Control method and model training method for intelligent driving vehicle, and vehicle

Abstract

Embodiments of the present application provide a kind of intelligent driving vehicle control method, model training method, device and vehicle.The method comprises: obtaining vehicle data and road constraint data of vehicle driving;Vehicle data and road constraint data are carried out data fusion processing, and fusion feature is obtained;Fusion feature is input into initial intelligent model, and fusion feature is processed based on the clustering module and the counter-imitation module of initial intelligent model, and the parameter of initial intelligent model is optimized by loss function, and driving style recognition model is obtained;Wherein, driving style recognition model is used to process vehicle data and road constraint data of vehicle driving, and driving style is obtained;Driving style is used to generate vehicle driving instruction;Vehicle driving instruction is used to control vehicle driving.The method is used to accurately identify driving style and generate more personalized and reliable vehicle driving instruction, to realize the effective control to vehicle driving.

Description

Technical Field

[0001] This application relates to the field of intelligent driving technology, and in particular to a control method, model training method, device and vehicle for intelligent driving vehicles. Background Technology

[0002] With the rapid development of intelligent driving technology, users' demand for personalized driving experiences is growing.

[0003] In existing technologies, the control method for intelligent driving vehicles involves collecting driver operation data, using machine learning algorithms to classify driving styles, and adjusting the vehicle's driving parameters based on the classification results.

[0004] However, the models used in the above methods cannot dynamically adapt to the long-term evolution of driving styles, resulting in a lag between the model output and the actual driving intention, which in turn affects the adaptability of the intelligent driving system to complex scenarios and the user experience. Summary of the Invention

[0005] This application provides a control method, model training method, device, and vehicle for intelligent driving vehicles, which can accurately identify driving styles and generate more personalized and reliable vehicle driving commands to achieve effective control of vehicle driving.

[0006] In a first aspect, embodiments of this application provide a method for training a driving style recognition model, comprising:

[0007] Acquire vehicle data and road constraint data; among which, vehicle data represents multimodal time-series data of vehicle motion state and driving operation; road constraint data represents the constraint information of road geometry and traffic rules on driving.

[0008] The vehicle data and road constraint data are fused to obtain fused features;

[0009] The fused features are input into the initial intelligent model. The clustering module and adversarial imitation module of the initial intelligent model process the fused features, and the parameters of the initial intelligent model are optimized through a loss function to obtain the driving style recognition model. The driving style recognition model is used to process vehicle data and road constraint data to obtain the driving style. The driving style is used to generate vehicle driving commands. The vehicle driving commands are used to control the vehicle's movement.

[0010] In one possible embodiment, vehicle data and road constraint data are fused to obtain fused features, including:

[0011] The road constraint data is encoded to obtain constraint-coded data;

[0012] The vehicle data is processed by statistical feature processing, sequence feature processing, and derived feature processing to obtain vehicle feature data.

[0013] Cross-modal fusion processing is performed on the constraint-coded data and vehicle feature data to obtain fused features.

[0014] In one possible embodiment, the fused features are input into an initial intelligent model. The clustering module and adversarial mimicry module of the initial intelligent model process the fused features, and the parameters of the initial intelligent model are optimized using a loss function to obtain a driving style recognition model, including:

[0015] The fused features are input into the initial intelligent model, and the fused features are processed based on the initial intelligent model to obtain features and discriminants; where the features represent the intermediate layer representation extracted by the internal modules of the model; the discriminants represent the probability value of the fused features belonging to the real samples;

[0016] Based on features and discriminants, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain a driving style recognition model.

[0017] In one possible embodiment, the fused features are input into an initial intelligent model, and the fused features are processed based on the initial intelligent model to obtain features and discriminants, including:

[0018] The fused features are input into the initial intelligent model, and the encoder based on the initial intelligent model processes the fused features to obtain temporal encoded features;

[0019] The clustering module based on the initial intelligent model performs clustering processing on the temporal coding features to obtain temporal coding features with clustering labels.

[0020] The adversarial imitation module based on the initial intelligent model processes the temporal coding features and the temporal coding features with clustering labels to obtain features and discriminants.

[0021] In one possible embodiment, based on features and discriminants, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain a driving style recognition model, including:

[0022] Based on features, discriminants, and loss functions, the loss value of the initial intelligent model is determined;

[0023] Based on the loss value, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain the driving style recognition model.

[0024] In one possible embodiment, before inputting the fused features into an initial intelligent model, processing the fused features based on the clustering module and adversarial mimicry module of the initial intelligent model, and optimizing the parameters of the initial intelligent model using a loss function to obtain the driving style recognition model, the method further includes:

[0025] The fused features are processed by feature extraction to obtain vertical features, horizontal features, and macro features. Among them, the vertical features represent the feature components of the vehicle's motion state and control behavior along the main driving direction; the horizontal features represent the feature components of the vehicle's motion state and control behavior perpendicular to the driving direction; and the macro features represent the features of high-level semantics within a preset time period.

[0026] Secondly, embodiments of this application provide a control method for an intelligent driving vehicle, including:

[0027] Acquire vehicle data and road constraint data;

[0028] Vehicle data and road constraint data are input into the driving style recognition model. The driving style recognition model is then used to process the vehicle data and road constraint data to obtain the driving style output by the driving style recognition model. The driving style recognition model is trained using the training method provided above.

[0029] Vehicle data, road constraint data, and driving style are input into the instruction generation model. The instruction generation model processes the vehicle data, road constraint data, and driving style to obtain the vehicle driving instructions output by the instruction generation model.

[0030] Control the vehicle's movement based on driving commands.

[0031] In one possible embodiment, vehicle data, road constraint data, and driving style are input into an instruction generation model. The instruction generation model processes the vehicle data, road constraint data, and driving style to obtain vehicle driving instructions output by the instruction generation model, including:

[0032] Vehicle data, road constraint data, and driving style are input into the large language model of the instruction generation model for processing to obtain macro instructions; among them, macro instructions represent the high-level instructions that the vehicle should follow in terms of driving intentions and behavioral decisions within a future preset time window.

[0033] Macro-level instructions, vehicle data, road constraint data, and driving style are input into the policy gradient module of the instruction generation model for processing to obtain vehicle driving instructions.

[0034] Thirdly, embodiments of this application provide a training device for a driving style recognition model, comprising:

[0035] The acquisition module is used to acquire vehicle data and road constraint data of vehicle movement; among which, vehicle data represents multimodal time-series data of vehicle motion state and driving operation; road constraint data represents the constraint information of road geometry and traffic rules on driving.

[0036] The processing module is used to perform data fusion processing on vehicle data and road constraint data to obtain fused features;

[0037] The training module is used to input the fused features into the initial intelligent model. The clustering module and the adversarial imitation module based on the initial intelligent model process the fused features and optimize the parameters of the initial intelligent model through a loss function to obtain the driving style recognition model. The driving style recognition model is used to process vehicle data and road constraint data to obtain the driving style. The driving style is used to generate vehicle driving commands. The vehicle driving commands are used to control the vehicle's movement.

[0038] In one possible embodiment, the processing module is configured to:

[0039] The road constraint data is encoded to obtain constraint-coded data;

[0040] The vehicle data is processed by statistical feature processing, sequence feature processing, and derived feature processing to obtain vehicle feature data.

[0041] Cross-modal fusion processing is performed on the constraint-coded data and vehicle feature data to obtain fused features.

[0042] In one possible embodiment, the training module is used for:

[0043] The fused features are input into the initial intelligent model, and the fused features are processed based on the initial intelligent model to obtain features and discriminants; where the features represent the intermediate layer representation extracted by the internal modules of the model; the discriminants represent the probability value of the fused features belonging to the real samples;

[0044] Based on features and discriminants, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain a driving style recognition model.

[0045] In one possible embodiment, the fused features are input into an initial intelligent model, and the fused features are processed based on the initial intelligent model to obtain features and discriminants, including:

[0046] The fused features are input into the initial intelligent model, and the encoder based on the initial intelligent model processes the fused features to obtain temporal encoded features;

[0047] The clustering module based on the initial intelligent model performs clustering processing on the temporal coding features to obtain temporal coding features with clustering labels.

[0048] The adversarial imitation module based on the initial intelligent model processes the temporal coding features and the temporal coding features with clustering labels to obtain features and discriminants.

[0049] In one possible embodiment, based on features and discriminants, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain a driving style recognition model, including:

[0050] Based on features, discriminants, and loss functions, the loss value of the initial intelligent model is determined;

[0051] Based on the loss value, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain the driving style recognition model.

[0052] In one possible embodiment, prior to the training module, the following is also included:

[0053] The extraction module is used to perform feature extraction processing on the fused features to obtain vertical features, horizontal features, and macro features. Among them, the vertical features represent the feature components of the vehicle's motion state and control behavior along the main driving direction; the horizontal features represent the feature components of the vehicle's motion state and control behavior perpendicular to the driving direction; and the macro features represent the features of high-level semantics within a preset time period.

[0054] Fourthly, embodiments of this application provide a control device for an intelligent driving vehicle, comprising:

[0055] The acquisition module is used to acquire vehicle data and road constraint data.

[0056] The recognition module is used to input vehicle data and road constraint data into the driving style recognition model, process the vehicle data and road constraint data based on the driving style recognition model, and obtain the driving style output by the driving style recognition model; wherein, the driving style recognition model is trained by the driving style recognition model training device provided above.

[0057] The generation module is used to input vehicle data, road constraint data, and driving style into the instruction generation model. Based on the instruction generation model, the vehicle data, road constraint data, and driving style are processed to obtain the vehicle driving instructions output by the instruction generation model.

[0058] The control module is used to control the vehicle's movement based on driving commands.

[0059] In one possible embodiment, the generation module is used for:

[0060] Vehicle data, road constraint data, and driving style are input into the large language model of the instruction generation model for processing to obtain macro instructions; among them, macro instructions represent the high-level instructions that the vehicle should follow in terms of driving intentions and behavioral decisions within a future preset time window.

[0061] Macro-level instructions, vehicle data, road constraint data, and driving style are input into the policy gradient module of the instruction generation model for processing to obtain vehicle driving instructions.

[0062] Fifthly, embodiments of this application provide a vehicle, including: a memory and a processor;

[0063] The memory stores the instructions that the computer executes;

[0064] The processor executes computer execution instructions stored in memory, causing the processor to perform the methods described above.

[0065] Sixthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided above.

[0066] In a seventh aspect, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method provided above.

[0067] This application provides a control method, model training method, device, and vehicle for intelligent driving vehicles. By acquiring vehicle data and road constraint data and fusing them, the model can simultaneously focus on vehicle motion state, driving operation characteristics, and constraint information such as road geometry and traffic rules during training, thereby improving the completeness and scenario adaptability of driving style representation. By inputting the fused features into the initial intelligent model, processing them based on clustering and adversarial imitation modules, and optimizing the model parameters through a loss function, the model's ability to identify differences, continuous changes, and potential behavioral patterns in driving styles can be improved. This allows the obtained driving style recognition model to more accurately output driving styles and generate more personalized, compliant, and reliable vehicle driving commands based on these driving styles, thereby achieving effective control of vehicle driving. Attached Figure Description

[0068] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0069] Figure 1 A flowchart illustrating a training method for a driving style recognition model provided in this application embodiment. Figure 1 ;

[0070] Figure 2 A flowchart illustrating a training method for a driving style recognition model provided in this application embodiment. Figure 2 ;

[0071] Figure 3 A flowchart illustrating step S203 in a training method for a driving style recognition model provided in an embodiment of this application;

[0072] Figure 4 A flowchart illustrating a control method for an intelligent driving vehicle provided in this application embodiment. Figure 1 ;

[0073] Figure 5 A flowchart illustrating a control method for an intelligent driving vehicle provided in this application embodiment. Figure 2 ;

[0074] Figure 6 A schematic diagram of the structure of a training device for a driving style recognition model provided in an embodiment of this application;

[0075] Figure 7 This is a schematic diagram of the structure of a control device for an intelligent driving vehicle provided in an embodiment of this application;

[0076] Figure 8 This is a structural schematic diagram of a vehicle provided in an embodiment of this application.

[0077] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0078] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0079] Driving style recognition and intelligent driving control technology belong to the field of intelligent connected vehicles and driver assistance systems. They are typically applied in complex traffic scenarios such as highways, urban expressways, main urban roads, construction zones, and congested areas.

[0080] In existing technologies, driving style recognition typically uses vehicle operation data as the core input source, such as vehicle speed, acceleration, braking intensity, steering wheel angle, accelerator pedal opening, and lane change frequency. This data is analyzed through short-time window statistics or time-series feature extraction, and then traditional machine learning models or shallow neural networks are used to classify driving styles. A typical approach is to treat a driving sequence as an independent sample, cluster or classify the collected local behavioral features, output style labels such as aggressive, conservative, or neutral, and map these labels to driver assistance control parameters, such as adjusting the target distance for adaptive cruise control, the response threshold for lane keeping assist, the triggering conditions for lane change assist, and the intervention timing of automatic braking.

[0081] However, their technical approaches generally have significant shortcomings: On the one hand, the recognition process relies heavily on single or short-term driving data, focusing on the static classification of local behavioral fragments. It is difficult to continuously track the evolution of driving styles over long-term use, and the changes in driver behavior at different times, under different mental states, or on different road types are often not captured in time, resulting in a lag between model output and actual driving intent. On the other hand, existing solutions do not make sufficient use of constraints such as road speed limits, lane rules, safe distances, and road geometry. They often treat driver behavior as a single signal independent of the road environment, ignoring the mandatory restrictions imposed on reasonable driving strategies by scenarios such as curves, ramps, construction areas, and congested areas. This can easily lead to recognition results that are biased towards local preferences and weaken compliance requirements. In addition, existing models typically only process vehicle data at the feature fusion level, failing to model road constraint data and vehicle behavior data in a unified manner. Therefore, it is difficult to accurately depict the actual driving intent in complex scenarios and to ensure that the output control strategy achieves a balance between personalization, reliability, and safety boundaries. In practical applications, these defects can cause the system to react slowly when the driver's style changes and make decisions that are not robust enough when road rules are strictly enforced, thereby affecting the availability and safety of driver assistance functions.

[0082] Therefore, the embodiments of this application provide a control method, model training method, device, and vehicle for intelligent driving vehicles, which can solve the above-mentioned problems.

[0083] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0084] Figure 1 A flowchart illustrating a training method for a driving style recognition model provided in this application embodiment. Figure 1 ,like Figure 1 As shown, the method includes:

[0085] S101. Obtain vehicle data and road constraint data.

[0086] Among them, vehicle data represents multimodal time-series data of vehicle motion state and driving operation, while road constraint data represents the road geometry and traffic rules that constrain driving.

[0087] For example, vehicle data can be understood as a set of data that changes continuously over time, generated by the vehicle itself, driver operations, and onboard perception links, and at least covers temporal information reflecting longitudinal motion, lateral motion, and maneuvering behavior.

[0088] Multimodal time-series data indicates that the data source is not a single sensor, but a heterogeneous sequence from different physical quantities, different sampling frequencies and different interfaces, such as vehicle speed, longitudinal acceleration, yaw rate, steering wheel angle, brake pedal opening, accelerator pedal opening, gear status, turn signal status and environmental perception status.

[0089] Road constraint data can be understood as a set of restrictive information corresponding to the current vehicle driving space and traffic rules. It includes both road geometry information and static or dynamic traffic rule information. Road geometry may include road curvature, slope, number of lanes, lane width, ramp connection relationship, curve radius and road segment topology relationship. Traffic rule constraints may include speed limit signs, no-entry rules, traffic light status, lane function attributes, safe distance constraints and construction control information, etc.

[0090] For example, the executing entity can be an in-vehicle domain controller, an intelligent driving controller, a central computing platform, or an edge server that communicates with the vehicle. Vehicle data can be collected through an in-vehicle sensor network, where an inertial measurement unit outputs acceleration and angular velocity, wheel speed sensors output vehicle speed and wheel speed difference, steering angle sensors output steering wheel angle or front wheel angle, pedal position sensors output brake pedal opening and accelerator pedal opening, and the vehicle stability system and powertrain controller output gear position, drive force request, braking pressure, and stability control status via a CAN (Controller Area Network) bus.

[0091] Road constraint data can be jointly provided by a high-precision map module, a navigation and positioning module, a roadside infrastructure communication unit, and a vehicle-to-everything (V2X) platform. The high-precision map provides lane-level geometry, speed limits, and lane connectivity. The navigation module provides the road attributes corresponding to the current location point. The roadside infrastructure or V2X platform provides real-time traffic lights, construction areas, congestion levels, and temporary traffic control information.

[0092] In one possible embodiment, target detection results, lane line recognition results, distance to the vehicle ahead, relative speed, and drivable area information output by radar, camera, lidar, and inertial navigation modules can also be incorporated into vehicle data to enhance the ability to describe driving scenarios. Correspondingly, road constraint data can also include real-time traffic flow, road event levels, and information on changes in traffic rules caused by weather.

[0093] Optionally, this step may also perform time alignment, coordinate unification, outlier removal, and missing value completion on the vehicle data.

[0094] Specifically, different sensors have different sampling periods. For example, CAN signals may have a sampling period of 10 milliseconds, camera perception results may have a sampling period of 40 milliseconds, and high-precision map refresh may have a sampling period of 100 milliseconds or be event-triggered. Therefore, a unified timestamp caching mechanism can be adopted, using the master control clock or GPS (Global Positioning System (Expectation-Maximization)) time as a reference, to resample the data from each source to a unified time base.

[0095] For continuous signals, linear interpolation, zero-order hold, or Kalman filtering estimation can be used for completion; for discrete state variables, nearest neighbor hold can be used. To eliminate the spatial differences between vehicle data and road constraint data, perception data, map data, and vehicle positioning data can be unified into the vehicle coordinate system or global map coordinate system. Then, a road segment matching relationship can be established based on the vehicle's current position and driving direction, thereby accurately associating the driving operation at a certain moment with the corresponding road constraint.

[0096] By synchronously collecting and standardizing vehicle behavior information and road environment constraint information, style judgment no longer relies solely on local driving actions themselves, but rather imbues each driving behavior with a clear road context, laying the foundation for subsequent unified modeling. This collection method addresses the problem in existing technologies where driving style recognition ignores road geometry and traffic rules, enabling the model to distinguish between proactive and aggressive driver behavior and necessary actions constrained by road conditions during training, thereby reducing style label bias.

[0097] S102. Perform data fusion processing on vehicle data and road constraint data to obtain fusion features.

[0098] For example, data fusion processing refers to the process of transforming vehicle data and road constraint data from different sources, dimensions, and sampling characteristics into a joint representation in the same feature space; fused features represent feature results in vector, matrix, or tensor form that can simultaneously characterize driver behavior patterns and external road constraints, and are used for subsequent model learning.

[0099] For example, vehicle data can first be preprocessed and feature extracted. For continuous time-series signals such as vehicle speed, acceleration, steering wheel angle, and pedal opening, statistical and dynamic features can be extracted within a preset time window, such as mean, variance, extreme values, rate of change, kurtosis, skewness, zero-crossing rate, slope, and sliding window trend, to characterize the driver's operational intensity and stability over a period of time. Derivative features can also be constructed, such as the ratio of acceleration to speed, the rate of change of brake opening, steering wheel angular velocity, speed difference before and after lane changes, and the coupling features of target distance and relative speed, thereby reflecting behavioral characteristics during following, overtaking, and avoidance. For environmental perception data, features such as distance to the vehicle in front, relative speed, target density in adjacent lanes, available lane change gap, and road boundary offset can be extracted.

[0100] Road constraint data is encoded. Discrete or enumerated information such as speed limits, number of lanes, road class, lane function attributes, and traffic rule status can be converted into numerical representations using one-hot encoding, embedding encoding, or label vector encoding. Continuous road attributes such as road curvature, slope, curve radius, lane width, and safety distance constraints can be directly input after normalization. If the road topology is complex, such as involving ramps, divergences, merges, and multi-lane connections, road elements can be abstracted into nodes and edges in a graph structure, and topological constraint features can be extracted through graph representation learning.

[0101] After each sub-feature is formed, cross-modal fusion can be further performed.

[0102] In one possible implementation, a feature concatenation method can be used to align the vehicle behavior feature vector and the road constraint feature vector by time step and then concatenate them to form a unified input.

[0103] In one possible implementation, a gating fusion mechanism can also be used. The gating unit learns the weight distribution of vehicle behavior features and road constraint features in the current scenario, so that the model pays more attention to road constraints in strongly constrained scenarios such as construction zones, ramp zones, and congested zones, and pays more attention to individual differences in driver operation in ordinary straight road sections.

[0104] In one possible implementation, an attention fusion mechanism can also be used, with the vehicle behavior sequence as the query vector and the road constraint sequence as the key-value pair, to calculate cross-modal attention weights, thereby enhancing the road constraint information most relevant to the current operation.

[0105] In one possible implementation, for the temporal modeling part, a recurrent neural network, a temporal convolutional network, or a transformer encoder can be used to encode the vehicle data within the time window to obtain a temporal behavior representation.

[0106] In one possible implementation, the road topology and rule information can be encoded using a graph neural network, a multilayer perceptron, or a dual-tower encoder, and then jointly mapped with the temporal behavior representation.

[0107] By placing vehicle behavior and road environment in a unified feature space, the model's learning object is expanded from how a single vehicle moves to how a vehicle moves under what constraints. This fundamentally improves the adaptability of style recognition to complex traffic scenarios and enables long-term style evolution and short-term scene changes to be jointly characterized in the same representation, thus providing a semantically consistent input foundation for subsequent joint training of clustering and adversarial imitation.

[0108] S103. Input the fused features into the initial intelligent model. Based on the clustering module and adversarial imitation module of the initial intelligent model, process the fused features and optimize the parameters of the initial intelligent model through the loss function to obtain the driving style recognition model.

[0109] Among them, the driving style recognition model is used to process vehicle data and road constraint data to obtain driving style; driving style is used to generate vehicle driving commands; vehicle driving commands are used to control vehicle driving.

[0110] For example, the initial intelligent model can be understood as an un-optimized model that has not yet been trained on the target driving dataset, which internally includes at least an encoding network for feature representation learning, a clustering module for style structure discovery, and an adversarial imitation module for improving style transferability and robustness.

[0111] The clustering module is used to perform unsupervised or weakly supervised pattern division on the fused features, so that samples with similar driving behaviors and similar road adaptation methods are aggregated in the feature space to form potential driving style categories.

[0112] The adversarial imitation module is used to train the model through generative and discriminative adversarial training, enabling the model to learn the distribution of real driving behavior and enhance its ability to distinguish samples with blurred style boundaries.

[0113] For example, the fused features are first input into a trajectory encoder or a temporal encoder to obtain a high-dimensional latent representation. This latent representation can then be fed into a differentiable clustering module, which calculates the soft assignment result based on the distance between the sample and the cluster center, and outputs the probability that each sample belongs to a different style cluster.

[0114] The distance metric in the clustering module can be Euclidean distance, cosine distance, or Mahalanobis distance. Soft assignment can be implemented using the Softmax (Softmax function) function to ensure that the clustering process can participate in backpropagation.

[0115] The adversarial imitation module can include a conditional generator and a discriminator. The conditional generator receives fused features or clustering labels and generates pseudo-sample trajectories, pseudo-style embeddings, or pseudo-control tendency representations that closely resemble the distribution of real driving samples. The discriminator then judges the authenticity of the input samples and outputs the probability that they belong to real samples or generated samples. Through alternating game training between the generator and the discriminator, the model can learn a smoother and more generalizable driving style representation.

[0116] To ensure the model training has a clear objective, this step also requires constructing a joint loss function and performing parameter optimization. The loss function can include one or more combinations of clustering loss, adversarial loss, reconstruction loss, and classification loss.

[0117] Clustering loss is used to constrain the degree of clustering of similar driving behaviors in the feature space and to widen the distance between different style clusters; adversarial loss is used to constrain the distribution consistency between generated samples and real samples, and to improve the robustness of the model to unseen driving scenarios; reconstruction loss is used to constrain the encoding results to retain the key semantics in the original fused features and to prevent style information loss due to excessive compression; classification loss is used to enhance the separability of style categories in labeled or pseudo-labeled scenarios.

[0118] For example, the joint loss can be expressed as a weighted sum of clustering loss, adversarial loss, reconstruction loss, and classification loss; where each weight coefficient can be set according to training stability and recognition target, so that the model achieves a balance between clear clustering structure, accurate distribution fitting and style interpretability.

[0119] During training, an alternating training strategy can be adopted. First, the generator and discriminator are fixed, and the encoder and clustering module parameters are updated to obtain stable cluster labels. Then, the clustering results are fixed, and the generator and discriminator are updated to approximate the real sample distribution. Alternatively, a time decay factor can be introduced to assign higher weights to newer driving samples, thereby enhancing the model's ability to respond to recent changes in driving style while retaining long-term habit information represented by historical samples. After multiple rounds of iterative optimization, a converged driving style recognition model is obtained.

[0120] After training, the model receives new vehicle and road constraint data during the inference phase. Following the same acquisition and fusion process as S101 and S102, it outputs driving style results. Driving style can be a discrete label such as aggressive, conservative, or neutral, or a continuous style score, style probability distribution, or style embedding vector. Understandably, the decision module generates vehicle driving instructions based on this driving style result, such as maintaining a small or large following distance, decelerating in advance, accelerating moderately, prioritizing lane changes, delaying lane changes, or maintaining the current lane.

[0121] Furthermore, driving commands can be mapped to specific control quantities, such as target vehicle speed, acceleration / deceleration limit, braking intensity, steering angle change threshold, and lane change trigger threshold, and sent to the vehicle controller for execution, thereby controlling the vehicle's longitudinal and lateral driving states.

[0122] Optionally, to avoid conflicts between personalized style output and road safety rules, a real-time constraint verification module can be introduced before the instruction is issued to verify whether the target vehicle speed exceeds the speed limit, whether the following distance is lower than the safety threshold, and whether the lane change request violates lane rules; if there is a conflict, the instruction is corrected before being output.

[0123] By using a clustering module to uncover the latent structure of driving styles, an adversarial imitation module to improve the ability of style representation to fit real complex driving distributions, and a joint loss function to achieve unified optimization, the final driving style recognition model can not only track long-term changes in driver behavior, but also maintain recognition stability under complex road constraints such as curves, ramps, construction zones, and congested areas. This results in the output of vehicle driving commands that balance personalization and compliance, thereby improving the reliability, safety, and human-machine collaboration of the intelligent driving control link.

[0124] This application provides a training method for a driving style recognition model. It collects various vehicle data during vehicle operation and road constraint data from driving scenarios, fuses these two types of heterogeneous data to obtain fused features, and then inputs these features into an initial intelligent model. The model's built-in clustering module accurately classifies different driving behavior categories, and an adversarial imitation module enhances the differentiation of different driving style features. Simultaneously, iteratively optimizing the training parameters of the initial intelligent model based on a loss function completes model convergence and accuracy calibration. This effectively addresses the technical pain points of traditional driving style recognition models, such as single data dimension, poor environmental adaptability, ambiguous style feature differentiation, low recognition accuracy, and weak generalization ability. It achieves excellent results in closely matching complex real-world driving conditions, accurately and differentially recognizing various user driving styles, adapting the model to various driving scenarios, and significantly improving recognition accuracy and practical application generalization performance. Furthermore, by combining the mapping and constraint verification processing from driving style to control commands, the final generated vehicle driving commands can meet personalized driving preferences while complying with road speed limits, safe distances, and traffic rules, thus providing a more stable, reliable, and compliant control foundation for the driver assistance system.

[0125] Figure 2 A flowchart illustrating a training method for a driving style recognition model provided in this application embodiment. Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the examples, a training method for a driving style recognition model is described in detail, the method including:

[0126] S201. Obtain vehicle data and road constraint data.

[0127] For example, this step can be referred to step S101, and will not be described again.

[0128] S202. Encode the road constraint data to obtain constraint-coded data; perform statistical feature processing, sequence feature processing, and derived feature processing on the vehicle data to obtain vehicle feature data; perform cross-modal fusion processing on the constraint-coded data and vehicle feature data to obtain fused features.

[0129] For example, road constraint data is used to characterize the constraints imposed on driving behavior by the road environment in which the vehicle is located. It may include one or more of the following: road speed limit, number of lanes, lane width, road curvature, slope, curve radius, no-entry rules, traffic signal status, construction area markings, congestion level, and temporary traffic control information.

[0130] When encoding road constraint data, discrete rule information can be mapped into vector representations, and numerical constraints can be normalized to form constraint encoding data that can participate in model calculations.

[0131] Vehicle data is used to characterize vehicle motion state and driving operation information. It can consist of vehicle speed, acceleration, steering wheel angle, brake pedal opening, accelerator pedal opening, gear status, and turn signal status. Statistical features, temporal dependency features, and derived features are extracted to enhance the ability to express driving intentions.

[0132] For example, statistical feature processing can calculate the mean, variance, maximum, minimum and rate of change of vehicle speed, acceleration and steering wheel angle within a preset time window to reflect the overall distribution characteristics of vehicle operation.

[0133] Sequence feature processing can model continuous time series based on a sliding window to extract the evolutionary patterns of operations such as acceleration, deceleration, turning, and lane changing.

[0134] Derivative feature processing can construct acceleration change rate, steering wheel angle change rate, longitudinal and lateral coupling features, braking and acceleration switching frequency, and driving aggression index based on the original data.

[0135] Cross-modal fusion processing first aligns the constraint coding data and vehicle feature data dimensionally, and then forms a unified representation through feature concatenation, gating fusion, or attention fusion, enabling road constraints and vehicle behavior to interact within the same feature space. The coding model, temporal network, or fusion network used here can be selected according to the onboard computing power configuration. In practical applications, other models of this component can also be selected, and this application embodiment does not limit this.

[0136] By encoding road constraint data into structured features and decomposing vehicle data into three types of features—statistical, sequential, and derived—and then performing cross-modal fusion, the fused features can simultaneously contain road rule constraints and driving behavior pattern information. This improves the adaptability of subsequent driving style recognition to complex road scenarios and enhances the accuracy of representing changes in individual driving tendencies, making the output results more in line with the safety and consistency requirements of actual driving environments.

[0137] In one example, feature extraction is performed on the fused features to obtain vertical features, horizontal features, and macroscopic features.

[0138] Among them, longitudinal features represent the feature components of the vehicle's motion state and control behavior along the main driving direction; lateral features represent the feature components of the vehicle's motion state and control behavior perpendicular to the driving direction; and macroscopic features represent the features of high-level semantics within a preset time period.

[0139] For example, the fusion feature is a unified representation of vehicle data and road constraint data after cross-modal fusion. The longitudinal feature is used to characterize longitudinal acceleration and deceleration, following and braking habits, the lateral feature is used to characterize steering, lane changing and lane keeping behavior, and the macro feature is used to summarize the overall semantic tendency of driving style over a period of time.

[0140] In one possible implementation, longitudinal features can be constructed from vehicle speed, longitudinal acceleration, braking intensity, accelerator pedal opening, following distance change, and longitudinal impact, and then normalized and statistically analyzed using a time window to form a longitudinal feature vector. Lateral features can be constructed from steering wheel angle, angle change rate, lateral acceleration, lane departure, lane change frequency, and lateral stability, and then sequence-encoded to form a lateral feature vector. Macroscopic features involve attention convergence or temporal summarization of multiple local segments within a preset time period to form feature representations capable of characterizing high-level semantics such as frequent lane changes, continuous following, rapid acceleration, or smooth cruising. Each feature can be encoded separately first, and then concatenated or weighted to form a feature tensor suitable for input into the initial intelligent model.

[0141] The system first extracts longitudinal, lateral, and macroscopic representations from vehicle operation data, and then uses these as refined inputs for fusion features to enhance the ability to simultaneously perceive local operational details and global behavioral semantics. Since different dimensional features correspond to the vehicle's driving intentions along the driving direction, perpendicular to the driving direction, and across time scales, this processing enables subsequent clustering and discrimination to more accurately distinguish between aggressive, conservative, or neutral styles, and reduces the bias caused by single short-term features.

[0142] By adopting this implementation method, the initial intelligent model obtains structured multi-scale feature inputs before training, which can improve the separability and stability of driving style representation, enhance the ability to characterize driving intentions in complex road scenarios, and make the final driving style recognition model more suitable for application environments where personalized control and road constraints coexist.

[0143] S203. Input the fused features into the initial intelligent model, process the fused features based on the initial intelligent model to obtain features and discriminants; based on the features and discriminants, optimize the parameters of the initial intelligent model through alternating backpropagation to obtain the driving style recognition model.

[0144] Among them, the intermediate layer representation extracted by the internal module of the feature representation model; the discriminant represents the probability value output by the initial intelligent model, which is used to represent the input belonging to a certain category or a real sample.

[0145] For example, the initial intelligent model is an untrained neural network architecture containing a backbone network for feature extraction, a clustering module, and an adversarial mimicry module. The clustering module is responsible for unsupervised or weakly supervised clustering of fused features, attempting to group similar driving style samples into the same category to discover potential style types. The adversarial mimicry module is based on a generative adversarial network and includes an encoder, a generator, and a discriminator. The discriminator distinguishes whether the input features come from the real driving data distribution or are representations generated or transformed by the model's internal generator, forcing the model to learn more essential and discriminative style representations.

[0146] Optionally, the adversarial imitation module can be constructed using the InfoGAIL model; InfoGAIL, or Information Maximization Generative Adversarial Imitation Learning, serves as a complete adversarial component here, consisting of a generator, a discriminator, and an information adversarial module. The generator is responsible for sampling latent variables from the prior distribution and generating fake driving style intermediate features. The discriminator receives either real or generated features and outputs two key discriminant metrics: the probability that the input is a real feature, and a prediction of the latent variable distribution used for mutual information calculation. The information adversarial module utilizes the lower bound of the mutual information between the latent variables reconstructed by the discriminator and the real posterior latent variables to force a strong correlation between the latent variables and the driving style features.

[0147] Features are intermediate layer vectors output by modules within the model, representing low-dimensional embeddings of driving styles. Discriminants are probability values ​​output by the model used to determine categories or authenticity, such as the probability of belonging to a certain style given by a style classifier, or the probability that a sample is a true style sample output by an adversarial discriminator.

[0148] For example, the fused features of a batch of driving segments are input into the initial intelligent model, and intermediate layer features are obtained through an optional shared feature extraction network. These features first enter the clustering module, where the similarity between the features and each learnable cluster center is calculated, and then normalized using Softmax or Student's t-distribution to obtain a K-dimensional soft label vector. This label vector serves as the latent variable of driving style for that sample, representing both the clustering result and the structured information of the style.

[0149] InfoGAIL's generator receives the label vector and optionally concatenates it with random noise, then maps it through a multi-layer neural network to generate fake-style features of the same dimension as the real features. The real intermediate-layer features are then paired with their corresponding cluster labels to form real pairs; the fake features output by the generator are paired with the same labels to form fake pairs.

[0150] These two pairs of paired data, along with the labels, are fed into the discriminator. The discriminator operates in two ways: first, it outputs a scalar with a value between 0 and 1, representing the probability that the feature-label pair comes from the distribution of real driving features, i.e., the true / false discrimination value; second, at the top of the discriminator, it reconstructs the distribution parameters of the labels using the input features, i.e., it outputs the reconstructed label probability vector.

[0151] The information adversarial module then uses the soft labels directly output by the clustering module as the posterior label distribution and calculates the KL divergence between it and the label distribution reconstructed by the discriminator. This divergence is an approximation of the lower bound of the mutual information between features and labels.

[0152] In the optimization phase, an alternating backpropagation strategy is employed. First, the parameters of the feature extraction network and clustering module are fixed, and the InfoGAIL component is trained: the discriminator is updated by minimizing the classification error of true / false pairs; the parameters of the adversarial module are updated by maximizing the lower bound of mutual information; simultaneously, the generator is updated by maximizing the probability of generated features being classified as true and maximizing the mutual information between generated features and labels, ensuring that the generated style features are realistic and strongly correlated with the labels. Then, all InfoGAIL parameters are fixed, and the feature extraction network and clustering module are optimized using clustering loss (e.g., minimizing the KL divergence between soft labels and auxiliary target distributions) and adversarial loss (to make the generated features able to confuse the discriminator). This forces the labels discovered by the clustering module to be utilized by the generator to generate realistic features, thereby verifying the effectiveness of its style representation. This alternating iterative process continues until convergence, and the model's feature extraction network and clustering module constitute the final driving style recognition model.

[0153] By employing the above method, the model can simultaneously learn the clustering structure and discrimination boundary of driving styles while uniformly representing vehicle behavior and road constraint information. This allows the output to retain individual driver preferences while also taking into account road rule constraints. Compared to methods that rely solely on static classification, this implementation improves the stability, robustness, and scene adaptability of driving style recognition, and provides a more reliable model foundation for subsequent vehicle driving command generation.

[0154] Figure 3 This is a flowchart illustrating step S203 of a training method for a driving style recognition model provided in an embodiment of this application. Figure 3 As shown, step S203 includes:

[0155] S2031. Input the fused features into the initial intelligent model. The encoder based on the initial intelligent model processes the fused features to obtain temporal encoded features. The clustering module based on the initial intelligent model clusters the temporal encoded features to obtain temporal encoded features with cluster labels. The adversarial imitation module based on the initial intelligent model processes the temporal encoded features and the temporal encoded features with cluster labels to obtain features and discriminants.

[0156] For example, temporal coding features are used to characterize the behavioral evolution of a vehicle within a continuous time window, and their dimensions can correspond to the temporal length of the fused features. Temporal coding features with clustering labels refer to a representation that adds cluster category identifiers or cluster probability information to the original temporal coding features, making it easier for subsequent modules to distinguish different driving style modes.

[0157] In this embodiment, the encoder can employ a Long Short-Term Memory (LSTM) network, gated recurrent units (GRUs), temporal convolutional networks (TCRs), or a Transformer encoding structure to perform time-dependent modeling of the fused input features. It can also combine timestamp encoding or positional encoding to enhance the perception of behavioral order. The clustering module can perform soft clustering or prototype clustering based on the encoded feature vectors to output the cluster label to which each sample belongs. The label can correspond one-to-one with the style cluster center. The adversarial imitation module can include a generator and a discriminator. The generator reconstructs or perturbs the temporally encoded features under the constraint of the cluster label, while the discriminator determines whether the input features are consistent with the real driving distribution and outputs the features and discriminant value.

[0158] For example, the fused features of a batch of driving segments are input into an initial intelligent model, which is then temporally modeled by an encoder. The encoder, such as a bidirectional LSTM or a Transformer encoder, processes the input sequence step by step or globally, capturing the dynamic evolution patterns during the driving process, and finally outputs a fixed-dimensional temporally encoded feature vector, which condenses the key style information of the entire driving segment.

[0159] The temporal encoded features are then fed into the clustering module. Internally, the clustering module maintains K learnable cluster center vectors. By calculating the similarity between the temporal encoded features and each center, and normalizing the results using a Student's t-distribution or a Softmax function, a K-dimensional soft label probability vector is generated. The temporal encoded features output in this step, bearing cluster labels, are essentially the result of aligning and pairing the original encoded features with the cluster labels. The label vector represents the latent variable of driving style for that sample, imbuing the features with style semantics.

[0160] The labeled feature is fed into the adversarial imitation module. The generator in this module, conditioned on the clustered label vector, optionally concatenates a small amount of random noise and generates a fake style feature with the same dimension as the real temporal encoded feature through a multilayer perceptron mapping.

[0161] The discriminator receives two types of paired inputs: one is the pairing of genuine temporal encoded features with their corresponding cluster labels, and the other is the pairing of fake features output by the generator with the same label. For each input pair, the discriminator performs two tasks: first, it outputs a scalar discriminant, representing the probability that the input pair belongs to the distribution of genuine driving style features; second, it reconstructs the label distribution at its output, outputting a reconstructed label probability vector as another discriminant.

[0162] The soft labels output by the clustering module are used as the posterior distribution, and the KL divergence is calculated with the label distribution reconstructed by the discriminator, serving as an approximation of the lower bound of the mutual information between features and latent variables. Thus, a complete forward propagation yields temporally encoded features and style label probabilities for clustering optimization, as well as discriminative quantities such as true / false probabilities and reconstructed label probabilities for adversarial training.

[0163] S2032. Based on features, discriminants, and loss functions, determine the loss value of the initial intelligent model; based on the loss value, optimize the parameters of the initial intelligent model through alternating backpropagation to obtain the driving style recognition model.

[0164] For example, the loss function is used to measure the deviation between the feature distribution and the preset style structure, as well as the difference between the discriminant and the target label. It can be composed of clustering loss, adversarial loss, imitation loss, classification loss and their weighted combination. The weight coefficients are set according to the training objective to balance the relationship between clustering constraints, discriminant constraints and imitation constraints.

[0165] For example, during the forward propagation phase, the model has already obtained features and various discriminant parameters, including the style probability labels output by the clustering module after soft assignment of temporally encoded features, and the true / false probabilities and reconstructed label probabilities given by the discriminator in the adversarial imitation module.

[0166] In the loss calculation stage, the clustering loss is first constructed: the soft label probability output by the clustering module is regarded as the current distribution, and the KL divergence is usually calculated with its own sharpened auxiliary target distribution, or a self-supervised method is used to make the same batch of samples clustered compactly and separated between classes, so as to obtain the clustering loss value.

[0167] Constructing adversarial generation loss: For the discriminator, the goal is to maximize the log probability of a real feature-label pair being judged as true, while maximizing the log probability of a fake feature-label pair being judged as false, thus generating the discriminator loss; for the generator, the goal is to minimize the log probability of a fake feature-label pair being judged as false, i.e., maximize the probability of being judged as true, thus generating the generator loss.

[0168] Constructing information maximization loss: InfoGAIL's information adversarial module uses the soft labels output by the clustering module as the posterior distribution and the labels reconstructed by the discriminator as the prior prediction distribution. It calculates the KL divergence between the two and uses it as the negative value of the lower bound of mutual information. Maximizing mutual information means minimizing this KL divergence loss.

[0169] These three losses are summed according to preset weights to obtain the total loss value for the current batch. Then, the alternating backpropagation optimization phase begins.

[0170] The first update involves fixing the parameters of the feature extraction encoder and clustering module, backpropagating the part of the total loss related to the adversarial imitation module, and letting the gradient flow through the discriminator and generator. The discriminator parameters are updated to improve its ability to distinguish between real and fake and to accurately reconstruct labels. At the same time, the generator parameters are updated to make the generated style features more realistic and highly bound to the given labels.

[0171] The second update involves fixing all parameters of the adversarial imitation module. Using the gradient of the clustering loss and the portion of the adversarial loss that the generator needs to use to deceive the discriminator, the parameters of the encoder and clustering module are updated. This forces the temporal encoded features extracted by the encoder to form cluster structures that are conducive to clustering, while also effectively supporting the generator's reproduction of its style. This process is repeated alternately, with each update recalculating the loss value based on the current batch of data and backpropagating, until the model's loss converges on the validation set or the preset number of iterations is reached. The final encoder plus clustering module constitutes the driving style recognition model.

[0172] This application provides a training method for a driving style recognition model. It collects vehicle data and road constraint data corresponding to driving scenarios, encodes the road constraint data to obtain constraint-encoded data, and simultaneously processes the vehicle data through multi-dimensional processing of statistical features, sequence features, and derived features to generate vehicle feature data that comprehensively covers driving behavior. Then, it performs cross-modal fusion processing on two different types of data to obtain integrated fusion features. These fusion features are then input into an initial intelligent model, which outputs the intermediate-layer features extracted by the model and the category probability discriminant. The model parameters are continuously optimized using an alternating backpropagation method based on the features and discriminant. This method solves the technical problems of traditional training methods, such as single feature extraction, low road environment correlation, poor modal data adaptability, slow model training convergence, and inaccurate style classification. It achieves the effects of accurately matching actual complex driving scenarios, fully considering the correlation between vehicle driving behavior and road constraints, strong model training stability, high driving style recognition accuracy, and excellent scenario adaptation and generalization ability.

[0173] Figure 4 A flowchart illustrating a control method for an intelligent driving vehicle provided in this application embodiment. Figure 1 ,like Figure 4 As shown, the method includes:

[0174] S401. Obtain vehicle data and road constraint data.

[0175] For example, an intelligent driving vehicle can be understood as an intelligent driving vehicle equipped with an environmental perception unit, a positioning module, an on-board controller, a communication unit, and an actuator, which can perform control actions based on perception and decision results in manual driving assistance, partial autonomous driving, or higher-level driving control modes.

[0176] Vehicle data can be understood as the operational information collected in real time by onboard sensors, transmitted via the vehicle bus, or output by the control unit during vehicle operation. It is used to characterize the multimodal temporal features of vehicle motion state and driving operation behavior.

[0177] Road constraint data can be understood as data that characterizes the current road traffic boundaries, traffic rules, road geometry, and safety restrictions, used to limit the driving range and compliance control boundaries of vehicles in the current scenario.

[0178] For example, the executing entity may be an in-vehicle domain controller, an intelligent driving controller, a central computing platform, or an edge computing node that communicates with the in-vehicle unit.

[0179] Vehicle data can be collected through vehicle speed sensors, longitudinal and lateral acceleration sensors, gyroscopes, steering wheel angle sensors, brake pedal position sensors, accelerator pedal opening sensors, wheel speed sensors, global navigation satellite positioning modules, inertial navigation modules, and the vehicle CAN bus interface, thereby obtaining information such as vehicle speed, acceleration, angular velocity, lateral displacement, longitudinal displacement, steering wheel operation frequency, pedal change amplitude, following distance, lane position deviation, and lane change trigger frequency.

[0180] Road constraint data can be obtained through high-precision map interfaces, vehicle camera recognition results, millimeter-wave radar and lidar perception results, road infrastructure communication units, vehicle-road cooperative OBU (On-Board Unit), cloud-based road information services, and traffic rule databases. This results in information such as speed limit signs, lane distribution, road geometry, curve radius, slope, lane width, ramp location, construction area markings, congested area information, lane change prohibition rules, safe distance thresholds, and constraint parameters for special road sections.

[0181] Optionally, vehicle data and road constraint data can be processed after collection, including timestamp alignment, coordinate system unification, noise filtering, and missing value compensation.

[0182] Timestamp alignment is used to ensure that data from different sources correspond to the same moment or the same time window. Coordinate system 1 is used to map the local coordinates of the vehicle to the global coordinates of the map to a unified reference system. Noise filtering can be done by moving average, Kalman filtering or median filtering. Missing value compensation can be done by interpolation, recent time backfilling or model estimation.

[0183] By simultaneously introducing vehicle behavior information and road constraint information during the data acquisition stage, driver operations are no longer regarded as isolated signals detached from the road scenario. This provides a data foundation from the source for subsequent recognition results to balance personalization and compliance. It helps to solve the technical problems of existing technologies that rely solely on single vehicle operation data, resulting in lagging style recognition, insufficient scene adaptability, and low decision-making stability under complex road conditions.

[0184] S402. Input the vehicle data and road constraint data into the driving style recognition model, process the vehicle data and road constraint data based on the driving style recognition model, and obtain the driving style output by the driving style recognition model.

[0185] The driving style recognition model is trained using the same training method as the driving style recognition model.

[0186] For example, the driving style recognition model can be understood as an intelligent model that jointly encodes, fuses, and models the temporal features of vehicle behavior and road constraint features, and outputs the driving tendency result. Its model parameters are not set manually and statically, but are obtained by joint training of clustering module and adversarial imitation module based on the fusion sample of vehicle data and road constraint data.

[0187] For example, vehicle data can first be organized into a multi-channel time-series trajectory, where each time step corresponds to at least vectors such as vehicle speed, longitudinal acceleration, lateral acceleration, steering wheel angle, braking input, throttle input, relative distance to the vehicle in front, and lane position; road constraint data is then converted into a constraint feature sequence or scene feature vector corresponding to the time window, including at least speed limit, lane permissible driving direction, road curvature, curve grade, safe distance requirement, construction restriction markers, and congestion level.

[0188] The model can utilize a trajectory encoder to extract temporal representations of vehicle behavior. The trajectory encoder can employ recurrent neural networks, temporal convolutional networks, Transformer encoders, or combinations thereof to extract the operational rhythm, speed change patterns, and following response patterns during driving. Road constraint data can be mapped to a unified feature space via a constraint encoder. The constraint encoder can employ multilayer perceptrons, embedding layers, graph structured coding networks, or attention coding structures to transform discrete rule information and continuous geometric parameters into a computable constraint representation. Subsequently, vehicle behavior features and road constraint features enter a cross-modal fusion module. This fusion module can employ feature concatenation, gating fusion, cross-attention mechanisms, or graph neural networks to ensure that the model considers whether the driver's actions occur within speed limits, on curved roads, in construction zones, or in congested environments when identifying driving styles.

[0189] The fused features are fed into the clustering module and the adversarial mimicry module in the model. The clustering module can be understood as a clustering structure capable of softly assigning samples and allowing the clustering results to participate in gradient backpropagation. Its role is to discover the style distribution of different driving behavior segments in a unified feature space, such as dividing samples into different regions within radical, neutral, conservative, or continuous spectrums. The adversarial mimicry module can be understood as a structure where a conditional generator and a discriminator engage in a game of mutual respect, enabling the model to learn the distribution of real driving behavior and improve its ability to represent style boundaries. The conditional generator generates behavior samples corresponding to a certain style based on clustering labels, road constraint representations, and historical behavior encodings. The discriminator distinguishes between real and generated samples and drives the encoder and generator to learn potential representations that more closely resemble real driving styles.

[0190] The currently collected vehicle data and road constraint data are fed into the above model. The output can be a discrete style label or a continuous style parameter. For example, a continuous value between 0 and 1 can represent the degree of style change from conservative to aggressive.

[0191] In this way, the system can not only determine which driving style the driver is closer to at a certain moment, but also continuously track the changing trend of the style over a long period of use.

[0192] By jointly modeling road constraint data and vehicle behavior data in a unified model and optimizing parameters through clustering and adversarial imitation learning mechanisms, the model can learn long-term driver behavior patterns while avoiding recognition results that deviate from the true intent in scenarios with strong rule constraints such as curves, ramps, construction zones, and speed limit change zones. This improves the dynamic adaptability, scenario robustness, and rule consistency of driving style recognition, providing reliable high-level semantic input for subsequent vehicle driving command generation.

[0193] S403. Input vehicle data, road constraint data, and driving style into the instruction generation model. Process the vehicle data, road constraint data, and driving style based on the instruction generation model to obtain the vehicle driving instructions output by the instruction generation model.

[0194] For example, the instruction generation model can be understood as a model structure that generates macro-strategy instructions or specific control instructions suitable for the vehicle control system based on the current vehicle operating state, road constraints, and driving style recognition results.

[0195] Driving style is used to characterize the following distance, acceleration and deceleration rhythm, lane change tendency, and aggressiveness of response that a vehicle should adopt while meeting road rules and safety boundaries.

[0196] For example, the instruction generation model may include a semantic policy generation layer, a parameter refinement layer, and a rule constraint correction layer. The semantic policy generation layer can generate macro-level driving strategies based on driving style and scenario state, such as maintaining the current lane and decelerating smoothly, changing lanes in advance when safety conditions are met, increasing following distance and limiting maximum acceleration, maintaining cruise control and suppressing frequent steering corrections; the parameter refinement layer further maps the macro-level strategies into executable control parameters such as target speed, acceleration threshold, deceleration limit, steering angle correction, minimum following distance, lane change trigger threshold, and lane keeping sensitivity.

[0197] The model can be implemented using a hybrid architecture of a large language model, a sequence decision network, a multi-task policy network, or a rule network and a neural network. The large language model is used to generate upper-level driving policies with semantic interpretation capabilities, the policy network is used to complete the mapping from semantics to parameters, and the rule constraint correction layer is used to check whether all instructions comply with road restrictions before output.

[0198] In one possible embodiment, the system first converts the driving style into several preference coefficients, such as target speed preference coefficient, lane change aggressiveness coefficient, comfort braking coefficient, and following conservative coefficient, and then performs boundary clipping on these preference coefficients in combination with speed limit value, curve radius, safe distance threshold, and construction area restrictions.

[0199] For example, when the driving style recognition result is biased towards the aggressive type, although the model can increase the target speed setpoint and shorten the target following distance, the rule constraint correction layer will limit the upper limit of the target speed and the lower limit of the minimum following distance based on the current road speed limit, the change of curvature ahead, and the minimum safe distance requirement, thereby preventing personalized control from exceeding the road's allowable range. When the recognition result is biased towards the conservative type, the model can correspondingly increase the safety margin, reduce the rate of change of acceleration, and delay the lane change trigger. To ensure the interpretability and executability of the output commands, the command generation model can also call the real-time constraint verification submodule after each command generation to detect whether the target speed exceeds the speed limit, whether the following distance is less than the safety threshold, whether the lane change intention conflicts with the lane rules, and whether the steering command conflicts with the road boundary. If any of these constraints are not met, the correction mechanism is triggered to readjust the control parameters or revert to a more conservative default control strategy.

[0200] The generated vehicle driving instructions can be a combination of macro-level instructions and specific instructions. For example, it can output a strategy description of "prioritizing lane keeping" along with the corresponding target speed, acceleration threshold, and lane centering control gain.

[0201] By introducing driving style recognition results as conditional inputs into the vehicle control decision-making chain, the control logic no longer reacts solely based on the real-time vehicle state and fixed rule templates, but can reflect the driver's long-term personalized driving preferences within the safety boundaries. At the same time, the road constraint correction mechanism ensures that the final output command always meets road rules and environmental constraints, thereby establishing a stable balance between comfort, personalization, and safety. This solves the problem in existing technologies where control strategies are difficult to adapt to both changes in driver style and complex road rules.

[0202] S404. Control vehicle movement based on vehicle driving commands.

[0203] For example, the vehicle controller can map command parameters such as target speed, acceleration threshold, minimum following distance, target lane selection, and steering correction amount to adaptive cruise control (ACC) parameters, automatic emergency braking (AEB) threshold, lane keeping assist (LKA) control parameters, lane change execution conditions, and electronic power steering control amount.

[0204] For example, if the output is a macro-level strategy instruction, it needs to be further converted into a low-level control objective by a strategy interpreter. For example, maintaining a safe distance and appropriately decelerating can be interpreted as reducing the cruise target speed, increasing the time-distance setting value, and increasing the sensitivity of forward collision risk monitoring; prioritizing lane change under certain conditions can be interpreted as starting lane change candidate search, determining the available space in adjacent lanes, and outputting a path tracking instruction when turning is feasible.

[0205] During execution, the system can use the feasibility assessment module to comprehensively consider the current tire adhesion status, vehicle yaw stability, road boundaries, adjacent lane occupancy, predicted trajectories of surrounding objects, and traffic rules to determine whether the current instruction is executable.

[0206] If conditions permit, the controller adjusts the torque output through the power system, applies the corresponding braking force through the braking system, outputs the required steering angle through the steering actuator, and coordinates the lateral and longitudinal coupling responses through the vehicle stability control system.

[0207] If conditions are not met, such as the target acceleration exceeding the current allowable value for road surface adhesion, insufficient lane change space, conflict between the target speed and the speed limit of the construction zone ahead, or the steering correction amount potentially causing the vehicle to cross the line, the system can trigger a secondary correction, delay execution, or downgrade to a more conservative control mode to avoid unsafe actions.

[0208] To ensure the effectiveness of closed-loop control, the vehicle will also transmit new vehicle status data and environmental perception results in real time after executing the command, so that the system re-enters the loop of data acquisition, style recognition, command generation and control execution, forming a continuously updated control closed loop.

[0209] In one possible embodiment, when the system receives a deceleration command, the controller gradually reduces the target vehicle speed according to a preset acceleration change rate limit, while increasing the safe following distance and dynamically adjusting the braking force distribution based on the change in the distance to the vehicle in front, so as to reduce the impact of abrupt deceleration on comfort; when the system receives a priority lane change command, the controller first confirms that the target lane line attributes, the speed difference of the following vehicles, and the target gap length meet the lane change requirements, then outputs the steering control quantity according to the planned trajectory, and restores lane keeping control after the lane change is completed.

[0210] By setting up a mapping mechanism, a feasibility judgment mechanism, and a closed-loop feedback mechanism between instructions and execution mechanisms, the vehicle driving instructions obtained by integrating vehicle behavior preferences and road constraints can be stably transformed into vehicle motion control actions. This ensures that personalized control strategies can be implemented and that the vehicle dynamics limits and road safety boundaries are not exceeded in complex traffic scenarios, thereby improving the stability, controllability, and reliability of vehicle control in practical applications.

[0211] This application provides a method for controlling an intelligent driving vehicle. It collects vehicle data corresponding to the vehicle's driving state and road constraint data corresponding to the driving scenario in real time. These two types of core data are input into a pre-trained driving style recognition model. The model intelligently analyzes and processes multi-dimensional data to accurately identify and output a driving style adapted to the current scenario and vehicle condition. Simultaneously, the vehicle data, road constraint data, and the identified driving style are input into an instruction generation model. The model then generates adaptive vehicle driving instructions through comprehensive multi-factor computation. Finally, based on the generated compliant driving instructions, the method precisely controls the vehicle's driving state. This solves the problems of traditional intelligent driving control modes being rigid and unable to adapt to different road scenarios and diverse driving preferences. It achieves a good effect of intelligent driving conforming to actual road conditions, matching the personalized driving habits of drivers and passengers, and balancing vehicle driving safety, driving comfort, and intelligent driving control adaptability.

[0212] Figure 5 A flowchart illustrating a control method for an intelligent driving vehicle provided in this application embodiment. Figure 2 ,like Figure 5 As shown, in this embodiment... Figure 4 Based on the embodiments, the flow of a control method for an intelligent driving vehicle is described in detail, the method including:

[0213] S501. Obtain vehicle data and road constraint data for vehicle movement.

[0214] For example, this step can be referred to step S401, and will not be described again.

[0215] S502. Input the vehicle data and road constraint data into the driving style recognition model, process the vehicle data and road constraint data based on the driving style recognition model, and obtain the driving style output by the driving style recognition model.

[0216] For example, this step can be referred to step S402, and will not be repeated here.

[0217] S503. Input vehicle data, road constraint data, and driving style into the large language model of the instruction generation model for processing to obtain macro instructions; input macro instructions, vehicle data, road constraint data, and driving style into the policy gradient module of the instruction generation model for processing to obtain vehicle driving instructions.

[0218] Among them, macro-level instructions represent high-level instructions that the vehicle should follow in terms of driving intentions and behavioral decisions within a future preset time window.

[0219] For example, a large language model is a language model pre-trained on massive amounts of text and possessing common-sense reasoning and long-range planning capabilities. Here, it is used for high-level driving intention reasoning and decision-making. Macro instructions are high-level instructions that represent the driving intentions and behavioral decisions that the vehicle should follow within a preset time window in the future. They are usually presented in the form of structured natural language or intention vectors, such as "Smoothly change lanes to the right lane and decelerate to below 40 km / h in preparation for a right turn within the next 8 seconds."

[0220] The policy gradient module is an action generation network built on reinforcement learning methods. It adopts an Actor-Critic architecture and is optimized using the policy gradient algorithm. It is responsible for refining high-level macro instructions into executable trajectories or control sequences.

[0221] For example, the vehicle data at the current moment, the road constraint data within a local area, and the current driver's driving style identifier or vector are input together into the large language model module of the instruction generation model.

[0222] The large language model internally transforms multi-source heterogeneous inputs into labeled sequences or feature embeddings. For example, it converts numerical vehicle data into structured descriptive statements, encodes map constraints into spatial relationship text, and injects driving styles in the form of cue words or style vectors. Based on its internalized driving common sense and planning capabilities acquired through adversarial training, the large language model performs contextual reasoning and long-term decision-making to generate a natural language description—a macro-instruction—representing the vehicle's driving intentions and behavioral decisions within a preset future time window. This macro-instruction not only specifies the tactical objective but also implicitly includes constraints on execution style and safety boundaries.

[0223] The model then enters the next level of refined execution. The generated macro-level instructions, along with the original vehicle data, road constraint data, and driving style, are transmitted as state input to the policy gradient module. Within the policy gradient module, the macro-level instructions are first encoded into intent feature vectors, and then concatenated or fused with the current environment state vector and style condition vector through cross-attention to form a complete decision state representation.

[0224] The policy network infers and outputs vehicle driving commands to be executed based on the state representation, such as the planned trajectory points and expected speed curve for the next few seconds, or directly outputs the steering wheel angle and acceleration / deceleration control values. The value network simultaneously provides an estimate of the expected reward of the current state-action pair for subsequent policy improvement. Throughout the process, the large language model is responsible for providing the logical deduction of global coordination and high-level decision-making, while the policy gradient module is responsible for implementing fine-grained trajectory optimization and closed-loop control in a high-dimensional continuous action space. The two work together serially to complete the end-to-end mapping from perceptual input to final control commands.

[0225] By decomposing driving decisions into two levels—macro-intention generation and policy refinement—the model can first generate high-level decisions that conform to the road scenario and driving preferences. Then, the policy gradient module completes the fine-grained control parameter solution, thereby reducing the search complexity when directly generating control quantities. Since the macro-instructions contain the driving intentions within a preset future time window, the system can take into account road rules, current traffic conditions, and changes in driving style in advance. This improves the compliance, consistency, and personalized matching capabilities of vehicle driving instructions, and enhances the stability and usability of intelligent driving control in complex road scenarios.

[0226] S504. Control vehicle movement based on vehicle driving commands.

[0227] For example, this step can be referred to step S404, and will not be repeated here.

[0228] This application provides a method for controlling intelligent driving vehicles. It acquires vehicle data corresponding to the vehicle's driving state and road constraint data under the current scenario in real time, inputs these two types of core data into a driving style recognition model, and outputs the driving style of the current driving scenario. Then, it simultaneously inputs the vehicle data, road constraint data, and driving style into a large language model of an instruction generation model, relying on the model's semantic and scenario understanding capabilities to generate macro-instructions that fit the style. Subsequently, it inputs the macro-instructions, various basic data, and driving style into the strategy gradient module of the instruction generation model for refined computation and processing, ultimately generating vehicle driving instructions. This method effectively solves the technical pain points of traditional intelligent driving control modes, such as homogenization, poor adaptability, inability to match different road conditions and driving preferences, and rigid and inflexible driving control. It achieves intelligent driving vehicle driving control that aligns with personalized driving habits, adapts to various road constraint scenarios, provides smooth and natural driving control, and makes precise decisions that meet actual driving needs, significantly improving the safety, comfort, and driving experience of intelligent driving.

[0229] Figure 6 This is a schematic diagram of the structure of a training device for a driving style recognition model provided in an embodiment of this application, as shown below. Figure 6As shown, the training device 60 for a driving style recognition model provided in this embodiment includes:

[0230] The acquisition module 601 is used to acquire vehicle data and road constraint data of the vehicle's movement; wherein, the vehicle data represents the multimodal time series data of the vehicle's motion state and driving operation; the road constraint data represents the constraint information of the road geometry and traffic rules on the vehicle's movement.

[0231] Processing module 602 is used to perform data fusion processing on vehicle data and road constraint data to obtain fusion features;

[0232] The training module 603 is used to input the fused features into the initial intelligent model. The clustering module and the adversarial imitation module based on the initial intelligent model process the fused features and optimize the parameters of the initial intelligent model through a loss function to obtain the driving style recognition model. The driving style recognition model is used to process the vehicle data and road constraint data of the vehicle to obtain the driving style. The driving style is used to generate vehicle driving commands. The vehicle driving commands are used to control the vehicle's driving.

[0233] In one possible embodiment, the processing module 602 is configured to:

[0234] The road constraint data is encoded to obtain constraint-coded data;

[0235] The vehicle data is processed by statistical feature processing, sequence feature processing, and derived feature processing to obtain vehicle feature data.

[0236] Cross-modal fusion processing is performed on the constraint-coded data and vehicle feature data to obtain fused features.

[0237] In one possible embodiment, training module 603 is used for:

[0238] The fused features are input into the initial intelligent model, and the fused features are processed based on the initial intelligent model to obtain features and discriminants; where the features represent the intermediate layer representation extracted by the internal modules of the model; the discriminants represent the probability value of the fused features belonging to the real samples;

[0239] Based on features and discriminants, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain a driving style recognition model.

[0240] In one possible embodiment, the fused features are input into an initial intelligent model, and the fused features are processed based on the initial intelligent model to obtain features and discriminants, including:

[0241] The fused features are input into the initial intelligent model, and the encoder based on the initial intelligent model processes the fused features to obtain temporal encoded features;

[0242] The clustering module based on the initial intelligent model performs clustering processing on the temporal coding features to obtain temporal coding features with clustering labels.

[0243] The adversarial imitation module based on the initial intelligent model processes the temporal coding features and the temporal coding features with clustering labels to obtain features and discriminants.

[0244] In one possible embodiment, based on features and discriminants, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain a driving style recognition model, including:

[0245] Based on features, discriminants, and loss functions, the loss value of the initial intelligent model is determined;

[0246] Based on the loss value, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain the driving style recognition model.

[0247] In one possible embodiment, prior to training module 603, the following is also included:

[0248] The extraction module 604 is used to perform feature extraction processing on the fused features to obtain longitudinal features, lateral features and macro features; among them, the longitudinal features represent the feature components of the vehicle's motion state and control behavior along the main driving direction; the lateral features represent the feature components of the vehicle's motion state and control behavior perpendicular to the driving direction; and the macro features represent the features of high-level semantics within a preset time period.

[0249] This embodiment provides a training device for a driving style recognition model, which can execute the method provided in the above-described method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0250] Figure 7 This is a schematic diagram of the structure of a control device for an intelligent driving vehicle provided in an embodiment of this application, as shown below. Figure 7 As shown, the control device 70 for an intelligent driving vehicle provided in this embodiment includes:

[0251] The acquisition module 701 is used to acquire vehicle data and road constraint data of the vehicle's movement;

[0252] The recognition module 702 is used to input vehicle data and road constraint data into the driving style recognition model, process the vehicle data and road constraint data based on the driving style recognition model, and obtain the driving style output by the driving style recognition model; wherein, the driving style recognition model is trained by the driving style recognition model training device provided above.

[0253] The generation module 703 is used to input vehicle data, road constraint data and driving style into the instruction generation model, and process the vehicle data, road constraint data and driving style based on the instruction generation model to obtain the vehicle driving instructions output by the instruction generation model.

[0254] The control module 704 is used to control the vehicle's movement based on vehicle driving commands.

[0255] In one possible embodiment, the generation module 703 is used for:

[0256] Vehicle data, road constraint data, and driving style are input into the large language model of the instruction generation model for processing to obtain macro instructions; among them, macro instructions represent the high-level instructions that the vehicle should follow in terms of driving intentions and behavioral decisions within a future preset time window.

[0257] Macro-level instructions, vehicle data, road constraint data, and driving style are input into the policy gradient module of the instruction generation model for processing to obtain vehicle driving instructions.

[0258] This embodiment provides a control device for an intelligent driving vehicle, which can execute the method provided in the above-described method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0259] Figure 8 This is a schematic diagram of the structure of a vehicle provided in an embodiment of this application. Figure 8 As shown, the vehicle 80 provided in this embodiment includes at least one processor 801 and a memory 802. Optionally, the vehicle 80 also includes a communication component 803. The processor 801, memory 802, and communication component 803 are connected via a bus 804.

[0260] In a specific implementation, at least one processor 801 executes computer execution instructions stored in memory 802, causing at least one processor 801 to perform the above-described method.

[0261] The specific implementation process of processor 801 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0262] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0263] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0264] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0265] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0266] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0267] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0268] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0269] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0270] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0271] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0272] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0273] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0274] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A training method for a driving style recognition model, characterized in that, include: Acquire vehicle data and road constraint data; wherein, the vehicle data represents multimodal time-series data of vehicle motion state and driving operation; the road constraint data represents the constraint information of road geometry and traffic rules on driving. The vehicle data and the road constraint data are fused together to obtain fused features; The fused features are input into an initial intelligent model. The clustering module and adversarial imitation module of the initial intelligent model process the fused features, and the parameters of the initial intelligent model are optimized through a loss function to obtain a driving style recognition model. The driving style recognition model is used to process vehicle data and road constraint data to obtain a driving style. The driving style is used to generate vehicle driving commands. The vehicle driving commands are used to control vehicle driving.

2. The method according to claim 1, characterized in that, The vehicle data and the road constraint data are fused to obtain fused features, including: The road constraint data is encoded to obtain constraint-coded data; The vehicle data is subjected to statistical feature processing, sequence feature processing, and derived feature processing to obtain vehicle feature data. The constraint-coded data and the vehicle feature data are subjected to cross-modal fusion processing to obtain the fused features.

3. The method according to claim 1, characterized in that, The fused features are input into an initial intelligent model. The clustering module and adversarial mimicry module of the initial intelligent model process the fused features, and the parameters of the initial intelligent model are optimized using a loss function to obtain a driving style recognition model, including: The fused features are input into an initial intelligent model, and the fused features are processed based on the initial intelligent model to obtain features and a discriminant; wherein, the features represent the intermediate layer representation extracted by the internal module of the model; the discriminant represents the probability value of the fused features belonging to the real sample; Based on the features and the discriminant, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain a driving style recognition model.

4. The method according to claim 3, characterized in that, The fused features are input into an initial intelligent model, and the fused features are processed based on the initial intelligent model to obtain features and discriminants, including: The fused features are input into the initial intelligent model, and the encoder of the initial intelligent model processes the fused features to obtain temporal encoded features; The clustering module based on the initial intelligent model performs clustering processing on the temporal coding features to obtain temporal coding features with clustering labels. The adversarial imitation module based on the initial intelligent model processes the temporal coding features and the temporal coding features with clustering labels to obtain features and discriminants.

5. The method according to claim 3, characterized in that, Based on the features and the discriminant, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain a driving style recognition model, including: Based on the features, the discriminant and the loss function, the loss value of the initial intelligent model is determined; Based on the loss value, the parameters of the initial intelligent model are optimized through alternating backpropagation to obtain a driving style recognition model.

6. The method according to any one of claims 1-5, characterized in that, Before obtaining the driving style recognition model, the following steps are included: First, the fused features are input into an initial intelligent model. Then, the clustering module and adversarial mimicry module of the initial intelligent model process the fused features. Finally, the parameters of the initial intelligent model are optimized using a loss function. The fused features are subjected to feature extraction processing to obtain longitudinal features, lateral features, and macro features; wherein, the longitudinal features represent the feature components of the vehicle's motion state and control behavior along the main driving direction; the lateral features represent the feature components of the vehicle's motion state and control behavior perpendicular to the driving direction; and the macro features represent the features of high-level semantics within a preset time period.

7. A control method for an intelligent driving vehicle, characterized in that, include: Acquire vehicle data and road constraint data; The vehicle data and the road constraint data are input into the driving style recognition model. The vehicle data and the road constraint data are processed based on the driving style recognition model to obtain the driving style output by the driving style recognition model. The driving style recognition model is trained by the driving style recognition model training method as described in any one of claims 1-6. The vehicle data, road constraint data, and driving style are input into the instruction generation model. The vehicle data, road constraint data, and driving style are processed based on the instruction generation model to obtain the vehicle driving instructions output by the instruction generation model. Based on the vehicle driving commands, control the vehicle's movement.

8. The method according to claim 7, characterized in that, The vehicle data, road constraint data, and driving style are input into the instruction generation model. Based on the instruction generation model, the vehicle data, road constraint data, and driving style are processed to obtain the vehicle driving instructions output by the instruction generation model, including: The vehicle data, the road constraint data, and the driving style are input into the large language model of the instruction generation model for processing to obtain macro instructions; wherein, the macro instructions represent high-level instructions for the vehicle to follow in terms of driving intentions and behavioral decisions within a future preset time window; The macro-level instructions, vehicle data, road constraint data, and driving style are input into the policy gradient module of the instruction generation model for processing to obtain vehicle driving instructions.

9. A training device for a driving style recognition model, characterized in that, include: The acquisition module is used to acquire vehicle data and road constraint data of the vehicle's movement; wherein, the vehicle data represents multimodal time-series data of the vehicle's motion state and driving operation; the road constraint data represents the constraint information of the road geometry and traffic rules on the vehicle's movement. The processing module is used to perform data fusion processing on the vehicle data and the road constraint data to obtain fusion features; The training module is used to input the fused features into the initial intelligent model, process the fused features based on the clustering module and the adversarial imitation module of the initial intelligent model, and optimize the parameters of the initial intelligent model through a loss function to obtain a driving style recognition model; wherein, the driving style recognition model is used to process vehicle data and road constraint data to obtain driving style; the driving style is used to generate vehicle driving commands; the vehicle driving commands are used to control vehicle driving.

10. A control device for an intelligent driving vehicle, characterized in that, include: The acquisition module is used to acquire vehicle data and road constraint data. The recognition module is used to input the vehicle data and the road constraint data into the driving style recognition model, and process the vehicle data and the road constraint data based on the driving style recognition model to obtain the driving style output by the driving style recognition model; wherein, the driving style recognition model is trained by the training device for the driving style recognition model as described in any one of claims 1-6; The generation module is used to input the vehicle data, the road constraint data, and the driving style into the instruction generation model, and process the vehicle data, the road constraint data, and the driving style based on the instruction generation model to obtain the vehicle driving instructions output by the instruction generation model. The control module is used to control the vehicle's movement based on the vehicle driving commands.

11. A vehicle, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the training method as described in any one of claims 1-6, or the method as described in any one of claims 7-8.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the training method as described in any one of claims 1-6, or the method as described in any one of claims 7-8.

13. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the training method as described in any one of claims 1-6, or the method as described in any one of claims 7-8.

Images