles The historical trajectory of moments, Indicates surrounding vehicles The historical trajectory of a moment.

[0042] Among them, such as Figure 3 As shown, in the embodiments disclosed in this invention, the lane-changing model is pre-trained using initial lane-changing samples. The purpose is to enable the model to fully learn the vehicle motion patterns and interaction features between vehicles contained in real data. During the pre-training process, an open-loop batch training method guided by a teacher is adopted, with the root mean square error as the loss function to ensure effective convergence of the model in terms of trajectory prediction accuracy.

[0043] Step S30: Use the target lane-changing sample set to perform closed-loop training on the pre-trained model to obtain the target lane-changing model, and use the target lane-changing model to simulate the lane-changing vehicle under test to obtain the simulated lane-changing trajectory of the vehicle under test.

[0044] Because of the open-loop training method, the pre-trained model exhibits exposure bias when actually generating lane-changing trajectories. Specifically, the input during training is the actual historical trajectory, while the input during inference is its own predicted trajectory, leading to an inconsistency between the distribution during testing and training. This results in error accumulation and weakens the model's ability to stably reproduce lane-changing behavior. Therefore, in the embodiments disclosed in this invention, closed-loop transfer learning is employed to further train the pre-trained model in a closed-loop manner, enhancing its ability to learn high-risk lane-changing behavior features, thereby constructing a closed-loop transfer learning (CLTL) system capable of generating high-risk lane-changing behavior.

[0045] Specifically, the target lane change sample set is input into the pre-trained model, wherein the number of samples in the target lane change sample set is greater than the number of samples that cause the pre-trained model to converge. The pre-trained model predicts each target lane change sample and outputs the simulated displacement of the vehicle under test under each target lane change sample. For each simulated displacement, the vehicle under test is controlled to perform a lane-changing simulation based on the simulated displacement to obtain the lane-changing position of the vehicle under test; Determine whether the simulation time of the lane change simulation of the vehicle under test is equal to the actual lane change time. If not, replace the simulation time with the actual lane change time, calculate the mean square error between the lane change simulation trajectory and the actual lane change trajectory, update the neural network parameters of the pre-trained model according to the mean square error, and obtain the target lane change model after the pre-trained model converges.

[0046] Among them, such as Figure 4As shown, in order to eliminate the exposure bias caused by open-loop training, a closed-loop mechanism is adopted in the training phase. That is, the model output is used as the input for the next time step at each time step, so that the distribution of training data is consistent with the inference process, thereby effectively reducing error accumulation.

[0047] Furthermore, since closed-loop training requires trajectory updates and historical information extraction at each time step, batch computation is difficult, resulting in low training efficiency. To accelerate training, this method introduces a transfer learning optimization strategy. By using a pre-trained model as the base model, closed-loop training is employed to fine-tune the parameters of the pre-trained model, thus addressing the issues of low efficiency and slow convergence in closed-loop training. In the embodiments disclosed in this invention, the source task can be considered as the task achievable by the pre-trained model, i.e., outputting the future trajectory of the lane-changing vehicle; while the target task is the simulation of lane-changing behavior. Based on holistic transfer learning, the parameters of the pre-trained model are used as initial values ​​for closed-loop training, and all parameters in the model are fine-tuned through closed-loop training.

[0048] To address the error accumulation problem caused by biases exposed during open-loop training, this invention introduces closed-loop training during the model optimization phase. This ensures that the input during training remains consistent with that during inference, effectively reducing distribution differences and minimizing the generation of abnormal trajectories. Through a transfer learning strategy, the parameters of the pre-trained model are used as initial values ​​for closed-loop training, enabling overall fine-tuning of all parameters. This shortens the model convergence time and improves the training efficiency of models for high-risk lane-changing behaviors.

[0049] Finally, the trained model is used to simulate lane-changing for the vehicle under test, outputting the simulated lane-changing trajectory (i.e., high-risk lane-changing behavior). The high-risk lane-changing behavior trajectory generated by this invention not only conforms to vehicle dynamics constraints but also possesses sufficient behavioral diversity. It can be directly used for extreme condition testing of autonomous driving systems and the construction of dangerous scenarios in virtual traffic simulation platforms, effectively improving the realism and applicability of the lane-changing behavior model in complex traffic environments.

[0050] Furthermore, to demonstrate the necessity and effectiveness of the closed-loop transfer learning method disclosed in this invention, neural network models were constructed using Scaled Dot-product Attention (SDPA) and Additive Attention (+A), respectively. The accuracy of models SDPA-LSTM-CL (CL, Closed-loop) and +A-LSTM-CL obtained solely through closed-loop learning, and models SDPA-LSTM-CLTL and +A-LSTM-CLTL obtained through closed-loop transfer learning, in generating lane-changing behavior, were compared. Closed-loop transfer learning uses the parameters of the pre-trained models as initial values ​​to accelerate model convergence and improve generalization ability.

[0051] To evaluate the error between the model and the actual lane-changing trajectory, the Simulation Average Displacement Error (SADE), Simulation Final Displacement Error (SFDE), and the corresponding lateral indices SLADE (a lateral index of SADE) and SLFDE (a lateral index of SFDE) were used to evaluate the model. The specific results are shown in Tables 1, 2, and 3. Table 1: Accuracy of Pre-trained Models for Lane Changing Behavior Simulation

[0052] Table 2: Accuracy of models trained using closed-loop method for lane-changing behavior simulation

[0053] Table 3: Accuracy of models using closed-loop transfer learning for lane-changing behavior simulation

[0054] As shown in Table 1, there is a significant error between the simulated trajectories generated by the pre-trained models and the actual trajectories. Specifically, almost all pre-trained models have a SADE greater than 20 meters and an SFDE exceeding 60 meters, which clearly fails to meet the requirements of traffic simulation and may even lead to the generation of abnormal trajectories or unreasonable traffic behaviors. Further calculation and analysis of lateral motion performance indicators revealed that almost all pre-trained models have an SLADE of approximately 2 meters. Considering that the standard lane width is approximately 3.6 meters, an SLADE of 2 meters indicates a significant difference between the lateral motion trend of the simulated trajectory and the actual trajectory. In summary, the large longitudinal and lateral errors and high collision rates highlight the inadequacy of the pre-trained models in simulating lane-changing behavior, failing to reasonably and accurately reflect the characteristics of real lane-changing behavior.

[0055] Comparing Tables 1 and 2 reveals that after adopting closed-loop learning, the SADE and SFDE metrics significantly decreased compared to the pre-trained model, with the SADE of all models using closed-loop learning dropping below 10 meters. However, the lateral motion accuracy metrics SLADE and SLFDE did not decrease significantly; in fact, the model + A-LSTM-CL even showed a slight increase in the SLADE metric. This indicates that using closed-loop learning alone has limited effectiveness in reducing lateral motion errors.

[0056] Comparing Tables 1, 2, and 3, the longitudinal and lateral errors of all models decreased after adopting closed-loop transfer learning. Taking the model +A-LSTM-CLTL, which showed the most significant improvement, as an example, its SADE dropped to below 3 meters, a decrease of 95.30% compared to +A-LSTM and 48.54% compared to +A-LSTM-CL. Regarding lateral error, the SLADE of +A-LSTM-CLTL decreased to 0.58 meters, a decrease of 73.27% compared to +A-LSTM and 77.86% compared to +A-LSTM-CL.

[0057] Furthermore, such as Figure 5 As shown, the simulation trajectories of the pre-trained model, the model using closed-loop learning, and the model using closed-loop transfer learning are illustrated. Figure 5 (a) shows the simulated trajectory of a typical lane change. Figure 5 (b) shows the simulated trajectory of continuous lane changes. Figure 5 (c) shows the simulated trajectory of consecutive reverse lane changes. Figure 5 Figure (d) shows the simulated trajectory without lane changing. A comparison reveals significant deviations between the trajectories generated by the pre-trained model and the model using only closed-loop learning and the actual trajectories. The endpoint of the simulated trajectory generated by the pre-trained model differs considerably from the endpoint of the actual trajectory; while the model using only closed-loop learning approximates the actual trajectory in longitudinal displacement, it clearly shows the vehicle traveling along a straight path, which is almost impossible in real traffic.

[0058] Furthermore, in another embodiment of the present invention, the SDPA-LSTM-CLTL model trained using general lane-changing behavior sample data and the SDPA-LSTM-CLTLH model trained using high-risk lane-changing behavior sample dataset are used to simulate lane-changing behavior and generate lane-changing trajectories. Risk indicators are used for risk assessment. Specifically, the risk indicators are: (First indicator) the minimum time difference between the lane-changing vehicle and the following vehicle in the target lane to the conflict point; (Second indicator) the lateral distance between the lane-changing vehicle and the following vehicle in the target lane when the time difference between them reaches the conflict point is minimized. The specific results are shown in Table 4. Table 4: Risk Quantitative Assessment Table for Lane Changing Model

[0059] As shown in Table 4, the mean values ​​of the SDPA-LSTM-CLTLH model on both the first and second risk indicators are lower than those of SDPA-LSTM-CLTL and also lower than those of the actual lane-changing behavior samples. This indicates that training the SDPA-LSTM-CLTLH model on a high-risk lane-changing behavior sample dataset makes it perform better in simulating and generating high-risk lane-changing behaviors. These experimental results also indirectly demonstrate that closed-loop transfer learning can accurately capture features from high-risk samples, further validating the effectiveness of the closed-loop transfer learning method.

[0060] This invention simultaneously performs time-series encoding on the historical trajectories of lane-changing vehicles and multiple surrounding vehicles, introduces an attention mechanism to adaptively allocate interaction weights, captures the interaction relationships between vehicles that change over time, improves the model's accuracy in capturing complex traffic dynamics, and makes the output results more accurate.

[0061] Furthermore, such as Figure 6 As shown, based on the above-mentioned lane-change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics, the present invention also provides a lane-change trajectory generation system based on trajectory start and end point adjustment and vehicle dynamics, wherein the lane-change trajectory generation system based on trajectory start and end point adjustment and vehicle dynamics includes: The sample construction module 51 is used to acquire multiple initial lane change samples, perform risk assessment on each initial lane change sample, and perform risk enhancement processing on the trajectory start point and / or trajectory end point of each low-risk lane change sample, so as to finally obtain a target lane change sample set that meets the requirements. The model pre-training module 52 is used to construct an initial lane-changing model using a long short-term memory encoder, an interactive feature model and a long short-term memory decoder, and to perform open-loop training on the initial lane-changing model based on all the initial lane-changing samples to obtain a pre-trained model. The training and simulation module 53 is used to input the target lane-changing sample set into the pre-trained model for closed-loop training, use the trained model to simulate lane-changing of the vehicle under test, and output the simulated lane-changing trajectory of the vehicle under test.

[0062] Furthermore, such as Figure 7 As shown, based on the above-mentioned method and system for generating lane change trajectories based on trajectory start and end point adjustment and vehicle dynamics, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 7 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0063] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 may also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a lane-change trajectory generation program 40 based on trajectory start and end point adjustment and vehicle dynamics. This lane-change trajectory generation program 40 can be executed by the processor 10 to implement the lane-change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics in this application.

[0064] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the lane change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics.

[0065] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components of the terminal communicate with each other via a system bus.

[0066] In one embodiment, when the processor 10 executes the lane change trajectory generation program 40 based on trajectory start and end point adjustment and vehicle dynamics in the memory 20, the following steps are performed: Multiple initial lane change samples are obtained, a risk assessment is performed on each initial lane change sample, and risk enhancement processing is applied to the trajectory start point and / or trajectory end point of each low-risk lane change sample to finally obtain a target lane change sample set that meets the requirements. An initial lane-changing model is constructed using a long short-term memory encoder, an interactive feature model, and a long short-term memory decoder. The initial lane-changing model is then trained in an open loop based on all the initial lane-changing samples to obtain a pre-trained model. The target lane-changing sample set is input into the pre-trained model for closed-loop training. The trained model is used to simulate lane-changing for the vehicle under test and outputs the simulated lane-changing trajectory of the vehicle under test.

[0067] Specifically, obtaining multiple initial lane-changing samples includes: Construct a vehicle lane-changing trajectory generation model and obtain the total lateral displacement, lane-changing speed, and lane-changing time of the lane-changing vehicle; The total lateral displacement, the lane-changing speed, and the lane-changing time are input into the vehicle lane-changing trajectory generation model for simulation, and all initial lane-changing samples of the lane-changing vehicle during the test period are output. Assess the risk of each initial lane change sample, and define all initial lane change samples whose risk assessment value is higher than a preset risk threshold as target lane change samples; For all initial lane change samples whose risk assessment value is not higher than the preset risk threshold, risk enhancement processing is performed on all initial lane change samples according to the trajectory starting point of each initial lane change sample to obtain target lane change samples that meet the requirements. All the target lane change samples are encapsulated to obtain the target lane change sample set.

[0068] The initial lane-changing sample includes lateral displacement and longitudinal displacement. The lateral displacement and the longitudinal displacement are expressed as follows:

[0069] in, Indicates the current time t Lateral displacement, Indicates the current time t longitudinal displacement, Indicates the total lateral displacement. Indicates lane change time. A coefficient representing lane-changing time. This represents the lane-changing characteristic coefficient. Indicates the lane-changing speed.

[0070] Specifically, for all initial lane-changing samples whose risk assessment value is not higher than the preset risk threshold, risk enhancement processing is performed on all initial lane-changing samples based on the trajectory starting point of each initial lane-changing sample to obtain target lane-changing samples that meet the requirements. This process includes: Obtain the lane change start point, lane change end point, and longitudinal speed of all the initial lane change samples. Randomly select multiple initial lane change samples, pre-set the lane change start point of each initial lane change sample by a distance, and keep the lane change end point and longitudinal speed unchanged to improve the risk level of each initial lane change sample and obtain the corresponding target lane change sample. Randomly select multiple initial lane change samples, pre-set the lane change termination point of each initial lane change sample by a distance, and keep the lane change start point and the longitudinal speed unchanged, so as to increase the risk threshold of each initial lane change sample and obtain the corresponding target lane change sample.

[0071] Specifically, the process of constructing an initial lane-changing model using a long short-term memory encoder, an interaction feature model, and a long short-term memory decoder, and then performing open-loop training on the initial lane-changing model based on all the initial lane-changing samples to obtain a pre-trained model, includes: A long short-term memory encoder, an interaction feature model, and a long short-term memory decoder are constructed respectively, and an initial lane-changing model is constructed based on the long short-term memory encoder, the interaction feature model, and the long short-term memory decoder; The target lane change sample set is input into the initial lane change model for initial prediction, and the corresponding prediction result is output. A loss function is constructed using the target lane-changing sample set and the corresponding prediction results. The initial lane-changing model is then trained using the loss function to obtain a pre-trained model.

[0072] The construction of a Long Short-Term Memory (LSTM) encoder, an interaction feature model, and a LSM decoder, and the construction of an initial lane-changing model based on the LSM encoder, the interaction feature model, and the LSM decoder, specifically includes: A long short-term memory encoder, an interaction feature model, and a long short-term memory decoder are constructed respectively. Identify the lane-changing vehicle and multiple surrounding vehicles of the lane-changing vehicle, and construct historical temporal features of the lane-changing vehicle and all surrounding vehicles based on the long short-time memory encoder; Based on the interaction feature model, construct the interaction relationship information between the lane-changing vehicle and all the surrounding vehicles; Based on the historical time series features and the interaction relationship information, the predicted coordinate sequence of the lane-changing vehicle is constructed according to the long short-term memory decoder; Based on the predicted coordinate sequence, an initial lane-changing model is constructed between the lane-changing vehicle and all surrounding vehicles: ; ; ; in, Indicates the predicted trajectory. express and The nonlinear relationship between them This indicates the historical trajectory of vehicles changing lanes. Indicates spatial relationships, Indicates the step size of the historical trajectory. Indicates the step size in traffic simulation. Indicates a time step. Indicates vehicles changing lanes at Coordinates at time, Indicates vehicles changing lanes at Coordinate position at time, Indicates vehicles changing lanes at Coordinate position at time, Indicates surrounding vehicles The historical trajectory of moments, Indicates surrounding vehicles The historical trajectory of moments, Indicates surrounding vehicles The historical trajectory of a moment.

[0073] Specifically, the step of using the target lane-changing sample set to perform closed-loop training on the pre-trained model to obtain the target lane-changing model includes: The target lane change sample set is input into the pre-trained model, wherein the number of samples in the target lane change sample set is greater than the number of samples that cause the pre-trained model to converge. The pre-trained model predicts each target lane change sample and outputs the simulated displacement of the vehicle under test under each target lane change sample. For each simulated displacement, the vehicle under test is controlled to perform a lane-changing simulation based on the simulated displacement to obtain the lane-changing position of the vehicle under test; Determine whether the simulation time of the lane change simulation of the vehicle under test is equal to the actual lane change time. If not, replace the simulation time with the actual lane change time, calculate the mean square error between the lane change simulation trajectory and the actual lane change trajectory, update the neural network parameters of the pre-trained model according to the mean square error, and obtain the target lane change model after the pre-trained model converges.

[0074] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a lane-change trajectory generation program based on trajectory start and end point adjustment and vehicle dynamics, and the lane-change trajectory generation program based on trajectory start and end point adjustment and vehicle dynamics, when executed by a processor, implements the steps of the lane-change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics as described above.

[0075] In summary, this invention provides a method and related equipment for generating lane-changing trajectories based on trajectory start and end point adjustment and vehicle dynamics. The method includes: acquiring multiple initial lane-changing samples; performing a risk assessment on each initial lane-changing sample; and performing risk enhancement processing on the trajectory start and / or trajectory end points of each low-risk lane-changing sample to finally obtain a target lane-changing sample set that meets the requirements; constructing an initial lane-changing model using a long short-term memory encoder, an interaction feature model, and a long short-term memory decoder; performing open-loop training on the initial lane-changing model based on all the initial lane-changing samples to obtain a pre-trained model; inputting the target lane-changing sample set into the pre-trained model for closed-loop training; using the trained model to simulate lane-changing for the vehicle under test; and outputting the simulated lane-changing trajectory of the vehicle under test. This invention also performs temporal encoding on the historical trajectories of the lane-changing vehicle and its surrounding multiple vehicles, introduces an attention mechanism to adaptively allocate interaction weights, captures the time-varying interaction relationships between vehicles, improves the model's accuracy in capturing complex traffic dynamic features, and makes the output results more accurate.

[0076] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0077] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0078] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for generating lane-changing trajectories based on trajectory start and end point adjustment and vehicle dynamics, characterized in that, The lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics includes: Multiple initial lane change samples are obtained, a risk assessment is performed on each initial lane change sample, and risk enhancement processing is applied to the trajectory start point and / or trajectory end point of each low-risk lane change sample to finally obtain a target lane change sample set that meets the requirements. An initial lane-changing model is constructed using a long short-term memory encoder, an interactive feature model, and a long short-term memory decoder. The initial lane-changing model is then trained in an open loop based on all the initial lane-changing samples to obtain a pre-trained model. The pre-trained model is trained in a closed loop using the target lane-changing sample set to obtain the target lane-changing model. The target lane-changing model is then used to simulate the lane-changing of the vehicle under test to obtain the simulated lane-changing trajectory of the vehicle under test.

2. The lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics according to claim 1, characterized in that, The acquisition of multiple initial lane-changing samples specifically includes: Construct a vehicle lane-changing trajectory generation model and obtain the total lateral displacement, lane-changing speed, and lane-changing time of the lane-changing vehicle; The total lateral displacement, the lane-changing speed, and the lane-changing time are input into the vehicle lane-changing trajectory generation model for simulation, and all initial lane-changing samples of the lane-changing vehicle during the test period are output. Assess the risk of each initial lane change sample, and define all initial lane change samples whose risk assessment value is higher than a preset risk threshold as target lane change samples; For all initial lane change samples whose risk assessment value is not higher than the preset risk threshold, risk enhancement processing is performed on all initial lane change samples according to the trajectory starting point of each initial lane change sample to obtain target lane change samples that meet the requirements. All the target lane change samples are encapsulated to obtain the target lane change sample set.

3. The lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics according to claim 2, characterized in that, The initial lane-changing sample includes: lateral displacement and longitudinal displacement; The lateral displacement and the longitudinal displacement are expressed as follows: in, Indicates the current time t Lateral displacement, Indicates the current time t longitudinal displacement, Indicates the total lateral displacement. Indicates lane change time. A coefficient representing lane-changing time. This represents the lane-changing characteristic coefficient. Indicates the lane-changing speed.

4. The lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics according to claim 3, characterized in that, For all initial lane-changing samples whose risk assessment value is not higher than the preset risk threshold, risk enhancement processing is performed on all initial lane-changing samples based on the trajectory starting point of each initial lane-changing sample to obtain target lane-changing samples that meet the requirements. Specifically, this includes: Obtain the lane change start point, lane change end point, and longitudinal speed of all the initial lane change samples. Randomly select multiple initial lane change samples, pre-set the lane change start point of each initial lane change sample by a distance, and keep the lane change end point and longitudinal speed unchanged to improve the risk level of each initial lane change sample and obtain the corresponding target lane change sample. Randomly select multiple initial lane change samples, pre-set the lane change termination point of each initial lane change sample by a distance, and keep the lane change start point and the longitudinal speed unchanged, so as to increase the risk threshold of each initial lane change sample and obtain the corresponding target lane change sample.

5. The lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics according to claim 1, characterized in that, The process involves constructing an initial lane-changing model using a long short-term memory encoder, an interaction feature model, and a long short-term memory decoder. Then, the initial lane-changing model is trained in an open-loop manner using all the initial lane-changing samples to obtain a pre-trained model. This process specifically includes: A long short-term memory encoder, an interaction feature model, and a long short-term memory decoder are constructed respectively, and an initial lane-changing model is constructed based on the long short-term memory encoder, the interaction feature model, and the long short-term memory decoder; The target lane change sample set is input into the initial lane change model for initial prediction, and the corresponding prediction result is output. A loss function is constructed using the target lane-changing sample set and the corresponding prediction results. The initial lane-changing model is then trained using the loss function to obtain a pre-trained model.

6. The lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics according to claim 5, characterized in that, The step of constructing a Long Short-Term Memory (LSTM) encoder, an interaction feature model, and a LSM decoder, and then constructing an initial lane-changing model based on the LSM encoder, the interaction feature model, and the LSM decoder, specifically includes: A long short-term memory encoder, an interaction feature model, and a long short-term memory decoder are constructed respectively. Identify the lane-changing vehicle and multiple surrounding vehicles of the lane-changing vehicle, and construct historical temporal features of the lane-changing vehicle and all surrounding vehicles based on the long short-time memory encoder; Based on the interaction feature model, construct the interaction relationship information between the lane-changing vehicle and all the surrounding vehicles; Based on the historical time series features and the interaction relationship information, the predicted coordinate sequence of the lane-changing vehicle is constructed according to the long short-term memory decoder; Based on the predicted coordinate sequence, an initial lane-changing model is constructed between the lane-changing vehicle and all surrounding vehicles: ； ； ； in, Indicates the predicted trajectory. express and The nonlinear relationship between them This indicates the historical trajectory of vehicles changing lanes. Indicating spatial relationships, Indicates the step size of the historical trajectory. Indicates the step size in traffic simulation. Indicates a time step. Indicates vehicles changing lanes at Coordinate position at time, Indicates vehicles changing lanes at Coordinate position at time, Indicates vehicles changing lanes at Coordinate position at time, Indicates surrounding vehicles The historical trajectory of moments, Indicates surrounding vehicles The historical trajectory of moments, Indicates surrounding vehicles The historical trajectory of a moment.

7. The lane-changing trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics according to claim 1, characterized in that, The step of using the target lane-changing sample set to perform closed-loop training on the pre-trained model to obtain the target lane-changing model specifically includes: The target lane change sample set is input into the pre-trained model, wherein the number of samples in the target lane change sample set is greater than the number of samples that cause the pre-trained model to converge. The pre-trained model predicts each target lane change sample and outputs the simulated displacement of the vehicle under test under each target lane change sample. For each simulated displacement, the vehicle under test is controlled to perform a lane-changing simulation based on the simulated displacement to obtain the lane-changing position of the vehicle under test; Determine whether the simulation duration of the lane change simulation of the vehicle under test is equal to the actual lane change duration. If not, replace the simulation duration with the actual lane change duration, calculate the mean square error between the lane change simulation trajectory and the actual lane change trajectory, update the neural network parameters of the pre-trained model based on the mean square error, and obtain the target lane change model after the pre-trained model converges.

8. A lane-changing trajectory generation system based on trajectory start and end point adjustment and vehicle dynamics, characterized in that, The lane-change trajectory generation system based on trajectory start and end point adjustment and vehicle dynamics is used to implement the lane-change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics as described in any one of claims 1-7, wherein the lane-change trajectory generation system based on trajectory start and end point adjustment and vehicle dynamics includes: The sample construction module is used to acquire multiple initial lane change samples, perform risk assessment on each initial lane change sample, and perform risk enhancement processing on the trajectory start point and / or trajectory end point of each low-risk lane change sample, so as to finally obtain a target lane change sample set that meets the requirements. The model pre-training module is used to construct an initial lane-changing model using a long short-term memory encoder, an interactive feature model, and a long short-term memory decoder, and to perform open-loop training on the initial lane-changing model based on all the initial lane-changing samples to obtain a pre-trained model. The training and simulation module is used to input the target lane-changing sample set into the pre-trained model for closed-loop training, use the trained model to simulate lane-changing for the vehicle under test, and output the simulated lane-changing trajectory of the vehicle under test.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a lane change trajectory generation program based on trajectory start and end point adjustment and vehicle dynamics, stored in the memory and executable on the processor. When the lane change trajectory generation program based on trajectory start and end point adjustment and vehicle dynamics is executed by the processor, it implements the steps of the lane change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a lane-change trajectory generation program based on trajectory start and end point adjustment and vehicle dynamics. When the lane-change trajectory generation program based on trajectory start and end point adjustment and vehicle dynamics is executed by a processor, it implements the steps of the lane-change trajectory generation method based on trajectory start and end point adjustment and vehicle dynamics as described in any one of claims 1-7.

Images