Intelligent driving experience generalization optimization method across scenes

Abstract

The application discloses a cross-scene intelligent driving experience generalization optimization method, relates to the technical field of automatic driving, obtains scene data corresponding to multiple driving scenes, performs feature extraction on the scene data, and obtains driving scene feature information containing scene topology features, vehicle dynamics features and driving behavior features; a scene feature embedding space is constructed based on the driving scene feature information, different driving scenes are mapped to feature vectors in the scene feature embedding space, and the similarity relationship between different driving scenes is determined according to the feature vectors; and according to the similarity relationship between different driving scenes, the control parameter optimization information corresponding to the historical driving scenes is associated and analyzed. The application realizes cross-scene optimization knowledge migration and experience reuse by constructing a driving scene feature embedding space and an optimization knowledge base, quickly generates a new scene optimization scheme, reduces the demand for long-tail scene data, and improves the intelligent driving experience optimization efficiency and scene adaptability.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, specifically to a method for generalizing and optimizing intelligent driving experience across different scenarios. Background Technology

[0002] Autonomous driving systems face complex and ever-changing road environments and numerous unstructured driving scenarios during actual operation. These scenarios typically differ in vehicle type, motion state, interaction relationships, environmental conditions, and driving behavior requirements. To improve the safety, comfort, and user experience of intelligent driving systems, existing technologies usually require data collection, scenario analysis, control strategy optimization, and performance verification for specific driving scenarios. For example, for vehicle entry scenarios, existing autonomous driving optimization processes typically require collecting driving data under different vehicle types, entry speeds, and relative distances. By analyzing vehicle motion states, driver perception indicators, and the changing patterns of control parameters, key parameters in the autonomous driving control strategy are adjusted to obtain an optimized solution that meets user experience requirements.

[0003] In practical applications, autonomous vehicles may encounter numerous and complex driving scenarios, such as trucks cutting in, buses cutting in, non-motorized vehicles suddenly entering the driving area, and target vehicle interactions in special road environments. Although there are certain similarities and correlations between different scenarios, traditional optimization methods usually model and adjust parameters independently for a single scenario, making it difficult to fully utilize the optimization experience accumulated in existing scenarios.

[0004] Existing technologies typically employ a closed-loop optimization approach focused on a single driving scenario. This involves separate processes for data collection, scenario modeling, driving experience evaluation, control parameter adjustment, and real-vehicle verification for each specific scenario. Under this approach, each new driving scenario requires the creation of new data samples and optimization processes. For example, control parameter optimization experience gained for a truck cutting in is usually only applicable to that specific scenario. When facing other similar scenarios such as a bus cutting in or a bicycle suddenly darting out, it is still necessary to collect a large amount of data and retrain the model or adjust the parameters. Furthermore, due to the vast number of driving scenarios in real-world road environments, and the low frequency of some long-tail scenarios, the number of available effective samples is limited, making it difficult to support a comprehensive data analysis and optimization process.

[0005] Therefore, existing technologies suffer from knowledge isolation between scenarios, making it impossible to establish relationships between different driving scenarios or to transfer optimization experience from existing scenarios to new similar scenarios. This results in repeated investment of large amounts of data collection and testing costs in the process of optimizing autonomous driving experience, leading to low optimization efficiency and difficulty in meeting the needs of autonomous driving systems for rapid generalization and continuous optimization in complex open environments.

[0006] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this application is to provide a generalized optimization method for intelligent driving experience across different scenarios, in order to solve the problems mentioned in the background technology.

[0008] To achieve the above objectives, this application provides the following technical solution: a cross-scenario intelligent driving experience generalization optimization method, comprising the following steps: Acquire scene data corresponding to multiple driving scenarios, extract features from the scene data, and obtain driving scene feature information including scene topology features, vehicle dynamics features, and driving behavior features; A scene feature embedding space is constructed based on driving scene feature information. Different driving scenes are mapped to feature vectors in the scene feature embedding space, and the similarity relationship between different driving scenes is determined based on the feature vectors. Based on the similarity between different driving scenarios, the control parameter optimization information corresponding to historical driving scenarios is correlated and analyzed to establish a mapping relationship between scenario feature information and driving experience optimization parameters, and to construct a cross-scenario optimization knowledge transfer model. For a new driving scenario to be optimized, obtain the new scenario feature information corresponding to the new driving scenario, input the new scenario feature information into the cross-scenario optimization knowledge transfer model, and generate the initial experience optimization parameters corresponding to the new driving scenario based on the optimization knowledge of similar driving scenarios. Collect actual driving data corresponding to new driving scenarios, and use the actual driving data to adjust the initial experience optimization parameters to obtain a driving experience optimization scheme suitable for new driving scenarios.

[0009] Driving scenario feature information includes scenario topology features, vehicle dynamics features, and driving behavior features. Scenario topology features describe the positional and interactive relationships between the autonomous vehicle and surrounding traffic participants. Vehicle dynamics features describe changes in speed, acceleration, and trajectory during vehicle movement. Driving behavior features describe the vehicle control response process, driving experience evaluation information, and behavioral change patterns during vehicle interaction. By fusing different types of features, a multi-dimensional feature vector is generated to characterize the overall state of the driving scenario.

[0010] The scene feature embedding space is obtained by transforming driving scene data through an encoding network. The encoding network includes an autoencoder structure, which compresses and encodes vehicle kinematic sequences, environmental information, and interactive behavior information in the driving scene, converting the original high-dimensional scene data into a low-dimensional latent feature vector, and retaining the key correlation features between different driving scenes that affect the optimization results of driving experience. Thus, the low-dimensional latent feature vector is used to represent the intrinsic relationship between driving scenes.

[0011] When determining the similarity between different driving scenarios, the distance between the feature vectors corresponding to different driving scenarios is calculated. The distance calculation method includes at least one of Euclidean distance calculation, cosine distance calculation, and Mahalanobis distance calculation. By analyzing the spatial positional relationship between the feature vectors of different driving scenarios, the degree of correlation between historical driving scenarios and driving scenarios to be optimized is determined, and similar driving scenarios corresponding to driving scenarios to be optimized are selected based on the degree of correlation.

[0012] The cross-scenario optimization knowledge transfer model is built through machine learning or deep learning models. The input information includes driving scenario feature vectors, and the output information includes driving experience optimization parameters, including deceleration sensitive boundaries, safe distance thresholds, and braking response parameters. Through the training process, the model learns the correlation between changes in different driving scenario features and the optimization results of control parameters, transforming the fixed optimization parameters originally set for a single driving scenario into dynamic optimization parameters that adjust with changes in scenario features.

[0013] When generating initial experience optimization parameters for a new driving scenario, multiple target scenarios with similar characteristics to the new driving scenario are retrieved from the historical driving scenario knowledge base, and the control parameter optimization information corresponding to the target scenarios is obtained. By analyzing the feature distance relationship between multiple target scenarios and the new driving scenario, the optimization experience corresponding to different target scenarios is weighted and assigned. The initial experience optimization parameters for the new driving scenario are generated based on the weighted calculation results.

[0014] Actual driving data includes vehicle status data, target object motion data, and driving experience evaluation data. Vehicle status data includes vehicle speed, vehicle acceleration, and vehicle trajectory. Target object motion data includes changes in target object speed, relative distance, and motion trend. Driving experience evaluation data includes braking comfort evaluation, safety risk evaluation, and occupant comfort evaluation. By using actual driving data, the initial experience optimization parameters are corrected to adapt the optimization parameters to the actual driving needs corresponding to the new driving scenario.

[0015] The driving scenario optimization knowledge base stores scenario feature information, control parameter optimization information, and driving experience evaluation results corresponding to multiple historical driving scenarios. These historical driving scenarios include vehicle entry scenarios, vehicle following scenarios, obstacle avoidance scenarios, and road merging scenarios. By continuously adding validated new driving scenario optimization data, the scenario associations and optimization parameter information in the driving scenario optimization knowledge base are updated, enabling the continuous accumulation of driving experience optimization knowledge.

[0016] When optimizing the driving experience for vehicle entry scenarios, the driving scenario feature information includes the target vehicle type, the target vehicle speed, the vehicle's speed, the relative distance between the target vehicle and the vehicle, and the movement change trend of the target vehicle during the entry process. By analyzing the correlation between the above feature information and the deceleration sensitive boundary, the deceleration control parameters corresponding to different vehicle type entry scenarios are predicted using a cross-scenario optimization knowledge transfer model, thereby realizing the transfer of optimization experience between vehicle entry scenarios.

[0017] When adapting parameters for the cross-scenario optimization knowledge transfer model, a small number of new driving scenario samples are used to adjust the model. The scene feature association rules learned in the historical driving scenarios are retained. At the same time, the newly added driving data is used to correct the optimization parameters related to the current driving scenario, so that the cross-scenario optimization knowledge transfer model can adapt to the driving experience optimization needs of different vehicle types, different motion states and different traffic environments.

[0018] The technical effects and advantages provided by this application in the above technical solution are as follows: This application constructs a driving scenario feature embedding space, transforming different driving scenarios into feature representations capable of similarity analysis. Based on the relationships between scenario features, it achieves optimized knowledge transfer, enabling the intelligent driving experience optimization process to derive optimization parameters from optimized scenarios to similar unknown scenarios. By leveraging control parameter optimization experience accumulated from historical driving scenarios, initial optimization schemes for new scenarios can be quickly generated, reducing the need for repetitive data collection, parameter analysis, and optimization verification for each new driving scenario. This expands the applicability of intelligent driving experience optimization schemes and improves scenario coverage in complex driving environments.

[0019] This application establishes an optimization knowledge base based on scene feature information, continuously accumulating optimization results, control parameters, and scene relationships across different driving scenarios. This allows driving experience optimization knowledge to be continuously reused and updated in subsequent scenario applications. For long-tail driving scenarios with limited data, only adjustments to the initial optimization parameters based on a small amount of actual scenario data are needed to achieve accurate adaptation to the target scenario. This reduces data acquisition costs during the optimization process for long-tail scenarios while improving the continuous learning ability and optimization efficiency of the autonomous driving system when facing new scenarios. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1 This is a flowchart of the method in this application. Detailed Implementation

[0022] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0023] like Figure 1 As shown, this application provides a cross-scenario intelligent driving experience generalization optimization method, including the following steps: Acquire scene data corresponding to multiple driving scenarios, extract features from the scene data, and obtain driving scene feature information including scene topology features, vehicle dynamics features, and driving behavior features; A scene feature embedding space is constructed based on driving scene feature information. Different driving scenes are mapped to feature vectors in the scene feature embedding space, and the similarity relationship between different driving scenes is determined based on the feature vectors. Based on the similarity between different driving scenarios, the control parameter optimization information corresponding to historical driving scenarios is correlated and analyzed to establish a mapping relationship between scenario feature information and driving experience optimization parameters, and to construct a cross-scenario optimization knowledge transfer model. For a new driving scenario to be optimized, obtain the new scenario feature information corresponding to the new driving scenario, input the new scenario feature information into the cross-scenario optimization knowledge transfer model, and generate the initial experience optimization parameters corresponding to the new driving scenario based on the optimization knowledge of similar driving scenarios. Collect actual driving data corresponding to new driving scenarios, and use the actual driving data to adjust the initial experience optimization parameters to obtain a driving experience optimization scheme suitable for new driving scenarios.

[0024] Specific Implementation Method 1: In this implementation method, data from multiple historical driving scenarios that have already undergone intelligent driving experience optimization are first collected, organized, and structured for storage, establishing a driving scenario optimization knowledge base for cross-scenario experience optimization. This driving scenario optimization knowledge base is used to store verified driving experience optimization results, control parameter adjustment processes, and corresponding scenario feature information under different driving scenarios, providing a historical experience basis for generating optimization solutions for subsequent unknown driving scenarios. Historical driving scenarios may include, but are not limited to, various common interaction scenarios in autonomous driving processes such as vehicle entry scenarios, vehicle following scenarios, obstacle avoidance scenarios, road merging scenarios, and cross-target interaction scenarios. For example, in vehicle entry scenarios, complete optimization data for typical driving scenarios such as "a sedan entering at high speed from the left" and "an SUV entering at low speed from the right" can be collected, including the target vehicle type, target vehicle motion state, autonomous vehicle response process, driving control parameter adjustment results, and corresponding driving experience evaluation results.

[0025] Specifically, in constructing the driving scenario optimization knowledge base, multi-dimensional scenario data corresponding to each historical driving scenario is collected. Among this, vehicle motion state data describes the dynamic changes of the vehicle itself and surrounding traffic participants during the driving process, including target vehicle type, target vehicle size information, distance changes of the target vehicle relative to the vehicle, speed changes of the target vehicle relative to the vehicle, the vehicle's current speed, acceleration changes, vehicle trajectory, lateral offset, and longitudinal motion. By collecting this vehicle motion state data, the movement patterns of different traffic participants during the interaction process can be accurately reflected, providing a data foundation for subsequent analysis of the similarity between driving scenarios.

[0026] Simultaneously, road environment data corresponding to the driving scenario is collected to describe the external traffic environment in which the vehicle is located. Road environment data can include road type information, road structure information, number of lanes, positional relationships between lanes, information about the vehicle's current lane, information about the target vehicle's current lane, road curvature, traffic flow density, road speed limits, and environmental status information. For example, although both vehicle entry scenarios on highways and vehicle entry scenarios on urban roads involve vehicle entry behavior, differences in road structure, vehicle speed range, and traffic participant density may lead to different driving control strategies. By collecting road environment information, driving scenario characteristics can reflect not only the vehicle's own motion state but also the overall traffic environment in which the driving scenario occurs.

[0027] Furthermore, driving interaction data is collected to describe the interaction between the autonomous vehicle and surrounding traffic participants, as well as changes in the driving experience. Driving interaction data can include the interaction behavior between the target vehicle and the vehicle itself, the relative motion trends between vehicles, and the driver's or passenger's evaluation of the autonomous driving behavior. For example, during a vehicle's approach, it can record whether the vehicle takes deceleration measures, the timing of deceleration triggering, the deceleration change process, changes in the safe distance between the vehicle and the target vehicle, and the passenger's evaluation of braking comfort. By analyzing the driving interaction data, the correlation between vehicle motion changes and driving experience can be established, providing a basis for subsequent control parameter optimization.

[0028] After collecting historical driving scenario data, the raw data corresponding to different driving scenarios are preprocessed and feature extracted, transforming the raw driving scenario data into driving scenario feature information that can be used for computational analysis. Driving scenario feature information can include scenario topology features, vehicle dynamics features, and driving behavior features. Scenario topology features describe the positional and interactive relationships between different traffic participants in the driving scenario, such as the relative positional relationship between the target vehicle and the vehicle itself, the relationship between the vehicle and the road area, lane changes, and the connectivity between traffic participants. Vehicle dynamics features describe the physical changes during vehicle motion, such as vehicle speed, acceleration, deceleration, changes in trajectory, and relative motion trends. Driving behavior features describe the vehicle control behavior and changes in driving experience, such as braking response timing, changes in control intensity, driving comfort, and changes in safety risks.

[0029] After obtaining the driving scene feature information, scene feature embedding technology is further used to map the feature space of different driving scenes. Specifically, each driving scene is represented as a high-dimensional feature vector containing multi-dimensional information, enabling different driving scenes to be represented and compared mathematically. To reduce the data complexity in the high-dimensional feature space, autoencoders, deep neural networks, or other feature dimensionality reduction algorithms can be used to encode the vehicle kinematic sequence, environmental information, and interaction behavior information in the driving scene, transforming the original high-dimensional driving scene features into a low-dimensional latent feature vector. This low-dimensional latent feature vector can retain the key change patterns in the driving scene while reducing the influence of irrelevant information on scene similarity judgment.

[0030] For example, for two historical driving scenarios, "a sedan entering from the left at high speed" and "an SUV entering from the right at low speed," feature information such as target vehicle type, average relative speed, average relative distance, average vehicle speed, and changes in vehicle trajectory can be extracted and converted into corresponding scenario feature vectors using an encoding model. By analyzing the distance relationship between the feature vectors of the two scenarios, the similarity between the two driving scenarios in terms of vehicle interaction behavior, motion state, and control requirements can be determined. When the feature vectors of the two driving scenarios are close in distance, it indicates that the two driving scenarios have a high feature correlation, and optimization experience from historical driving scenarios may be applicable to new similar driving scenarios.

[0031] Furthermore, based on the positional relationships of different driving scenarios in the scene feature embedding space, a similarity evaluation mechanism is established between driving scenarios. Specifically, Euclidean distance, cosine distance, Mahalanobis distance, or other similarity calculation methods can be used to calculate the distance between the feature vectors corresponding to different driving scenarios, thereby obtaining a ranking of the similarity between different driving scenarios. In this way, driving scenarios are no longer distinguished solely by manually defined scene categories, but rather a more accurate scene association relationship is established based on the topological, dynamic, and behavioral information contained within the driving scenarios themselves.

[0032] After establishing the correlation between driving scenarios, the mapping relationship between driving scenario feature information and driving experience optimization parameters is analyzed based on the control parameter optimization information corresponding to historical driving scenarios. In traditional methods, driving experience optimization parameters are usually directly associated with specific scenarios. For example, a fixed deceleration sensitive boundary is set for the "car high-speed entry" scenario, and another set of fixed parameters is set for the "SUV low-speed entry" scenario. Optimization experience across different scenarios is difficult to reuse. In this embodiment, the control parameter optimization information and driving scenario feature information are correlated and modeled, so that the control parameters are no longer independent parameters under a fixed scenario, but rather a dynamically adjusted functional relationship based on changes in driving scenario features.

[0033] Specifically, machine learning or deep learning models can be used to establish a mapping relationship between scene feature information and driving experience optimization parameters. The model takes as input the complete scene feature vector corresponding to a historical driving scenario and outputs the corresponding experience optimization parameters, such as deceleration sensitive boundaries, safe distance thresholds, and braking response parameters. By training on multiple historical driving scenario data, the model can learn the influence of changes in different scene features on the optimization results of control parameters. For example, by analyzing multiple source driving scenarios such as "car high-speed entry" and "SUV low-speed entry," the model can learn the changing patterns between factors such as target vehicle size, vehicle speed, and relative distance and deceleration sensitive boundaries, thereby establishing a predictive relationship from driving scenario features to control parameter optimization results.

[0034] By employing the above methods, a cross-scenario optimization knowledge transfer model is constructed. This allows driving experience optimization knowledge to move beyond a single driving scenario and instead achieve knowledge association and experience transfer between different driving scenarios based on scenario characteristics. When a new driving scenario emerges, this cross-scenario optimization knowledge transfer model can be used to predict suitable experience optimization parameters based on the feature information corresponding to the new driving scenario. This provides initial optimization directions for scenarios with scarce data, reducing the need to re-collect large amounts of data and repeat optimization processes, thus achieving cross-scenario generalization capabilities in the intelligent driving experience optimization process.

[0035] Specific Implementation Method Two: In this implementation method, when an autonomous vehicle encounters a new driving scenario to be optimized during operation, data collection and analysis are first performed on the new driving scenario to obtain scenario data related to the current driving interaction process. Based on a preset data processing flow, feature extraction is performed on the collected scenario data to generate corresponding new driving scenario feature information. The new driving scenario can be a scenario with scarce data in the historical driving scenario knowledge base that has not been fully optimized, or an unknown scenario that has some similarity to existing driving scenarios but differs in vehicle type, motion state, or environmental conditions. For example, in actual road operation, a driving scenario of "a truck cutting in from the left" may occur, while historical optimization data mainly focuses on scenarios such as "a sedan cutting in from the left at high speed" and "an SUV cutting in from the right at low speed." In this case, because the number of effective samples for the truck cutting in scenario is small, it is impossible to directly use traditional methods to complete sufficient data analysis and control parameter optimization. Therefore, it is necessary to utilize optimization experience from existing driving scenarios to achieve rapid generalization.

[0036] Specifically, after acquiring new driving scenario data of "a truck cutting in from the left," key influencing factors in this scenario are extracted, including target vehicle type, target vehicle size information, target vehicle speed, the current speed of the vehicle being driven, the relative distance between the target vehicle and the vehicle being driven, the direction of the target vehicle's entry, vehicle movement trends, and dynamic changes during vehicle interaction. Among these, the target vehicle type describes the category difference of the current interactive object, such as sedan, SUV, truck, bus, or other traffic participants; the target vehicle speed reflects the movement state and potential risk changes of the target object; the vehicle's speed represents the current driving state of the autonomous vehicle; the relative distance between vehicles describes the spatial relationship between them; and the movement trend describes the behavioral changes of the target vehicle, such as approaching, moving away, accelerating, decelerating, or cutting in laterally. By extracting the above multi-dimensional information, a new scenario feature vector that describes the inherent characteristics of the current driving scenario can be generated.

[0037] Furthermore, the obtained new scene feature vectors are input into the already constructed scene feature embedding space. Spatial matching is performed with the feature vectors corresponding to historical driving scenes to determine the degree of correlation between the new and historical driving scenes. Specifically, the scene feature information of the new driving scene is converted using the same data encoding method as the historical driving scene, enabling the new driving scene to be represented in the same feature space. Subsequently, the distance relationship between the feature vectors of the new and historical driving scenes is calculated, for example, using Euclidean distance, cosine distance, or other similarity calculation methods, to quantitatively analyze the feature similarity between different driving scenes. When the distance between two driving scenes in the scene feature embedding space is small, it indicates that the two driving scenes have high similarity in terms of vehicle motion patterns, interaction relationships, and control requirements. Therefore, the experience optimization knowledge in historical driving scenes can serve as an important reference for the new driving scene.

[0038] For example, when performing feature space analysis on the scenario of a truck entering from the left, it can be found that this new driving scenario has certain similarities with the scenarios of a sedan entering from the left at high speed and an SUV entering from the right at low speed in the historical knowledge base. Although the target vehicle type, vehicle size, and motion parameters differ in the three driving scenarios, all three belong to the lateral entry interaction behavior of vehicles, and all involve the problem of speed adjustment and safe distance control of the autonomous vehicle after the target vehicle enters the driving area. Therefore, by embedding scene features into the space, it is possible to break through the limitations of traditional matching based on scene names or manual classification methods, and determine more accurate similarity relationships based on the motion features, interaction features, and risk change patterns contained in the driving scenario itself.

[0039] After identifying multiple historical driving scenarios that are highly similar to the new driving scenario, the optimization experience corresponding to these similar driving scenarios is input into a cross-scenario optimization knowledge transfer model for analysis. The optimization experience from historical driving scenarios can include validated control parameter adjustment results, driving experience evaluation results, and control parameter variation patterns. For example, for a vehicle entry scenario, optimization experience can include deceleration sensitive boundaries, safe distance thresholds, and braking response strategies under different target vehicle types, relative speeds, and vehicle distances. The cross-scenario optimization knowledge transfer model comprehensively analyzes the optimization knowledge from multiple similar driving scenarios based on the correlation between the features of the new driving scenario and the features of multiple historical driving scenarios, rather than directly copying the fixed parameters of a single historical scenario.

[0040] Specifically, the cross-scenario optimization knowledge transfer model can assign weights to optimization parameters corresponding to similar driving scenarios based on the feature distance between different historical driving scenarios and the new driving scenario. For example, the scenario of "a sedan cutting in from the left at high speed," which is closer in feature distance to "a truck cutting in from the left," can have a higher weight, while the scenario of "an SUV cutting in from the right at low speed," which is relatively farther in feature distance, can have a lower weight. By performing weighted analysis on optimization experience from multiple historical driving scenarios, initial experience optimization parameters that are more consistent with the characteristics of the new driving scenario can be obtained.

[0041] For example, in vehicle entry scenarios, historical optimization data shows that the target vehicle's size, speed, and relative distance affect the deceleration sensitivity boundary required by the autonomous vehicle. Through cross-scenario optimization knowledge transfer model analysis, the initial deceleration sensitivity boundary for the "truck entering from the left" scenario can be predicted based on characteristics such as the target vehicle's large size, moderate speed, and medium-to-long distance from the vehicle. For instance, the model predicts an initial deceleration sensitivity boundary of 2.2 m / s² for this new driving scenario. This prediction is not derived from replicating parameters from a single historical scenario, but rather from comprehensive reasoning based on causal optimization patterns across multiple similar driving scenarios.

[0042] By employing the above method, when faced with new driving scenarios with limited data or those not yet fully optimized, there is no need to re-collect large amounts of driving data and retrain the optimization model as in traditional methods. Instead, optimization experience from existing driving scenario knowledge bases can be used to quickly generate initial experience optimization solutions. This process reduces the dependence of the new driving scenario optimization process on large amounts of sample data and improves the efficiency of the autonomous driving experience optimization process.

[0043] Furthermore, since the cross-scenario optimization knowledge transfer model establishes an optimization parameter mapping relationship based on scenario feature information, this model is not only applicable to vehicle-entry scenarios but can also be extended to other driving scenarios with similar behavioral patterns. For example, in following scenarios, existing optimization experience with ordinary vehicles can be transferred to large vehicle following scenarios; in obstacle avoidance scenarios, existing experience with avoiding fixed obstacles can be transferred to avoiding dynamic targets. Through similarity judgment based on scenario feature embedding space and optimization knowledge transfer mechanism, the autonomous driving experience optimization scheme can be rapidly generalized from known scenarios to unknown similar scenarios, improving the adaptability and continuous optimization capability of the autonomous driving system in the face of complex open road environments.

[0044] Specific Implementation Method 3: In this implementation method, after using a cross-scenario optimization knowledge transfer model to perform feature analysis on the new driving scenario and generate corresponding initial experience optimization parameters, a small amount of data collection and parameter fine-tuning process is further carried out for the new driving scenario to achieve precise adaptation of the optimization scheme from the general transfer result to the target driving scenario. This implementation method is mainly applied to long-tail driving scenarios in real road environments where the amount of data is small, the frequency of occurrence is low, but experience optimization is still required. By first using optimization knowledge from historical driving scenarios to generate initial optimization results, and then using a small amount of target scenario data for rapid correction, the problem of traditional technologies requiring the re-collection of large amounts of data, retraining of the model, and repeated implementation of the complete optimization process for each new scenario can be avoided.

[0045] Specifically, when an autonomous vehicle encounters a new driving scenario to be optimized, it first utilizes the already established cross-scenario optimization knowledge transfer model to predict initial optimization parameters for the new driving scenario. For example, in practical applications, a new driving scenario of "a truck cutting in from the left" may arise. Since the historical driving scenario knowledge base mainly accumulates optimization experience for vehicle entry scenarios such as "a sedan cutting in from the left at high speed" and "an SUV cutting in from the right at low speed," the data sample size for truck entry scenarios is relatively small. Therefore, it is impossible to directly obtain accurate driving experience optimization parameters through traditional methods. In this case, feature analysis of the "truck cutting in from the left" scenario is performed through the aforementioned scenario feature embedding space to obtain feature information such as target vehicle type, target vehicle speed, vehicle speed, relative distance between vehicles, changes in the entry trajectory, vehicle interaction process, and driving experience-related indicators. This new scenario feature information is then input into the cross-scenario optimization knowledge transfer model.

[0046] The cross-scenario optimization knowledge transfer model performs transfer analysis on the optimization experience of control parameters in existing similar driving scenarios based on the correlation between the feature information of new driving scenarios and the optimization knowledge of historical driving scenarios. For example, the model can use the target vehicle speed change, vehicle relative distance change, and deceleration control law in the historical scenario of "a sedan cutting in from the left at high speed" and refer to the vehicle size influencing factors and driving experience evaluation results in the scenario of "an SUV cutting in from the right at low speed" to comprehensively reason about the optimization knowledge in different historical scenarios and predict the initial experience optimization parameters corresponding to the scenario of "a truck cutting in from the left".

[0047] Taking deceleration-sensitive boundary optimization as an example, traditional optimization methods typically require collecting a large amount of road test data for truck entry scenarios. This involves analyzing vehicle safety, occupant comfort, and driving naturalness under different braking intensities to gradually determine the optimal deceleration range. In this implementation, however, by utilizing optimization patterns from similar driving scenarios, an initial deceleration-sensitive boundary applicable to truck entry scenarios can be directly predicted. For instance, the cross-scenario optimization knowledge transfer model, after analyzing factors such as vehicle type, vehicle mass characteristics, speed, and relative distance, predicts the initial deceleration-sensitive boundary for the current "truck entering from the left" scenario to be 2.2 m / s². This initial parameter, serving as the basis for subsequent precise optimization, can significantly shorten the time cycle from discovery to optimization of new driving scenarios.

[0048] After obtaining the initial experience optimization parameters, a small amount of actual driving data is collected for new driving scenarios to fine-tune and adapt the initial optimization parameters. Specific data collection methods may include real-world road testing, closed-course testing, autonomous driving simulation testing, or data playback testing based on historical operating data. During the testing process, vehicle status data and driving experience evaluation data corresponding to the new driving scenarios are collected, including vehicle braking response process, vehicle deceleration change curve, distance change between the vehicle and the target object, collision risk indicators, vehicle stability indicators, and occupant comfort evaluation results.

[0049] For example, in the scenario of a truck cutting in from the left, only a small number of actual driving samples are needed, such as 10 cases of this type of scenario data, to adjust the initial optimization parameters output by the cross-scenario optimization knowledge transfer model. For each test, it is analyzed whether the current initial deceleration sensitive boundary can meet the comprehensive requirements of driving safety and comfort. If the test results show that the predicted deceleration sensitive boundary of 2.2 m / s² leads to premature braking response or insufficient comfort, the corresponding parameters are reduced based on the actual test data; if the test results show insufficient braking response, the corresponding parameter range is increased. By iteratively adjusting the model parameters or optimization parameters with a small amount of target scenario data, the prediction results gradually approach the actual needs of the current new driving scenario.

[0050] Furthermore, during parameter fine-tuning, transfer learning can be used to adapt and update the cross-scenario optimization knowledge transfer model. Specifically, the general optimization rules learned from historical driving scenarios are retained, while only a small amount of new driving scenario data is used to adjust the parameters in the model that are relevant to the current scenario. This allows the model to maintain both its generalization ability for existing scenarios and its adaptability to new scenarios. For example, after analyzing 10 test data cases of "trucks cutting in from the left," the initial deceleration sensitive boundary predicted by the model can be adjusted from 2.2 m / s² to a more accurate 2.0 m / s², thereby obtaining optimization results that meet the actual driving experience requirements of truck cutting in scenarios.

[0051] After adjusting the parameters for the new driving scenario, the optimized driving experience solution is validated. The validation process includes multiple aspects such as safety validation, comfort validation, and stability validation. Safety validation determines whether the optimized control parameters can guarantee the vehicle's safe response in complex traffic environments; comfort validation evaluates whether the vehicle's acceleration and deceleration processes meet occupant experience requirements; and stability validation determines whether the optimized parameters are applicable to similar driving scenarios with different speeds, distances, and traffic conditions. The validated optimized parameters are saved as the final driving experience optimization solution for this new driving scenario.

[0052] Furthermore, validated new driving scenario data, scenario feature information, and corresponding optimization parameters are re-stored in the driving scenario optimization knowledge base, enabling continuous accumulation and iterative updates of driving scenario knowledge. For example, after optimizing the "truck cutting in from the left" scenario, the corresponding scenario feature vector, deceleration sensitive boundary, safe distance parameters, and driving experience evaluation results can all be added to the knowledge base as new optimization knowledge. When subsequent scenarios such as buses cutting in, large engineering vehicles cutting in, or other driving scenarios with similar motion characteristics occur, the newly added knowledge can be used to further improve the prediction accuracy of the cross-scenario optimization knowledge transfer model.

[0053] This implementation method utilizes scene feature embedding technology to establish feature associations between different driving scenarios. It retrieves transferable optimization experience from historical driving scenarios through scene similarity calculation, and then uses transfer learning methods to establish a mapping relationship between scene features and driving experience optimization parameters. Finally, it combines a small amount of new scene data to complete rapid fine-tuning, effectively generalizing the autonomous driving experience optimization scheme from known driving scenarios to unknown similar driving scenarios. Compared to traditional methods that repeatedly collect large amounts of data and re-optimize for a single scenario, this implementation method reduces the data requirements in the optimization process of long-tail driving scenarios, improves optimization efficiency, and enhances the continuous learning and experience optimization capabilities of the intelligent driving system when facing complex open road environments.

[0054] The foregoing has only described certain exemplary embodiments of this application by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of this application. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of this application.

Claims

1. A cross-scenario intelligent driving experience generalization optimization method, characterized in that, Includes the following steps: Acquire scene data corresponding to multiple driving scenarios, extract features from the scene data, and obtain driving scene feature information including scene topology features, vehicle dynamics features, and driving behavior features; A scene feature embedding space is constructed based on driving scene feature information. Different driving scenes are mapped to feature vectors in the scene feature embedding space, and the similarity relationship between different driving scenes is determined based on the feature vectors. Based on the similarity between different driving scenarios, the control parameter optimization information corresponding to historical driving scenarios is correlated and analyzed to establish a mapping relationship between scenario feature information and driving experience optimization parameters, and to construct a cross-scenario optimization knowledge transfer model. For a new driving scenario to be optimized, obtain the new scenario feature information corresponding to the new driving scenario, input the new scenario feature information into the cross-scenario optimization knowledge transfer model, and generate the initial experience optimization parameters corresponding to the new driving scenario based on the optimization knowledge of similar driving scenarios. Collect actual driving data corresponding to new driving scenarios, and use the actual driving data to adjust the initial experience optimization parameters to obtain a driving experience optimization scheme suitable for new driving scenarios.

2. The cross-scenario intelligent driving experience generalization optimization method according to claim 1, characterized in that, Driving scenario feature information includes scenario topology features, vehicle dynamics features, and driving behavior features. Scenario topology features describe the positional and interactive relationships between the autonomous vehicle and surrounding traffic participants. Vehicle dynamics features describe changes in speed, acceleration, and trajectory during vehicle movement. Driving behavior features describe the vehicle control response process, driving experience evaluation information, and behavioral change patterns during vehicle interaction. By fusing different types of features, a multi-dimensional feature vector is generated to characterize the overall state of the driving scenario.

3. The cross-scenario intelligent driving experience generalization optimization method according to claim 1, characterized in that, The scene feature embedding space is obtained by transforming driving scene data through an encoding network. The encoding network includes an autoencoder structure, which compresses and encodes vehicle kinematic sequences, environmental information, and interactive behavior information in the driving scene, converting the original high-dimensional scene data into a low-dimensional latent feature vector, and retaining the key correlation features between different driving scenes that affect the optimization results of driving experience.

4. The cross-scenario intelligent driving experience generalization optimization method according to claim 1, characterized in that, When determining the similarity between different driving scenarios, the distance between the feature vectors corresponding to different driving scenarios is calculated. The distance calculation method includes at least one of Euclidean distance calculation, cosine distance calculation, and Mahalanobis distance calculation. By analyzing the spatial positional relationship between the feature vectors of different driving scenarios, the degree of correlation between historical driving scenarios and driving scenarios to be optimized is determined, and similar driving scenarios corresponding to driving scenarios to be optimized are selected based on the degree of correlation.

5. The cross-scenario intelligent driving experience generalization optimization method according to claim 1, characterized in that, The cross-scenario optimization knowledge transfer model is built through machine learning or deep learning models. The input information includes driving scenario feature vectors, and the output information includes driving experience optimization parameters, including deceleration sensitive boundaries, safe distance thresholds, and braking response parameters. Through the training process, the model learns the correlation between changes in different driving scenario features and the optimization results of control parameters, transforming the fixed optimization parameters originally set for a single driving scenario into dynamic optimization parameters that adjust with changes in scenario features.

6. The cross-scenario intelligent driving experience generalization optimization method according to claim 5, characterized in that, When generating initial experience optimization parameters for a new driving scenario, multiple target scenarios with similar characteristics to the new driving scenario are retrieved from the historical driving scenario knowledge base, and the control parameter optimization information corresponding to the target scenarios is obtained. By analyzing the feature distance relationship between multiple target scenarios and the new driving scenario, the optimization experience corresponding to different target scenarios is weighted and assigned. The initial experience optimization parameters for the new driving scenario are generated based on the weighted calculation results.

7. The cross-scenario intelligent driving experience generalization optimization method according to claim 6, characterized in that, Actual driving data includes vehicle status data, target object motion data, and driving experience evaluation data. Vehicle status data includes vehicle speed, vehicle acceleration, and vehicle trajectory. Target object motion data includes changes in target object speed, relative distance, and motion trend. Driving experience evaluation data includes braking comfort evaluation, safety risk evaluation, and occupant comfort evaluation. The initial experience optimization parameters are corrected by using actual driving data.

8. The cross-scenario intelligent driving experience generalization optimization method according to claim 1, characterized in that, The driving scenario optimization knowledge base stores scenario feature information, control parameter optimization information, and driving experience evaluation results corresponding to multiple historical driving scenarios. These historical driving scenarios include vehicle entry scenarios, vehicle following scenarios, obstacle avoidance scenarios, and road merging scenarios. By continuously adding validated new driving scenario optimization data, the scenario associations and optimization parameter information in the driving scenario optimization knowledge base are updated.

9. The cross-scenario intelligent driving experience generalization optimization method according to claim 1, characterized in that, When optimizing the driving experience for vehicle-entry scenarios, the driving scenario feature information includes the target vehicle type, the target vehicle speed, the vehicle's speed, the relative distance between the target vehicle and the vehicle, and the motion change trend of the target vehicle during the entry process. By analyzing the correlation between the above feature information and the deceleration sensitive boundary, the deceleration control parameters corresponding to different vehicle type entry scenarios are predicted using a cross-scenario optimization knowledge transfer model.

10. The cross-scenario intelligent driving experience generalization optimization method according to claim 1, characterized in that, When adapting parameters for the cross-scenario optimization knowledge transfer model, a small number of new driving scenario samples are used to adjust the model. The scene feature association rules learned in the historical driving scenarios are retained. At the same time, the newly added driving data is used to correct the optimization parameters related to the current driving scenario, so that the cross-scenario optimization knowledge transfer model can adapt to the driving experience optimization needs of different vehicle types, different motion states and different traffic environments.

Images