Adaptive vehicle driving control method, system and computer readable storage medium based on scene feature matching

Abstract

This invention discloses an adaptive vehicle driving control method, system, and computer-readable storage medium based on scene feature matching. The method includes the following steps: real-time acquisition of traffic environment data around the vehicle; recording of the driver's active driving data in various driving scenarios; extraction of key environmental feature parameters in the driving scenarios to construct a scene-behavior mapping database; in autonomous driving mode, real-time calculation of the relative motion state vector of the current traffic environment and similarity matching with historical relative motion state vectors in the scene-behavior mapping database; if the similarity exceeds a preset confidence threshold, calling the active driving operation data corresponding to the historical vector to generate an initial adaptive control strategy and performing safety boundary verification; if successful, sending instructions to the vehicle actuators; if unsuccessful, correcting the initial adaptive control strategy or switching to a default safety control strategy.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle driving technology, and in particular to an adaptive vehicle driving control method, system, and computer-readable storage medium based on scene feature matching. Background Technology

[0002] With the rapid development of artificial intelligence, sensor technology, and vehicle-to-everything (V2X) technology, autonomous driving technology has evolved from assisted driving (Level 2) to conditional autonomous driving (Level 3) and higher levels. In Level 3 and above autonomous driving modes, the autonomous driving system completely takes over the longitudinal control (acceleration, deceleration) and lateral control (steering, lane changing) of the vehicle within a specific Design Operating Domain (ODD). The driver only needs to take over when requested by the system. This transformation aims to free up driver attention and improve traffic efficiency and safety.

[0003] However, in practical application and promotion, high-level autonomous driving systems face the serious challenge of a poor "human-machine co-driving" experience. Existing autonomous driving control algorithms typically employ a fixed set of parameters or a limited number of modes (such as "economy mode" or "sport mode"). This "one-size-fits-all" control approach has the following significant drawbacks: it lacks the ability to adapt to the individual driving habits of drivers. Different drivers develop drastically different driving styles due to varying driving experience and psychological expectations. Some drivers prefer aggressive driving, tending to respond quickly, accelerate rapidly, decelerate sharply, and frequently overtake and change lanes; while others prefer conservative driving, seeking smoothness and comfort, tending to accelerate and decelerate gently and minimize lane changes. Because existing autonomous driving systems have fixed control parameters, they cannot identify and mimic the style characteristics of specific drivers. Furthermore, existing self-learning or adaptive technologies are often based solely on simple statistical laws (such as average following distance and average acceleration). This leads to scenario generalization errors when the system learns driver habits, meaning it incorrectly applies certain driving habits in inappropriate scenarios, not only reducing passenger comfort but also potentially causing safety hazards. Therefore, how to break through the limitations of traditional fixed-parameter control logic and design a system that can deeply perceive specific traffic scenarios and accurately learn drivers' personalized habits has become an urgent technical problem to be solved in the field of autonomous driving technology. Summary of the Invention

[0004] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes a vehicle driving control method that integrates the driver's individual habits into autonomous driving decisions while ensuring driving safety. This achieves a leap from "mechanical rule-based driving" to "humanized adaptive driving," effectively improving the driving experience.

[0005] The present invention further proposes a vehicle driving control system.

[0006] The present invention further proposes a computer-readable storage medium.

[0007] According to a first aspect of the present invention, an adaptive vehicle driving control method based on scene feature matching includes the following steps: Step S1: Real-time acquisition of traffic environment data around the vehicle using an onboard sensor array, construction of a three-dimensional traffic environment model, and prediction of the movement trajectories of surrounding traffic participants; Step S2: In manual driving mode, recording the driver's active driving operation data under various preset driving scenarios, wherein the active driving operation data includes at least steering wheel angle and angular velocity, lateral acceleration, longitudinal acceleration, yaw rate, accelerator pedal depth, and brake pedal depth; Step S3: Extracting key environmental feature parameters under the preset driving scenarios and constructing a scene-behavior mapping database; wherein the key environmental feature parameters are represented as relative motion state vectors, the vectors containing... This includes the relative speed, relative distance, and relative acceleration between the vehicle and surrounding traffic participants; Step S4: In autonomous driving mode, calculate the relative motion state vector of the current traffic environment in real time and perform similarity matching with the historical relative motion state vectors in the scene-behavior mapping database; Step S5: If the similarity between the current relative motion state vector and a certain historical vector in the database exceeds a preset confidence threshold, then call the active driving operation data corresponding to the historical vector to generate an initial adaptive control strategy; Step S6: Perform safety boundary verification on the initial adaptive control strategy. If the verification passes, convert it into a final control command and send it to the vehicle actuator; if the verification fails, then modify the initial adaptive control strategy or switch to the default safety control strategy.

[0008] Therefore, the method of this invention strengthens the perception foundation by constructing a three-dimensional environment model and predicting trajectories, and establishes a "scene-behavior" mapping library using human driving data. In autonomous driving, the system matches the current environment with historical scenarios in real time, directly reuses the driver's personalized operating strategies, and is supplemented by strict safety boundary verification. If the verification passes, human-like control is executed; if it fails, it is immediately corrected or switched to a safe mode. This solution successfully integrates the driver's personalized habits into autonomous driving decision-making while ensuring driving safety, achieving a leap from "mechanical rule-based driving" to "humanized adaptive driving," and effectively improving the driving experience.

[0009] According to some embodiments of the present invention, the traffic environment data collected in step S1 includes at least: the position, speed, and acceleration information of surrounding vehicles; the position and movement trajectory information of pedestrians; the geometric shape and type information of lane lines; traffic sign recognition information; and traffic light status information; in step S3, based on the spatial positional relationship and movement trend between the vehicle and surrounding traffic participants, the relative motion state vector is divided into at least three types of scene feature vectors: the first type is a longitudinal following scene vector, whose dimensions include at least: the relative longitudinal speed, relative longitudinal distance, and relative longitudinal acceleration between the vehicle and the vehicle in front; the second type is a lateral parallel scene vector, whose dimensions include at least: the relative lateral speed, relative lateral distance, and relative longitudinal speed between the vehicle and vehicles traveling in the same direction in adjacent lanes; the third type is a trajectory intersection scene vector, whose dimensions include at least: the relative lateral speed, relative longitudinal speed, relative distance, and relative acceleration between the vehicle and vehicles intersecting at the predicted intersection point.

[0010] According to some embodiments of the present invention, the similarity matching in step S4 specifically includes: calculating the distance metric between the current real-time scene feature vector and the historical scene feature vector in the database; setting a dynamic tolerance threshold, and determining that the scene matching is successful when the distance metric is less than or equal to the threshold; wherein the dynamic tolerance threshold is dynamically adjusted at least according to the current road adhesion coefficient or weather conditions.

[0011] According to some embodiments of the present invention, in the longitudinal following scenario, step S5 specifically includes: when it is detected that the relative longitudinal speed, relative longitudinal distance, and relative longitudinal acceleration of the vehicle and the vehicle in front all fall within the preset threshold range of the driver's historical active driving data, extracting the accelerator pedal opening or brake pedal pressure value of the driver at that historical moment; mapping the extracted accelerator pedal opening or brake pedal pressure value to the vehicle's desired longitudinal acceleration, as the core parameter of the initial adaptive control strategy.

[0012] According to some embodiments of the present invention, in the lateral parallel scenario or trajectory intersection scenario, step S5 specifically includes: when the lateral relative speed, lateral relative distance, lateral relative acceleration, longitudinal relative speed, longitudinal relative distance, and longitudinal relative acceleration of the adjacent lane traffic participant and the vehicle all fall within the preset threshold range of the driver's historical active driving data, extracting the driver's steering wheel angle and angular velocity at that historical moment; mapping the extracted steering wheel angle and angular velocity to the vehicle's desired lateral displacement trajectory or yaw rate, as the core parameters of the initial adaptive control strategy.

[0013] According to some embodiments of the present invention, the safety boundary verification in step S6 specifically includes: obtaining the current dynamic limit parameters of the vehicle and the regulatory restriction parameters under the current traffic scenario, and constructing a safe feasible region; determining whether the control parameters in the initial adaptive control strategy are within the safe feasible region; if they are within the safe feasible region, they are directly output; if they exceed the safe feasible region, a projection algorithm is used to truncate the control parameters to the boundary value of the safe feasible region, or the control parameters are output after smoothing and filtering.

[0014] According to some embodiments of the present invention, an online iterative update step is further included: during the operation of the autonomous driving mode, the driver's takeover behavior is monitored; if driver takeover is detected, the traffic environment feature vector at the moment of takeover and the driver's actual operation data after takeover are recorded to form a new sample pair; the new sample pair is stored in the scenario-behavior mapping database, and the historical active driving data under the scenario is updated using a time decay weight algorithm; wherein, the time decay weight algorithm assigns a higher weight coefficient to the most recent data to correct the subsequently generated adaptive control strategy.

[0015] According to some embodiments of the present invention, the specific calculation formula of the time decay weight algorithm is as follows: for the i-th historical data under the same scene feature, its weight ;in, The current system time. The record time of this historical data is λ, and the preset forgetting factor is λ. When generating the adaptive control strategy, the operating parameters of multiple historical data are fused using a weighted average method. The calculation formula is as follows:

[0016] According to a second aspect of the present invention, a vehicle driving control system includes: a perception module, configured with at least one of a lidar, a millimeter-wave radar, and a camera, configured to collect traffic environment data around the vehicle to construct a three-dimensional traffic environment model and predict a trajectory; a storage module, configured to store a scene-behavior mapping database, the database storing historical relative motion state vectors constructed by the method and corresponding active driving operation data; a processing module, connected to the perception module and the storage module, configured to execute the steps of the method, including scene similarity matching, control strategy generation, and safety boundary verification; and an execution module, connected to the processing module, configured to drive the vehicle's throttle, braking, and steering systems according to a final control command.

[0017] According to a third aspect of the present invention, a computer-readable storage medium thereon stores a computer program that, when executed by a processor, implements the method described thereon.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a schematic diagram of the vehicle driving control method according to an embodiment of the present invention. Detailed Implementation

[0020] The embodiments of the present invention are described in detail below. The embodiments described with reference to the accompanying drawings are exemplary. The embodiments of the present invention are described in detail below.

[0021] The following is for reference. Figure 1 An adaptive vehicle driving control method based on scene feature matching according to an embodiment of the present invention is described. For example... Figure 1 As shown, the adaptive vehicle driving control method based on scene feature matching according to the first aspect of the present invention includes the following steps: Step S1: Collect real-time traffic environment data around the vehicle using onboard sensor arrays, construct a three-dimensional traffic environment model, and predict the movement trajectories of surrounding traffic participants; Step S2: In manual driving mode, record the driver's active driving operation data under various preset driving scenarios. The active driving operation data includes at least the steering wheel angle and angular velocity, lateral acceleration, longitudinal acceleration, yaw rate, accelerator pedal depth and brake pedal depth. Step S3: Extract key environmental feature parameters under the preset driving scenario and construct a scenario-behavior mapping database; the key environmental feature parameters are represented as relative motion state vectors, which contain the relative speed, relative distance and relative acceleration between the vehicle and surrounding traffic participants; Step S4: In autonomous driving mode, calculate the relative motion state vector of the current traffic environment in real time and perform similarity matching with the historical relative motion state vector in the scene-behavior mapping database; Step S5: If the similarity between the current relative motion state vector and a certain historical vector in the database exceeds a preset confidence threshold, then call the active driving operation data corresponding to the historical vector to generate an initial adaptive control strategy. Step S6: Perform safety boundary verification on the initial adaptive control strategy. If the verification passes, convert it into a final control command and send it to the vehicle actuator. If the verification fails, modify the initial adaptive control strategy or switch to the default safety control strategy.

[0022] Specifically, in this embodiment, the adaptive vehicle driving control method based on scene feature matching first collects real-time traffic environment data around the vehicle using an onboard sensor array. The collected data is then used to construct a three-dimensional traffic environment model, and this model is used to predict the movement trajectories of surrounding traffic participants, thus providing an accurate environmental perception basis for subsequent decision-making. While the vehicle is in manual driving mode, the system continuously records the driver's active driving operation data under various preset driving scenarios. This operation data covers key indicators reflecting the driver's operating habits, specifically including steering wheel angle and angular velocity, lateral acceleration, longitudinal acceleration, yaw rate, accelerator pedal depth, and brake pedal depth, thereby forming raw data samples of the driver's personalized driving style.

[0023] At the same time, the system extracts key environmental feature parameters from preset driving scenarios to build a mapping database between driving scenarios and behaviors. The key environmental feature parameters are specifically represented as relative motion state vectors, which contain the relative speed, relative distance, and relative acceleration between the vehicle and surrounding traffic participants. By associating and storing specific environmental features with the driver's actual operation data, a historical experience database that can be retrieved and learned is established.

[0024] When the vehicle switches to autonomous driving mode, the system calculates the relative motion state vector in real time under the current traffic environment and performs a similarity match between this real-time vector and historical relative motion state vectors stored in the scene behavior mapping database to find the closest historical driving scenario under the current conditions. If the similarity between the current relative motion state vector and a historical vector in the database exceeds a preset confidence threshold, the system directly calls the active driving operation data corresponding to the successfully matched historical vector, and generates an initial adaptive control strategy accordingly, enabling the vehicle's control actions to mimic the driver's operating habits in similar scenarios.

[0025] Finally, the system performs a safety boundary check on the generated initial adaptive control strategy to ensure that the control commands comply with vehicle dynamics limits and traffic safety regulations. If the check passes, the strategy is converted into the final control command and sent to the vehicle actuators to drive the vehicle. If the check fails, it indicates that the initial strategy has a safety risk, and the system immediately corrects the initial adaptive control strategy or directly switches to the default conservative safety control strategy, thereby ensuring the vehicle's driving safety while meeting the driver's personalized expectations.

[0026] Therefore, the method of this invention strengthens the perception foundation by constructing a three-dimensional environment model and predicting trajectories, and establishes a "scene-behavior" mapping library using human driving data. In autonomous driving, the system matches the current environment with historical scenarios in real time, directly reuses the driver's personalized operating strategies, and is supplemented by strict safety boundary verification. If the verification passes, human-like control is executed; if it fails, it is immediately corrected or switched to a safe mode. This solution successfully integrates the driver's personalized habits into autonomous driving decision-making while ensuring driving safety, achieving a leap from "mechanical rule-based driving" to "humanized adaptive driving," and effectively improving the driving experience.

[0027] In some embodiments of the present invention, the traffic environment data collected in step S1 includes at least: the position, speed and acceleration information of surrounding vehicles; the position and movement trajectory information of pedestrians; the geometry and type information of lane lines; traffic sign recognition information; and traffic light status information.

[0028] Specifically, the traffic environment data collected by the vehicle-mounted sensor array covers multi-dimensional dynamic and static information, including the position, speed, and acceleration information of surrounding vehicles to understand the movement status of nearby vehicles, the position and movement trajectory information of pedestrians to predict potential conflict risks, the geometric shape and type information of lane lines to clarify road traffic rules, traffic sign recognition information to obtain speed limits, prohibition and other control requirements, and traffic light status information to respond to intersection passage instructions in real time. These data together constitute the basic input for building a three-dimensional traffic environment model and predicting the movement trajectory of traffic participants, providing solid and comprehensive information support for subsequent accurate prediction of traffic participant movement trajectory, extraction of key scene features, and generation of safe and reliable adaptive control strategies.

[0029] In some embodiments of the present invention, in step S3, based on the spatial positional relationship and movement trend between the vehicle and surrounding traffic participants, the relative motion state vector is divided into at least three types of scene feature vectors: the first type is a longitudinal following scene vector, whose dimensions include at least: the relative longitudinal speed, relative longitudinal distance, and relative longitudinal acceleration between the vehicle and the vehicle in front; the second type is a lateral parallel scene vector, whose dimensions include at least: the relative lateral speed, relative lateral distance, and relative longitudinal speed between the vehicle and vehicles traveling in the same direction in adjacent lanes; the third type is a trajectory intersection scene vector, whose dimensions include at least: the relative lateral speed, relative longitudinal speed, relative distance, and relative acceleration between the vehicle and vehicles intersecting at the predicted intersection point.

[0030] It is understandable that, based on the spatial positional relationship and movement trend between the vehicle and surrounding traffic participants, the relative motion state vector is refined into three core scene feature vectors: For longitudinal following scenarios, a first-class vector is constructed that includes the relative longitudinal velocity, relative longitudinal distance, and relative longitudinal acceleration between the vehicle and the vehicle in front, in order to accurately depict the dynamic game between the vehicles in front and behind. For lateral parallel scenarios, a second type of vector is constructed that includes the relative lateral velocity, relative lateral distance, and relative longitudinal velocity between the vehicle and vehicles traveling in the same direction in adjacent lanes, in order to monitor lateral interaction risks in real time. For trajectory intersection scenarios, a third type of vector is constructed that includes the relative lateral velocity, relative longitudinal velocity, relative distance, and relative acceleration of the vehicle and the intersecting vehicle at the predicted intersection point, in order to quantify the conflict situation at the intersection. The above multi-dimensional scenario segmentation strategy can decouple the complex traffic environment into concrete typical driving conditions, providing structured and standardized feature inputs for subsequent high-confidence similarity matching and personalized strategy reuse in the "scenario-behavior" mapping database.

[0031] In some embodiments of the present invention, the similarity matching in step S4 specifically includes: calculating the distance metric between the current real-time scene feature vector and the historical scene feature vector in the database; setting a dynamic tolerance threshold, and determining that the scene matching is successful when the distance metric is less than or equal to the threshold. The dynamic tolerance threshold is dynamically adjusted at least according to the current road adhesion coefficient or weather conditions.

[0032] Specifically, during the similarity matching process, the distance metric between the current real-time scene feature vector and the historical scene feature vector in the database is first quantitatively calculated to assess their similarity. Then, a dynamic tolerance threshold is introduced as a matching criterion. When the calculated distance metric is less than or equal to this threshold, the scene is considered successfully matched, and the corresponding strategy is reused. It's important to note that this dynamic tolerance threshold is not fixed but can be adaptively adjusted in real-time based on changes in environmental factors such as road adhesion coefficient and weather conditions. For example, it automatically widens or tightens the matching tolerance on low-adhesion surfaces or in adverse weather conditions, thereby ensuring that the accuracy, robustness, and safety of scene matching are effectively guaranteed under various complex environmental conditions.

[0033] The distance metric represents the degree of difference between the current real-time scene feature vector and the historical scene feature vector in the multi-dimensional feature space, and its value directly reflects the similarity between the two. This value can be calculated using algorithms such as Euclidean distance, Mahalanobis distance, or weighted norm. Each dimension of the feature is assigned a different weight coefficient according to its impact on driving safety, ensuring that key safety indicators dominate the similarity assessment. When the calculated distance metric falls within a preset dynamic tolerance threshold, it indicates that the current traffic condition highly matches a historical experience in the database, thereby triggering the migration and reuse of the corresponding control strategy. Each traffic scene can be understood as an abstract vector composed of multiple data dimensions (e.g., an array of relative speed, relative distance, acceleration, etc.), that is, each scene is considered a "point" in a multi-dimensional space. The distance metric calculates the straight-line distance (or path length under a specific rule) between the "current scene point" and the "historical scene point." The smaller the distance metric value, the closer the features of the two scenes are, and the higher the similarity (i.e., the closer the "distance"); the larger the value, the greater the difference between the two scenes, and the lower the similarity.

[0034] The dynamic tolerance threshold is not a fixed constant, but a functional variable that evolves in real time with environmental perception information. Specifically, the system monitors the current road adhesion coefficient and weather conditions in real time: under high adhesion coefficient and good weather conditions, the system adopts a more stringent baseline threshold, requiring the current scene characteristics to be highly consistent with the records in the historical database, to ensure the refinement and comfort of the control strategy; under low adhesion coefficient (such as rain, snow, icy roads) or severe weather (such as heavy fog, strong light interference), considering the increase in sensor measurement noise and the nonlinear changes in vehicle dynamic characteristics, the system automatically adaptively corrects the dynamic tolerance threshold (for example, appropriately relaxing the upper limit of distance measurement, or adjusting the weight ratio of each feature dimension); this dynamic adjustment mechanism effectively avoids matching failures (i.e., "no solution" state) caused by sudden environmental changes, and prevents the incorrect reuse of unsuitable historical strategies under extreme conditions, thereby maximizing the reuse rate of historical driving experience and the robustness of the system while ensuring driving safety.

[0035] In some embodiments, the dynamic tolerance threshold can be calculated in real time by a preset mapping function. The input variables of the mapping function include at least the real-time estimated road adhesion coefficient, and the output variable is the current distance metric tolerance value. When the road adhesion coefficient is lower than the preset safety threshold, the mapping function exhibits a non-linear growth trend to adaptively relax the allowable error range of scene matching.

[0036] In some embodiments, setting the dynamic tolerance threshold includes: selecting a subset of historical scenes from the scene database that are consistent with the current weather conditions and road adhesion coefficient labels; calculating the statistical distribution characteristics (such as standard deviation or specific quantiles) of the distance metric between the historical scene feature vectors and the cluster centers in the subset; and using a preset multiple of the statistical distribution characteristics as the current dynamic tolerance threshold, thereby ensuring that the threshold always conforms to the statistical laws of data under the current environment.

[0037] In some embodiments, the setting of the dynamic tolerance threshold is also based on the vehicle dynamics safety boundary in the current environment: first, the maximum feasible deceleration and safe braking distance of the vehicle under extreme conditions are calculated based on the current road adhesion coefficient; then, the maximum allowable deviation in each dimension of the scene feature vector is derived in reverse, so that the predicted trajectory after reusing the historical strategy is still within the safety boundary; the maximum deviation is converted into a value in the distance metric space as the upper limit of the dynamic tolerance threshold to prevent the reuse of high-risk historical driving strategies on low-adhesion surfaces.

[0038] In some embodiments, the dynamic tolerance threshold can also be generated by fusing multi-source environmental perception information: the system acquires the road adhesion coefficient, weather visibility level and confidence score of vehicle-mounted sensors in real time; normalizes each parameter and assigns dynamic weights, wherein the weight of the sensor confidence score increases with the severity of the environment; calculates the environmental uncertainty index by weighted summation, and scales the benchmark threshold proportionally according to the index to obtain the final dynamic tolerance threshold.

[0039] In some embodiments of the present invention, in a longitudinal following scenario, step S5 specifically includes: when the relative longitudinal speed, relative longitudinal distance, and relative longitudinal acceleration of the vehicle and the vehicle in front are all detected to fall within the preset threshold range of the driver's historical active driving data, extracting the accelerator pedal opening or brake pedal pressure value of the driver at that historical moment; mapping the extracted accelerator pedal opening or brake pedal pressure value to the vehicle's desired longitudinal acceleration as the core parameter of the initial adaptive control strategy.

[0040] For longitudinal following scenarios, the system first enters a multi-dimensional feature matching stage, monitoring three key kinematic indicators of the vehicle and the vehicle ahead in real time: relative longitudinal speed, relative longitudinal distance, and relative longitudinal acceleration. The system compares these three real-time indicators with historical active driving data stored in the database. Only when all three fall within a preset threshold tolerance range is the current condition deemed highly similar to a historical driving segment, triggering the strategy extraction mechanism. This "three-parameter joint constraint" effectively avoids misjudgments that may arise from single-parameter matching (e.g., similar distance but completely different speed trends), ensuring the accuracy of scenario reuse. Once a match is successful, the system immediately enters the driving intent extraction stage, accurately capturing the driver's operational behavior data from the historical data record, specifically accelerator pedal opening (representing acceleration intent) or brake pedal pressure (representing deceleration intent). This data directly reflects the driver's personalized driving habits (e.g., aggressive or conservative) in similar following conditions. Finally, the system performs control parameter mapping and generation, converting the extracted pedal operation amount into a specific desired longitudinal acceleration through a vehicle dynamics model or calibration mapping curve. This desired acceleration not only incorporates the kinematic requirements for maintaining a safe following distance but also integrates the driver's comfort preferences and style characteristics, thus being established as the core parameter of the initial adaptive control strategy. Subsequent vehicle controllers will use this as a core benchmark, fine-tuning the execution based on the current vehicle state, thereby achieving human-like longitudinal control that conforms to both current traffic flow dynamics and the driver's personal habits.

[0041] The preset threshold range is not a fixed value, but can be adaptively adjusted in conjunction with the 'dynamic tolerance threshold' mechanism mentioned above. For example, in high-speed following scenarios, the threshold requirements for relative longitudinal distance are more stringent; while in low-speed congested scenarios, the tolerance for fluctuations in relative longitudinal acceleration is appropriately relaxed to adapt to the frequent start-stop operation characteristics of drivers in congested road conditions.

[0042] In some embodiments of the present invention, in lateral parallel scenarios or trajectory intersection scenarios, step S5 specifically includes: when the lateral relative speed, lateral relative distance, lateral relative acceleration, longitudinal relative speed, longitudinal relative distance, and longitudinal relative acceleration of the adjacent lane traffic participant and the vehicle all fall within the preset threshold range of the driver's historical active driving data, extracting the driver's steering wheel angle and angular velocity at that historical moment; mapping the extracted steering wheel angle and angular velocity to the vehicle's desired lateral displacement trajectory or yaw rate as the core parameters of the initial adaptive control strategy.

[0043] For lateral parallel scenarios (such as overtaking by adjacent vehicles) or trajectory intersection scenarios (such as merging and intersection passage), the system first enters a six-dimensional motion state joint matching stage. Given the complexity and high risk of lateral interaction scenarios, the system is no longer limited to single-dimensional judgment, but simultaneously monitors the complete kinematic characteristics of the vehicle and adjacent lane traffic participants in both the lateral and longitudinal planes. Specifically, the system calculates and compares the following six key indicators in real time: Lateral dimension: lateral relative velocity, lateral relative distance (i.e., lateral spacing), lateral relative acceleration; Longitudinal dimension: longitudinal relative velocity, longitudinal relative distance (i.e., distance between vehicles), longitudinal relative acceleration.

[0044] Only when the real-time data from these six dimensions simultaneously falls within the preset threshold range (i.e., dynamic tolerance threshold) in the driver's historical active driving database does the system determine that the current complex interaction situation is highly consistent with a certain historical segment in terms of "spatiotemporal situation." This multi-dimensional constraint effectively eliminates pseudo-similar scenarios where "the lateral distances appear similar but the longitudinal approach speeds are completely different," greatly improving the safety of strategy reuse. Once a match is successful, the system immediately enters the driver's control intention extraction stage, accurately capturing the driver's steering wheel angle and its angular velocity from the record of that historical moment. These two parameters not only reflect the driver's desired steering range but also imply the driver's operational agility and smoothness preference in that specific risk scenario (e.g., quick sharp steering to avoid an obstacle or slow and gentle lane adjustment).

[0045] Finally, the system performs lateral kinematic mapping and strategy generation. Using the vehicle's lateral dynamics model or a data-driven mapping function, the extracted steering angle and angular velocity are transformed into the vehicle's desired lateral displacement trajectory or desired yaw rate. The desired lateral displacement trajectory directly plans a smooth lane change or avoidance path that conforms to the driver's habits. The desired yaw rate serves as the tracking target for the upper-level controller (such as LKA or LCA systems), guiding the vehicle to complete actions at a specific rotational rate. This transformation result is then established as the core parameter of the initial adaptive control strategy, enabling the vehicle to maintain a safe interaction distance when facing complex lateral interactions while also showcasing the driver's personalized steering style.

[0046] In the process of matching six-dimensional motion states, the system can dynamically adjust the weight coefficients of each dimension according to the scene type. For example, in a horizontally parallel scene, the weights of 'lateral relative distance' and 'lateral relative velocity' are appropriately increased to ensure lateral safety margin; while in a trajectory intersection scene, the matching accuracy of 'vertical relative distance' and 'vertical relative velocity' is strengthened to accurately determine the intersection timing, further improving the robustness of scene recognition.

[0047] Furthermore, the extracted steering wheel angular velocity is a key feature characterizing the driver's operating style (aggressive / conservative). When mapping to the desired lateral displacement trajectory, the system not only considers the magnitude of the steering angle but also reconstructs the rate of curvature change of the historical trajectory through integral calculations combined with the steering angular velocity. Finally, it generates an ideal reference trajectory that conforms to the driver's habits and satisfies vehicle dynamics continuity constraints (such as lateral acceleration and jerk limits).

[0048] In addition, the generated initial adaptive control strategy (desired trajectory or yaw rate) undergoes a secondary collision risk verification before output. Based on the current real-time six-dimensional relative state, the system predicts whether executing the historical strategy would trigger a potential collision (especially considering the possible gear shifting behavior of adjacent vehicles). If the predicted risk exceeds the safety threshold, the system will automatically attenuate the amplitude of the desired trajectory or the peak value of the yaw rate, or switch to a rule-based minimum risk strategy, ensuring that while reusing historical experience, current driving safety is given absolute priority.

[0049] In some embodiments of the present invention, the safety boundary verification in step S6 specifically includes: obtaining the current dynamic limit parameters of the vehicle and the regulatory restriction parameters under the current traffic scenario, and constructing a safe feasible region; determining whether the control parameters in the initial adaptive control strategy are within the safe feasible region; if they are within the safe feasible region, they are directly output; if they exceed the safe feasible region, a projection algorithm is used to truncate the control parameters to the boundary value of the safe feasible region, or the control parameters are output after smoothing and filtering.

[0050] Specifically, step S6, security boundary verification, acts as a firewall connecting "personalized driving strategy" and "vehicle physical / regulatory security." Its key lies in constructing a multi-dimensional security feasible domain and imposing real-time constraints and corrections on the initial strategy. The specific execution process is as follows: First, the system integrates two types of key constraints: vehicle dynamics limit parameters and traffic scenario regulatory restrictions, and dynamically defines the space of control parameters allowed at the current moment.

[0051] Among them, the vehicle dynamics limit parameters are calculated based on real-time perceived road adhesion coefficient, vehicle load, center of gravity position, and tire characteristics to determine the vehicle's physical limits. Longitudinal: Maximum available driving force / braking force (limited by the friction circle) to prevent slippage or frequent ABS triggering. Lateral: Maximum lateral acceleration and maximum yaw rate to prevent vehicle instability (such as fishtailing or understeer). Coupling constraints: Considering the coupling of longitudinal and lateral forces (friction ellipse), i.e., the available steering ability will decrease accordingly under heavy braking. Traffic scenario regulatory limit parameters are extracted based on high-precision maps, traffic sign recognition, and traffic regulations, extracting soft / hard constraints at the legal and safety levels. Speed ​​limit: The legally mandated maximum / minimum speed for the current road segment. Lane boundaries: Lane line position and distance from the curb to ensure the vehicle does not cross the line. Safe following distance: The legally mandated minimum following distance or collision time threshold calculated based on the current vehicle speed and relative speed. Therefore, by mapping the above parameters to the control parameter space (e.g., the desired acceleration-desired yaw rate plane, or the desired trajectory-velocity plane), a closed safe and feasible domain is formed. This area is dynamic and changes in real time with road conditions and vehicle speed.

[0052] Then, the system substitutes the initial adaptive control strategy generated in step S5 (which includes core parameters such as desired longitudinal acceleration, desired lateral displacement trajectory, or yaw rate) into the above-mentioned safe and feasible domain model for judgment: Scenario 1: Completely within the domain. If all control parameters meet the dynamic limits and comply with regulations, the personalized strategy is safe and legal in the current environment. Action: Direct output. The system retains the driver's original driving style and executes the strategy without intervention.

[0053] Scenario 2: Partial or complete violation of the domain. If one or more parameters in the initial strategy exceed physical limits (e.g., requesting deceleration exceeding the road surface adhesion capacity) or violate regulations (e.g., lane change trajectory crossing a solid line or insufficient space in the target lane). Action: Activate the correction mechanism.

[0054] The correction mechanism addresses situations outside the safe and feasible region by employing one or a combination of the following two strategies to ensure the absolute safety of the output instructions: Strategy 1: Projection Algorithm Truncation – Project the excess control parameter vector along the shortest path (or a specific weight direction) to the boundary of the safe and feasible region. For example, if the driver's historical habits request... The deceleration is 0.8g, but the current limit on the wet surface is only -0.4g, so the system forcibly truncates the output value to... 0.4g.

[0055] Strategy Two: Smoothing Filtering – When direct truncation might cause a sudden change in control commands, leading to vehicle jerking or passenger discomfort, the system introduces a smoothing filter to generate a temporary trajectory that smoothly transitions from the current state to the boundary value within the safe and feasible domain. This not only satisfies safety constraints but also ensures the continuity and comfort of control, avoiding vehicle tremors caused by sudden policy changes.

[0056] The process of constructing the safe and feasible domain specifically includes building a dynamic friction ellipse based on the real-time estimated road adhesion coefficient. The system synthesizes the longitudinal acceleration request and lateral acceleration request from the initial adaptive control strategy into a total acceleration vector and determines whether the endpoint of this vector is located inside the dynamic friction ellipse. If it exceeds the ellipse boundary, a projection algorithm is used to compress the vector radially to the ellipse boundary. This reduces the amplitude of each component proportionally while preserving the driver's intended direction of operation (such as simultaneous braking and steering), ensuring that the combined tire force does not exceed the adhesion limit.

[0057] The construction of the safe feasible domain follows a hierarchical priority: the first level consists of hard constraints (such as vehicle dynamic stability limits and physical collision boundaries), and any parameters exceeding this level must be forcibly truncated; the second level consists of soft constraints (such as comfort limits, recommended speeds, and lane centering preferences). When the initial strategy only exceeds the soft constraints but does not violate the hard constraints, a gradual correction using smoothing filtering is prioritized to preserve the driver's personalized style to the greatest extent possible while ensuring safety; forced truncation is only implemented when smoothing still fails to meet the hard constraints.

[0058] In addition, the initial adaptive control strategy is frequently corrected by the safety boundary verification module (i.e., frequent truncation or filtering is triggered). The system marks this event as 'environment mismatch' or 'policy outdated' and feeds it back to the learning module. The learning module will reduce the weight of this historical data in similar scenarios or remove the data in subsequent updates, thereby prompting the adaptive control strategy to continuously evolve with environmental changes and reducing the frequency of correction.

[0059] In some embodiments of the present invention, an online iterative update step is also included: during the operation of the autonomous driving mode, the driver's takeover behavior is monitored; if the driver takeover is detected, the traffic environment feature vector at the moment of takeover and the driver's actual operation data after takeover are recorded to form a new sample pair; the new sample pair is stored in the scene-behavior mapping database, and the historical active driving data in the scene is updated using a time decay weight algorithm; wherein, the time decay weight algorithm assigns a higher weight coefficient to the most recent data to correct the subsequently generated adaptive control strategy.

[0060] To overcome the limitations of traditional offline training models, such as one-time learning and inability to adapt to changes, the system also includes an online iterative update mechanism. This mechanism utilizes the driver's takeover behavior in autonomous driving mode as a high-value feedback signal to achieve real-time evolution and self-correction of the driving style model. The specific implementation process is as follows: During autonomous driving operation, the system monitors steering wheel torque, pedal opening, and gear shift signals in real time. Once a driver actively intervenes and takes over vehicle control (identified as a "takeover event"), the system immediately triggers a data recording procedure. First, environmental feature capture: The system traces back and locks down traffic environment feature vectors within a preset time window (e.g., t0-Δt to t0) prior to the takeover moment (t0). This vector includes not only conventional relative speed, distance, and acceleration, but also information such as weather, road curvature, and neighboring vehicle behavior. This represents the triggering scenario that caused the autonomous driving strategy to fail or not meet the driver's expectations. Second, operation data recording: The system synchronously records the driver's actual operation data sequence (including steering wheel angle, accelerator / brake pedal curves, etc.) over a subsequent period after takeover (e.g., t0 to t0+T). This represents the driver's perceived "correct" or "preferred" response strategy in that specific scenario. Third, sample pair generation: The system packages the aforementioned environmental feature vectors with the actual operation data to form a new scenario-behavior sample pair.

[0061] The newly generated sample pairs are immediately stored in the scene-behavior mapping database. Unlike a static database, this database supports dynamic writing, ensuring that the latest driving interaction experiences can be obtained by the system in real time. For recurring similar scene labels, new samples will be added as the latest members of that scene cluster, enriching the data distribution under that condition.

[0062] When generating adaptive control strategies from historical data, the system no longer treats all historical data "equally." Instead, it introduces a time-decay weighting algorithm, assigning a dynamic weight coefficient to each historical data entry in the database. This coefficient is a function of the data's generation time and the current time. A recentity-first principle can be adopted; that is, according to the algorithm, the most recent data (i.e., newly recorded takeover correction data) will receive the highest weight coefficient, while the weight of older historical data gradually decreases over time. The weighted and updated database will be directly used for strategy extraction in the next step, S5. This means that each driver takeover is an online learning process. By continuously absorbing these "correction instructions," the system makes the generated initial adaptive control strategy increasingly accurate in matching the driver's current true intentions, ultimately reducing the frequency of takeovers and achieving seamless integration of human-machine co-driving.

[0063] In addition, before updating the strategy, the system performs a confidence assessment on new sample pairs. Only when the duration of the driver's operation after takeover exceeds the minimum effective duration (e.g., 2 seconds), and the operation trajectory is smooth and conforms to dynamic constraints, is the sample considered valid teaching data and included in the update process. This effectively filters out unintentional touches or atypical operations during emergency avoidance by the driver, ensuring the robustness of iterative updates.

[0064] In some embodiments of the present invention, the specific calculation formula of the time decay weight algorithm is as follows: for the i-th historical data under the same scene feature, its weight is... ;in, The current system time. The record time of this historical data is λ, and the preset forgetting factor is λ. When generating the adaptive control strategy, the operating parameters of multiple historical data are fused using a weighted average method. The calculation formula is as follows:

[0065] It is understandable that, for the i-th historical data in the scene feature library that belongs to the same scene (i.e., the environmental feature vector similarity is higher than a preset threshold), the system calculates its timeliness weight coefficient based on the time difference between it and the current time. .

[0066] in: is the weight coefficient of the i-th historical data point, with a value range of (0,1). Time (i.e., data that has just occurred), It possesses the greatest influence; as time goes by, The exponential decay approaches 0.

[0067] It is the current system timestamp, representing the real-time moment when the strategy was generated. It is the timestamp of the i-th historical data record, representing the moment when the driving behavior occurred. It is a preset forgetting factor, and >0. This parameter determines how quickly the system "forgets" old habits. Larger: This indicates that the system is extremely sensitive to recent data and quickly forgets older data. It is suitable for scenarios where driver habits change frequently or the environment is highly dynamic. Smaller: Indicates the system has a longer memory window, historical data has a lasting impact, suitable for scenarios where driver style is stable and smoothing of long-term statistical characteristics is required. Note: It can be dynamically adjusted according to the scenario type. For example, a larger value can be set in "emergency avoidance" scenarios. To quickly learn the latest safety procedures, and in the "cruise following" scenario, set a smaller [safety level]. To maintain stylistic consistency.

[0068] After extracting n historical data points for the matching scenario in step S5, the system does not use a simple arithmetic mean, but instead uses the aforementioned weighting coefficients to perform a weighted average fusion to calculate the final target control parameters. .

[0069] in, The generated target control parameters (such as desired longitudinal acceleration, desired steering wheel angle, desired yaw rate, etc.) will serve as the core inputs to the initial adaptive control strategy.

[0070] is the actual operation parameter value recorded in the i-th historical data (i.e., the driver's actual operation at that historical moment). n is the total number of valid historical data currently matched. It is the weighted sum of operations, reflecting the dominant role of recent high-weight data in the results. It is a normalization factor that ensures that the dimensions of the fused parameter values ​​are consistent with the original operating parameters, thus avoiding numerical divergence due to fluctuations in the amount of data.

[0071] Through the implementation of the above-described formula, this embodiment achieves the following technical advantages: Dynamic adaptability: When the driver most recently takes over and corrects a certain operation (resulting in new...) , ),because Its weight The weight of older data has decayed. Therefore, Will quickly to By moving closer together, the vehicle immediately "learns" the driver's latest intentions. Compared to using only the latest data (n=1), the weighted average method utilizes information from n data points. Even if there is a random error in the latest operation, due to the ∑ in the denominator... The existence of and the smoothing contribution of old data, It avoids drastic changes, ensuring continuous control and comfort. Through adjustment... These parameters can precisely control the system's learning rate and memory length, allowing it to adapt to different vehicle models and positioning.

[0072] According to a second aspect of the present invention, a vehicle driving control system includes: a perception module, a storage module, a processing module, and an execution module. The perception module is equipped with at least one of a lidar, a millimeter-wave radar, and a camera, and is configured to collect traffic environment data around the vehicle to construct a three-dimensional traffic environment model and predict the trajectory. The storage module stores a scene-behavior mapping database, which stores historical relative motion state vectors constructed by a method and corresponding active driving operation data. The processing module is connected to the perception module and the storage module, and is configured to execute method steps, including scene similarity matching, control strategy generation, and safety boundary verification. The execution module is connected to the processing module and is used to drive the vehicle's throttle, braking, and steering systems according to final control commands.

[0073] According to a third aspect of the present invention, a computer-readable storage medium has a computer program stored thereon that implements a method when executed by a processor.

[0074] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0075] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example.

[0076] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An adaptive vehicle driving control method based on scene feature matching, characterized in that, Includes the following steps: Step S1: Collect real-time traffic environment data around the vehicle using onboard sensor arrays, construct a three-dimensional traffic environment model, and predict the movement trajectories of surrounding traffic participants; Step S2: In manual driving mode, record the driver's active driving operation data under various preset driving scenarios. The active driving operation data includes at least the steering wheel angle and angular velocity, lateral acceleration, longitudinal acceleration, yaw rate, accelerator pedal depth and brake pedal depth. Step S3: Extract key environmental feature parameters under the preset driving scenario and construct a scenario-behavior mapping database; the key environmental feature parameters are represented as relative motion state vectors, which include the relative speed, relative distance and relative acceleration between the vehicle and surrounding traffic participants; Step S4: In autonomous driving mode, calculate the relative motion state vector of the current traffic environment in real time, and perform similarity matching with the historical relative motion state vector in the scene-behavior mapping database; Step S5: If the similarity between the current relative motion state vector and a certain historical vector in the database exceeds a preset confidence threshold, then call the active driving operation data corresponding to the historical vector to generate an initial adaptive control strategy. Step S6: Perform a safety boundary check on the initial adaptive control strategy. If the check passes, convert it into a final control command and send it to the vehicle actuator. If the check fails, modify the initial adaptive control strategy or switch to the default safety control strategy.

2. The method according to claim 1, characterized in that, The traffic environment data collected in step S1 includes at least: the position, speed, and acceleration information of surrounding vehicles; the position and movement trajectory information of pedestrians; the geometric shape and type information of lane lines; traffic sign recognition information; and traffic light status information. In step S3, based on the spatial positional relationship and movement trend between the vehicle and surrounding traffic participants, the relative motion state vector is divided into at least three types of scene feature vectors: The first category is longitudinal following scenario vectors, whose dimensions include at least: the relative longitudinal speed, relative longitudinal distance, and relative longitudinal acceleration between the vehicle and the vehicle in front; The second category is the lateral parallel scene vector, whose dimensions include at least: the relative lateral speed, relative lateral distance, and relative longitudinal speed of the vehicle and the vehicle traveling in the same direction in the adjacent lane. The third category is trajectory intersection scene vectors, whose dimensions include at least: the relative lateral velocity, relative longitudinal velocity, relative distance, and relative acceleration of the vehicle and the intersecting vehicle at the predicted intersection point.

3. The method according to claim 2, characterized in that, The similarity matching in step S4 specifically includes: Calculate the distance metric between the current real-time scene feature vector and the historical scene feature vector in the database; Set a dynamic tolerance threshold. When the distance metric value is less than or equal to this threshold, the scene is considered to be successfully matched. The dynamic tolerance threshold is dynamically adjusted at least according to the current road adhesion coefficient or weather conditions.

4. The method according to claim 2, characterized in that, In the longitudinal following scenario, step S5 specifically includes: When the relative longitudinal speed, relative longitudinal distance, and relative longitudinal acceleration of the vehicle and the vehicle in front are all detected to fall within the preset threshold range of the driver's historical active driving data, the accelerator pedal opening or brake pedal pressure value of the driver at that historical moment is extracted. The extracted accelerator pedal opening or brake pedal pressure value is mapped to the vehicle's desired longitudinal acceleration, serving as the core parameter of the initial adaptive control strategy.

5. The method according to claim 2, characterized in that, In the aforementioned horizontally parallel scenario or trajectory intersection scenario, step S5 specifically includes: When the lateral relative speed, lateral relative distance, lateral relative acceleration, longitudinal relative speed, longitudinal relative distance, and longitudinal relative acceleration of the adjacent lane traffic participants and the vehicle all fall within the preset threshold range of the driver's historical active driving data, the driver's steering wheel angle and angular velocity at that historical moment are extracted. The extracted steering wheel angle and angular velocity are mapped to the vehicle's desired lateral displacement trajectory or yaw rate, serving as the core parameters of the initial adaptive control strategy.

6. The method according to claim 1, characterized in that, The security boundary verification in step S6 specifically includes: Obtain the vehicle's current dynamic limit parameters and regulatory constraints under the current traffic scenario, and construct the safe and feasible domain; Determine whether the control parameters in the initial adaptive control strategy are within the safe and feasible region; If the control parameter is within the safe and feasible region, it is output directly; if it is outside the safe and feasible region, a projection algorithm is used to truncate the control parameter to the boundary value of the safe and feasible region, or the control parameter is output after smoothing and filtering.

7. The method according to claim 1, characterized in that, It also includes online iterative update steps: During autonomous driving mode operation, monitor driver takeover attempts; If driver takeover is detected, record the traffic environment feature vector at the moment of takeover and the driver's actual operation data after takeover to form a new sample pair; The new sample pairs are stored in the scenario-behavior mapping database, and the historical active driving data under the scenario is updated using a time decay weight algorithm. The time decay weighting algorithm assigns a higher weight coefficient to the most recent data to correct the subsequently generated adaptive control strategy.

8. The method according to claim 7, characterized in that, The specific calculation formula for the time decay weight algorithm is as follows: For the i-th historical data point under the same scene features, its weight ; in, The current system time. The record time of this historical data is λ, and the preset forgetting factor is λ. When generating the adaptive control strategy, a weighted average method is used to fuse the operating parameters from multiple historical data points. The calculation formula is as follows: 。 9. A vehicle driving control system, characterized in that, include: The perception module is equipped with at least one of lidar, millimeter-wave radar and camera, and is configured to collect traffic environment data around the vehicle to build a three-dimensional traffic environment model and predict the trajectory. The storage module is used to store a scene-behavior mapping database, wherein the database stores historical relative motion state vectors constructed by the method of any one of claims 1 to 8 and corresponding active driving operation data; The processing module, connected to the perception module and the storage module, is configured to perform the steps of the method according to any one of claims 1 to 8, including scene similarity matching, control strategy generation, and security boundary verification; An execution module, connected to the processing module, is used to drive the vehicle's throttle, braking, and steering systems according to the final control command.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method described in any one of claims 1 to 8.

Images