Intelligent driving control method, system and equipment based on environment perception and storage medium

Abstract

The invention discloses an intelligent driving control method, system and device based on environment perception, and a storage medium. The method comprises the following steps: performing 4D-Var assimilation and EnKF disturbance analysis on a current moment environment observation vector field corresponding to multi-source original data and a latest short-term forecast field; the disturbance analysis set and the SPDE-INLA physical field are input into an HSTAN model set, an environment state forecasting field in the future time is output, and the HSTAN model set comprises a plurality of HSTAN sub-models constructed based on a physical constraint U-RNN module; performing uncertainty quantification processing on the environment state forecast field; and carrying out data fusion on the environment probability forecast result, the vehicle perception result and the map static information, and carrying out driving parameter mapping on the fused comprehensive information to realize intelligent driving control. According to the invention, through data assimilation, disturbance prediction, uncertainty quantification, adaptive fusion and accurate mapping of driving parameters, the reliability of intelligent driving control is improved.

Description

Technical Field

[0001] This invention relates to the field of driving control technology, and in particular to an intelligent driving control method, system, device, and storage medium based on environmental perception. Background Technology

[0002] With the deep integration of the automotive industry and artificial intelligence, intelligent driving has become a core development direction in the transportation sector. Its core lies in improving driving safety, comfort, and traffic efficiency through precise environmental perception and efficient decision-making and control. Environmental perception is the core foundation of intelligent driving systems, especially the dynamic changes in meteorological environment (fog, crosswinds, rain, snow, road icing, etc.) and road traffic environment (traffic participants, road structure, etc.), which directly determine the driving stability and safety of intelligent driving vehicles.

[0003] Existing intelligent driving environmental perception and control technologies primarily rely on real-time perception from sensors such as onboard cameras and LiDAR, combined with static road information from high-precision maps to achieve environmental awareness and control parameter matching. However, this approach has significant limitations and struggles to adapt to complex dynamic environments. Specific problems include: First, the sensing range and timeliness are insufficient. The sensing range of vehicle-mounted sensors is limited to a few hundred meters around the vehicle body, and can only obtain the current environmental conditions, unable to predict future environmental changes. When faced with sudden weather disasters (such as dense fog, strong crosswinds) or sudden road conditions, the system is unable to adjust its decisions in time, which can easily lead to safety risks such as delayed braking and loss of steering control. Secondly, the accuracy and reliability of multi-source data fusion are unsatisfactory. In existing technologies, multi-source data such as meteorological, vehicle-mounted sensing, and map data are often in isolation or simply superimposed, without fully considering the uncertainties and reliability differences of each data source, which affects the accuracy of the fusion results. Third, environmental prediction models lack physical consistency constraints. Most existing prediction models are purely data-driven deep learning methods, and the prediction results may violate objective physical laws (such as sudden temperature changes or abnormal wind speeds), resulting in insufficient reliability. Fourth, the driving parameter mapping logic is not accurate or comprehensive. Existing technologies mostly map driving parameters based on a single environmental element, which can easily lead to invalid or erroneous control commands and reduce driving reliability.

[0004] Therefore, improving the reliability of intelligent driving control has become an urgent problem to be solved.

[0005] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention The main objective of this invention is to provide an intelligent driving control method, system, device, and storage medium based on environmental perception, aiming to address the technical problem of how to improve the reliability of intelligent driving control.

[0006] To achieve the above objectives, the present invention provides an intelligent driving control method based on environmental perception, the intelligent driving control method based on environmental perception comprising: Real-time acquisition of multi-source raw data within a preset range of the currently driven vehicle, and preprocessing of the multi-source raw data to obtain the environmental observation vector field at the current moment; The current environmental observation vector field and the latest short-term forecast field are assimilated using 4D-Var to obtain the current environmental analysis field; The current environmental analysis field is subjected to EnKF perturbation to obtain the EnKF perturbation analysis set; The EnKF perturbation analysis set and the SPDE-INLA physical field are input into the HSTAN model set, and the environmental state prediction field within the prediction time range is output. The HSTAN model set includes multiple HSTAN sub-models constructed based on the physically constrained U-RNN module. Uncertainty quantification is performed on the environmental state prediction field to generate environmental probability prediction results; The environmental probability prediction results, vehicle perception results, and map static information are fused together, and the fused comprehensive information is used to map driving parameters to achieve intelligent driving control.

[0007] Optionally, the step of performing 4D-Var assimilation of the current environmental observation vector field with the latest short-term forecast field to obtain the current environmental analysis field includes: Using the latest short-term forecast field as the background field, an objective function is constructed based on the current environmental observation vector field and the background field. The objective function is:

[0008] In the formula, For the environmental analysis variables to be solved, For background field variables, The background error covariance matrix, For the observation vector, For the observation operator, The observation error covariance matrix; The current environmental analysis field is obtained by minimizing the objective function to solve for the environmental analysis variables.

[0009] Optionally, the step of performing EnKF perturbation on the current environmental analysis field to obtain an EnKF perturbation analysis set includes: Based on the EnKF algorithm, a Gaussian-distributed perturbation term is added to the current environmental analysis field to generate multiple perturbation analysis samples; The EnKF algorithm:

[0010]

[0011] In the formula, For the i-th disturbance analysis sample, Let i be the i-th random perturbation vector. This represents the total number of samples in the perturbation analysis. An EnKF perturbation analysis set was constructed based on multiple perturbation analysis samples.

[0012] Optionally, the step of inputting the EnKF perturbation analysis set and the SPDE-INLA physical field into the HSTAN model set and outputting the environmental state forecast field within the prediction time range includes: The EnKF perturbation analysis set and the SPDE-INLA physics field are input into the HSTAN model set; Each EnKF perturbation analysis sample in the EnKF perturbation analysis set is transmitted to each HSTAN sub-model respectively; The physical constraint U-RNN module in each HSTAN sub-model performs time-series prediction and feature fusion on the corresponding EnKF perturbation analysis samples and the SPDE-INLA physical field to obtain the single-sample environmental state prediction result. The single-sample environmental state prediction results output by all HSTAN sub-models are averaged and fused to output the environmental state forecast field within the prediction time range.

[0013] Optionally, the uncertainty quantification of the environmental state forecast field to generate environmental probability forecast results includes: The environmental state prediction field and the environmental state prediction results of each sample are aligned in terms of spatiotemporal dimensions. Based on the environmental state prediction field, the predicted values ​​of the corresponding spatial locations are extracted from the environmental state prediction results of each sample through a mapping relationship according to the environmental variables corresponding to the spatial locations at each time within the prediction time range. Calculate the variance and quantiles based on the predicted values ​​for spatial location at each time point; The confidence interval and event probability are determined based on the quantiles. The variance, quantiles, confidence intervals, and event probabilities are integrated to obtain environmental probability prediction results.

[0014] Optionally, the data fusion of the environmental probability prediction results, vehicle perception results, and map static information includes: The environmental probability prediction results, vehicle perception results, and static map information are preprocessed. A basic credibility assessment is performed on the preprocessed environmental probability forecast results, vehicle perception results, and map static information to obtain the initial credibility of the probability forecast, the initial credibility of the vehicle perception, and the initial credibility of the map scene. Based on the current driving scenario information, the probability prediction weight coefficient, vehicle perception weight coefficient, and map scenario weight coefficient are determined according to the initial confidence level of the probability prediction, the initial confidence level of the vehicle perception, and the initial confidence level of the map scenario. The preprocessed environmental probability prediction results, vehicle perception results, and map static information are fused based on the probability prediction weight coefficient, the vehicle perception weight coefficient, and the map scene weight coefficient.

[0015] Optionally, the step of mapping driving parameters onto the fused integrated information to achieve intelligent driving control includes: The integrated information after fusion is analyzed based on the time dimension, spatial dimension and element dimension to obtain a multi-dimensional driving mapping element set, which includes various environmental elements of different dimensions and their corresponding confidence levels. Based on various environmental factors and their corresponding confidence levels, multi-dimensional driving control parameters and driving strategy suggestions are obtained through a preset environmental factor-driving parameter mapping rule base. Based on the environmental probability prediction results, driving commands are generated according to multi-dimensional driving control parameters and driving strategy suggestions for intelligent driving control.

[0016] Furthermore, to achieve the above objectives, the present invention also proposes an intelligent driving control system based on environmental perception, wherein the intelligent driving control system based on environmental perception includes: The data processing module is used to collect multi-source raw data within a preset range of the currently driven vehicle in real time, and to preprocess the multi-source raw data to obtain the environmental observation vector field at the current moment. The 4D-Var assimilation module is used to perform 4D-Var assimilation between the current environmental observation vector field and the latest short-term forecast field to obtain the current environmental analysis field. The EnKF perturbation module is used to perturb the current environmental analysis field using EnKF to obtain an EnKF perturbation analysis set. The HSTAN inference module is used to input the EnKF perturbation analysis set and the SPDE-INLA physical field into the HSTAN model set and output the environmental state forecast field within the prediction time range. The HSTAN model set includes multiple HSTAN sub-models constructed based on the physical constraint U-RNN module. The uncertainty quantification module is used to perform uncertainty quantification processing on the environmental state forecast field and generate environmental probability forecast results; The data processing module is also used to fuse the environmental probability prediction results, vehicle perception results and map static information. The driving control module is used to map driving parameters to the fused integrated information in order to achieve intelligent driving control.

[0017] Furthermore, to achieve the above objectives, the present invention also proposes an environment-aware intelligent driving control device, the device comprising: a memory, a processor, and an environment-aware intelligent driving control program stored in the memory and executable on the processor, the environment-aware intelligent driving control program being configured to implement the steps of the environment-aware intelligent driving control method described above.

[0018] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing an environment-aware intelligent driving control program, wherein the environment-aware intelligent driving control program, when executed by a processor, implements the steps of the environment-aware intelligent driving control method described above.

[0019] This invention first collects multi-source raw data within a preset range of the currently driving vehicle in real time, and preprocesses the multi-source raw data to obtain the environmental observation vector field at the current moment. Then, it performs 4D-Var assimilation on the current environmental observation vector field and the latest short-term forecast field to obtain the current environmental analysis field. The current environmental analysis field is then perturbed by EnKF to obtain the EnKF perturbation analysis set. Subsequently, the EnKF perturbation analysis set and the SPDE-INLA physical field are input into the HSTAN model group to output the environmental state forecast field within the prediction time range. The HSTAN model group includes multiple HSTAN sub-models constructed based on the physical constraint U-RNN module. Finally, the uncertainty of the environmental state forecast field is quantified to generate the environmental probability forecast result. The environmental probability forecast result, vehicle perception result and map static information are fused, and the fused comprehensive information is mapped to driving parameters to achieve intelligent driving control. This invention constructs a closed-loop architecture of perception-prediction-decision-execution. It achieves environmental state prediction within a future time range through ultra-fast data assimilation, EnKF perturbation, and HSTAN model set. Then, by combining vehicle perception results and static map information, it realizes the accurate and orderly transformation from environmental information to control parameters, realizes intelligent driving control, and thus improves the reliability of intelligent driving. Attached Figure Description

[0020] Figure 1This is a schematic diagram of the structure of an intelligent driving control device based on environmental perception, which is part of the hardware operating environment involved in the embodiments of the present invention. Figure 2 This is a flowchart illustrating the first embodiment of the intelligent driving control method based on environmental perception of the present invention. Figure 3 This is a structural block diagram of the first embodiment of the intelligent driving control system based on environmental perception of the present invention.

[0021] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0023] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of an intelligent driving control device based on environmental perception, which is part of the hardware operating environment involved in the embodiments of the present invention.

[0024] like Figure 1 As shown, the environment-aware intelligent driving control device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk storage device. Optionally, the memory 1005 may also be a storage system independent of the aforementioned processor 1001.

[0025] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the environmental perception-based intelligent driving control device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0026] like Figure 1As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and an intelligent driving control program based on environmental perception.

[0027] exist Figure 1 In the environmental perception-based intelligent driving control device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the environmental perception-based intelligent driving control device of the present invention can be set in the environmental perception-based intelligent driving control device. The environmental perception-based intelligent driving control device calls the environmental perception-based intelligent driving control program stored in the memory 1005 through the processor 1001 and executes the environmental perception-based intelligent driving control method provided in the embodiment of the present invention.

[0028] This invention provides an intelligent driving control method based on environmental perception, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the intelligent driving control method based on environmental perception of the present invention.

[0029] In this embodiment, the intelligent driving control method based on environmental perception includes the following steps: S1, collect multi-source raw data within a preset range of the currently driven vehicle in real time, and preprocess the multi-source raw data to obtain the environmental observation vector field at the current moment.

[0030] It is easy to understand that the execution subject of this embodiment can be an environment-aware intelligent driving control system with functions such as data processing, network communication and program execution, or other computer devices with similar functions. This embodiment does not limit it.

[0031] It should be noted that the environmental perception data collected by the vehicle-mounted sensors, the roadside equipment data transmitted by the vehicle network, the real-time meteorological data collected by the meteorological monitoring equipment, and the traffic flow data collected by the traffic flow monitoring equipment are used as multi-source raw data. The preset range is a circular area or a rectangular area with a preset radius (e.g., 50-500 meters) centered on the currently driving vehicle.

[0032] The preprocessing methods for multi-source raw data are: data cleaning, data standardization, data spatiotemporal alignment, and outlier removal. Among them, data cleaning is used to remove missing values ​​and noisy data from multi-source raw data, data standardization is used to convert multi-source raw data with different dimensions into the same dimension, and data spatiotemporal alignment is used to unify multi-source raw data with different collection times and collection locations into the same spatiotemporal coordinate system.

[0033] S2, perform 4D-Var assimilation of the current environmental observation vector field and the latest short-term forecast field to obtain the current environmental analysis field.

[0034] Furthermore, the current environmental observation vector field and the latest short-term forecast field are assimilated using four-dimensional variational data assimilation (4D-Var). The current environmental analysis field is obtained by using the latest short-term forecast field as the background field and constructing an objective function based on the current environmental observation vector field and the background field. The environmental analysis variables are solved by minimizing the objective function to obtain the current environmental analysis field.

[0035] The objective function (characterizing the deviation between the current environmental observation vector field and the background field) is:

[0036] In the formula, These are the environmental analysis variables to be solved (i.e., the variable representations of the current environmental analysis field). These are background field variables (i.e., variables representing the latest short-term forecast field). This is the background error covariance matrix, used to describe the error distribution characteristics of the background field variables. It is the observation vector (i.e., the variable representation of the environmental observation vector field at the current moment). The observation operator is used to map analytical variables from the model space to the observation space, achieving dimensionality matching between analytical and observed variables. This is the observation error covariance matrix, used to describe the error distribution characteristics of the observed variables.

[0037] It should also be noted that the background field can be a refined initial environmental field (such as wind field, temperature field, visibility field, and road surface adhesion coefficient field) that is derived by the system based on historical data or obtained from the cloud and covers several kilometers around the vehicle.

[0038] S3, apply EnKF perturbation to the current environmental analysis field to obtain the EnKF perturbation analysis set.

[0039] Furthermore, based on the Ensemble Kalman Filter (EnKF) algorithm, a Gaussian-distributed perturbation term is added to the current environmental analysis field to generate multiple perturbation analysis samples, and an EnKF perturbation analysis set is constructed based on the multiple perturbation analysis samples.

[0040] EnKF algorithm:

[0041]

[0042] In the formula, For the i-th disturbance analysis sample, Let i be the i-th random perturbation vector, which follows a standard Gaussian distribution with mean 0 and variance 1. This represents the total number of samples in the perturbation analysis.

[0043] S4, input the EnKF perturbation analysis set and the SPDE-INLA physical field into the HSTAN model set, and output the environmental state forecast field within the prediction time range. The HSTAN model set includes multiple HSTAN sub-models constructed based on the physical constraint U-RNN module.

[0044] In the specific implementation, The goal of the training phase is to train the core AI model in the inference process using massive amounts of historical observation data and high-quality reanalysis data. Its core idea is physical information machine learning, which, through the design of the loss function, forces the model's learning process to adhere to fundamental physical conservation laws.

[0045] Training process for the Hierarchical Spatio-Temporal Attention Network (HSTAN) sub-model: 1. Historical multi-source data includes heterogeneous data spanning long periods (such as several years): Vehicle-mounted perception history logs: Camera image sequences, radar point clouds, and vehicle sensor data (temperature, wiper status, ESP / ABS trigger signals, etc.) transmitted from a fleet of mass-produced vehicles. This data is rigorously labeled with timestamps and location information. Roadside infrastructure history data: Historical records from fixed weather stations, visibility meters, and road surface sensors. Publicly available historical meteorological data: Historical weather station observation data, radar-derived precipitation data, etc.

[0046] It should be understood that the scale, quality, and diversity of data directly determine the ceiling of the model. It needs to cover various weather phenomena (sunny, rainy, snowy, foggy, crosswinds) and various road scenarios (highways, urban areas, bridges, tunnels).

[0047] 2. Spatiotemporal alignment and feature engineering The original historical multi-source data is cleaned, aligned, and fused, and features suitable for model training are extracted.

[0048] Spatiotemporal alignment: Unifying all data to a standard spatiotemporal coordinate system (such as UTC time, WGS-84 latitude and longitude, or local planar coordinates). For example, accurately matching an image of a vehicle taken at a specific time and location with data from the nearest weather station at that time, as well as cloud cover data for that location retrieved from satellite.

[0049] Feature engineering: High-level features can be extracted from raw data. For example, "texture features" extracted from camera images can be used to quantify fog concentration; features extracted from radar scattering signals can indirectly reflect the characteristics of precipitation particles; and subtle intervention signals from vehicle ESP can serve as weak indicators of slippery road surfaces.

[0050] Construct spatiotemporal sequence samples. Use historical data over continuous time (e.g., the past 30 minutes) as the input of a sample, and use the state over a future time period (e.g., the next 15 minutes) as the label of the sample.

[0051] It's important to note that this step transforms the messy raw data into "clean, aligned, and meaningful" <input, label> pairs, which is fundamental for effective subsequent training. The accuracy of this alignment directly affects the model's ability to learn "causal relationships" rather than "correlation relationships."

[0052] 3. Further analysis of the meteorological field Reanalysis is a meteorological concept that uses data assimilation techniques to integrate all historical observation data (ground, radiosonde, satellite, etc.) into a physical numerical model, thereby deriving a global, long-term, gridded, and physically consistent atmospheric state dataset (such as ERA5, MERRA-2).

[0053] During training, it can serve as a supervisory signal. For example, precise wind speed and temperature fields provided by reanalysis data can be used as labels to supervise the model's learning of how to predict these fields from vehicle data.

[0054] More importantly, it serves as a carrier of physical laws. Its data itself satisfies fluid dynamics and thermodynamic equations, providing a computational basis for physical constraints (such as convection-diffusion equations).

[0055] 4. Physically Constrained U-RNN U-RNN is a physically constrained recurrent neural network.

[0056] U-Net architecture: used for efficient processing of spatial information (such as local meteorological feature maps extracted from vehicle images) and upsampling to output full-field predictions.

[0057] Physical constraint embedding: The model's architecture or activation function may be designed to implicitly satisfy certain physical properties (such as conservation). But more importantly, physical constraints are explicitly introduced through the next module—the loss function.

[0058] 5. Convection-diffusion equations and TV regularization These two are key physical constraints in the loss function, used to guide the model to learn solutions that conform to physical laws, rather than just fitting the data.

[0059] The convection-diffusion equation is a fundamental partial differential equation describing the transport of physical quantities such as heat, water vapor, and momentum in the atmosphere. In the loss function, the spatiotemporal derivatives of the fields predicted by the model (such as temperature and water vapor concentration fields) are calculated to check the extent to which they satisfy this equation. Even in areas without observational data, model predictions must roughly obey this physical law. This greatly enhances the model's generalization ability and the reasonableness of its extrapolations.

[0060] TV regularization: Total variational regularization. It penalizes excessive, unphysical minor oscillations or noise in the prediction field, causing the model to produce "piecewise smooth" results. This physically corresponds to the fact that many meteorological fields (such as temperature and pressure) typically change smoothly in space rather than with abrupt jumps. TV regularization effectively suppresses overfitting, making predictions more stable and more in line with visual and physical intuition.

[0061] 6. Loss Function Calculation Function: Defines the overall objective of model training optimization, which is a weighted sum of data fitting terms and physical constraint terms.

[0062] Detailed content: Total loss function L total Typically designed as follows: L total =λ data * L data +λ physics * L pde +λ tv * L tv L data Data fitting loss. This calculates the difference, such as mean squared error, between the model's predictions and the "true labels" (labels generated from spatiotemporal alignment and feature engineering, or partial data from reanalysis of the meteorological field). This is part of traditional supervised learning.

[0063] L pde Physical constraint loss. This refers to the portion calculated based on the convection-diffusion equations, measuring the degree to which the prediction violates physical laws.

[0064] L tvRegularization loss. This is the part calculated based on TV regularization, which measures the roughness of the predicted field.

[0065] Λ data , λ physics , λ tv These are hyperparameters used to balance the importance of the three terms. By adjusting them, the optimal balance can be found between "fitting the data" and "observing physics".

[0066] 7. Backpropagation optimization and 8. Model parameter update The standard deep learning training process involves iteratively optimizing model parameters based on the gradients calculated from the loss function.

[0067] Forward propagation: Input a batch of training data into the physically constrained U-RNN to obtain the predicted output.

[0068] Loss Calculation: Calculate the total loss using the complex loss function calculation method described above.

[0069] Backpropagation: Using automatic differentiation, the gradient of the total loss with respect to each trainable parameter of the model is calculated. This process goes through the entire network and also includes differentiation of the physical equations.

[0070] Parameter update: Using an optimizer (such as Adam), all weights and bias parameters in the physically constrained U-RNN are updated based on the calculated gradient, so that the model can produce lower loss in the next iteration.

[0071] Repeat the above steps until the model converges on the validation set, resulting in a well-trained model that is both accurate and physically sound.

[0072] Furthermore, the EnKF perturbation analysis set and the physical field of Stochastic Partial Differential Equation - Integrated Nested Laplace Approximation (SPDE-INLA) are input into the HSTAN model set; each EnKF perturbation analysis sample within the EnKF perturbation analysis set is transmitted to each HSTAN sub-model; the physical constraint U-RNN module within each HSTAN sub-model performs time-series prediction and feature fusion on the corresponding EnKF perturbation analysis samples and SPDE-INLA physical field to obtain single-sample environmental state prediction results; the single-sample environmental state prediction results output by all HSTAN sub-models are averaged and fused to output the environmental state forecast field within the prediction time range.

[0073] In its implementation, the EnKF perturbation analysis set contains N (e.g., 20-50) perturbation analysis samples, each of which is the current environmental analysis field after adding Gaussian perturbations, covering the uncertainty range of the current environmental state; the SPDE-INLA physical field contains at least one of the atmospheric dynamics equation, fluid dynamics equation, and traffic flow dynamics equation, providing the physical constraints for the inference stage.

[0074] Then, each perturbation analysis sample in the EnKF perturbation analysis set is distributed one-to-one to each parallel sub-model of the HSTAN model group (each sub-model is built on a physically constrained U-RNN architecture, with independent parameters and consistent functions). Parallel inference of sub-models: After receiving the corresponding perturbation analysis samples, each HSTAN sub-model completes inference in three steps in conjunction with the SPDE-INLA physical field to ensure that the output results have both temporal rationality and physical consistency. The specific steps are as follows: Step 1: Temporal Feature Extraction and Evolution Trend Capture. The sub-model first inputs the received perturbation analysis samples (perturbation version of the current environmental analysis field) into the RNN / LSTM unit of the physically constrained U-RNN. The unit extracts the temporal features of the environmental state (such as the gradual trend of temperature change and the magnitude of wind speed change over a period of time) by parsing the time dimension data in the samples frame by frame, and based on historical temporal patterns, preliminarily predicts the evolution direction of the environmental state within the future prediction time range, forming the initial temporal prediction result; Step 2: SPDE-INLA Physical Constraint Correction. The initial time-series prediction results are matched and verified with the core equations (atmospheric dynamics equations, fluid dynamics equations, etc.) in the SPDE-INLA physical field: If the initial results contain features that violate physical laws (such as temperature abrupt changes exceeding the thermodynamic allowable range, or the direction of fluid motion conflicting with the derivation direction of the dynamic equations), then the abnormal parameters in the initial results are corrected according to the physical equations, forcing the prediction results to conform to objective physical laws; if the initial results conform to physical constraints, then the core prediction features are retained. Step 3: Spatial Feature Fusion and Final Result Output. The temporal prediction results, after being corrected by physical constraints, are then input into the U-Net structure of the physically constrained U-RNN. The U-Net processes the spatial information in the samples (such as the spatial distribution of local environmental features and the correlation between environmental parameters in different regions), fusing the temporal prediction results with spatial features to generate a single-sample environmental state prediction result that combines temporal continuity and spatial rationality. This result includes complete environmental variable data for each time point within the prediction time range (such as temperature, wind speed, visibility, etc. at different locations at each time point).

[0075] Ensemble averaging fusion: Collect single-sample forecast results from all HSTAN sub-models and fuse multiple sets of forecast results using an ensemble averaging algorithm. The specific implementation is as follows: The first step is result collection and alignment. This involves collecting single-sample environmental state forecasts from the N parallel sub-models in the HSTAN model set, ensuring consistency in the spatiotemporal dimensions of all sub-sample forecasts (i.e., complete matching of time points and spatial coordinate grid points within the prediction time range, avoiding fusion errors due to dimensional deviations). The second step is time-by-time, grid-point-by-grid mean calculation. For each time t within the prediction time range, and for each coordinate grid point (x, y) in the spatial domain at that time, the predicted environmental variables (such as temperature and wind speed) from all N sub-models at that (t, x, y) location are extracted. The arithmetic mean of these predicted values ​​is calculated, and this mean is used as the final predicted value of the fused environmental state forecast field at (t, x, y). The third step is generating the fusion result. This involves iterating through all time points and spatial grid points, completing the mean calculation across the entire spatiotemporal domain, and then integrating the results to obtain the complete environmental state forecast field. This method utilizes the statistical properties of the arithmetic mean to reduce random errors caused by single-sample perturbations, while also mitigating the impact of outlier sub-samples on the overall result, thus enhancing the stability and reliability of the forecast results.

[0076] It should also be noted that, assuming the HSTAN model group contains N parallel sub-models, the predicted value of the environmental variable at time t and spatial coordinates (x, y) in the single-sample prediction result output by the i-th (i=1,2,...,N) sub-model is f. i (t,x,y) (Note: Forecast values ​​refer to the model's quantitative predictions of key environmental variables such as temperature, wind speed, and visibility, which are the basis for subsequent fusion and uncertainty quantification). The final forecast value of the fused environmental state forecast field at (t,x,y) is F(t,x,y). Then, the formula for ensemble averaging fusion is:

[0077] F(t,x,y) represents the final predicted environmental variables after fusion; N is the number of sub-models in the HSTAN model set (i.e., the number of perturbation analysis samples in the EnKF perturbation analysis set, e.g., 20≤N≤50); i is the sub-model index; f i (t,x,y) represents the predicted environmental variable value of the i-th sub-model at time t and spatial coordinates (x,y); 1 / N is the average coefficient, used to convert the summation result into an arithmetic mean.

[0078] By averaging and fusing the above sets, the environmental state forecast field within the prediction time range (e.g., 5-30 minutes) is finally obtained.

[0079] It should also be understood that the SPDE-INLA physical field includes at least one physical constraint from the atmospheric dynamics equation, the fluid dynamics equation, and the traffic flow dynamics equation, which is used to guide the model output to produce prediction results that conform to physical laws.

[0080] S5, perform uncertainty quantification on the environmental state forecast field to generate environmental probability forecast results.

[0081] Furthermore, the environmental state forecast field and the environmental state prediction results of each sample are aligned in terms of spatiotemporal dimensions. Based on the environmental state forecast field, the predicted values ​​of the corresponding spatial locations at each time point within the prediction time range are extracted from the environmental state prediction results of each sample through mapping relationships according to the environmental variables corresponding to the spatial locations at each time point. The variance and quantiles are calculated based on the predicted values ​​of the spatial locations at each time point. The confidence interval and event probability are determined based on the quantiles. The variance, quantiles, confidence intervals, and event probabilities are integrated to obtain the environmental probability forecast results.

[0082] In the specific implementation, the validity of the fused environmental state forecast field is verified, and the N single-sample forecast sets (N=20-50) of its generation source are extracted through reverse correlation to establish a spatiotemporal one-to-one mapping of "fusion result-single sample".

[0083] For each time t within the prediction period and at spatial coordinates (x, y), key environmental variables (visibility, wind speed, etc.) are extracted, and all single-sample forecast values ​​are calculated to determine two core indicators: variance. :

[0084] Among them, f i Let F be the predicted environmental variable value of a single sample in the i-th sub-model, F be the predicted environmental variable value of the fused environmental state forecast field, and N be the number of single samples (20-50). The larger the variance, the greater the difference in single-sample forecasts, the higher the uncertainty of environmental element prediction, and the lower the corresponding confidence level. Key quantiles: Calculate the 2.5%, 10%, 20%, 80%, 90%, and 97.5% quantiles to construct three-level confidence intervals. The smaller the span of the confidence interval, the higher the corresponding confidence level.

[0085] Based on quantiles, a three-level confidence interval is constructed (60%: 20%~80% quantile, 80%: 10%~90% quantile, 95%: 2.5%~97.5% quantile); the proportion of single samples predicting key driving impact events (dense fog, strong crosswinds, etc.) in the statistical forecast set is used as the event probability; the confidence intervals, event probabilities, variances, and other information of each spatiotemporal location are integrated to generate a complete environmental probability forecast result, and the credibility of each environmental state is labeled.

[0086] S6, the environmental probability prediction results, vehicle perception results and map static information are fused together, and the fused comprehensive information is mapped to driving parameters to achieve intelligent driving control.

[0087] Furthermore, data preprocessing (i.e., standardization, spatiotemporal alignment, and validity verification) is performed on the environmental probability forecast results, vehicle perception results, and map static information. A basic credibility assessment is then conducted on the preprocessed environmental probability forecast results, vehicle perception results, and map static information to obtain the initial credibility of the probability forecast, vehicle perception, and map scene. Based on the current driving scenario information, the probability forecast weight coefficient, vehicle perception weight coefficient, and map scene weight coefficient are determined according to the initial credibility of the probability forecast, vehicle perception, and map scene. Finally, the probability forecast weight coefficient, vehicle perception weight coefficient, and map scene weight coefficient are determined based on the probability forecast weight coefficient, vehicle perception weight coefficient, and map scene weight coefficient. The coefficients are used to fuse the preprocessed environmental probability prediction results, vehicle perception results, and static map information. The fused comprehensive information is analyzed based on the time dimension, spatial dimension, and element dimension to obtain a multi-dimensional driving mapping element set, which includes various environmental elements of different dimensions and their corresponding confidence levels. Based on the various environmental elements of different dimensions and their corresponding confidence levels, multi-dimensional driving control parameters and driving strategy suggestions are obtained through a preset environmental element-driving parameter mapping rule base. Based on the environmental probability prediction results, driving commands are generated according to the multi-dimensional driving control parameters and driving strategy suggestions for intelligent driving control.

[0088] It should be understood that format standardization involves uniformly converting probability forecasts, vehicle perception results, and map static information (i.e., high-precision map static information) into a raster format of 0.5m×0.5m to 5m×5m. Probability forecasts retain probability / confidence interval information. Vehicle perception results specifically include information about traffic participants (such as the position, speed, and direction of movement of other vehicles, pedestrians, and non-motorized vehicles), road environment perception information (such as lane line type, traffic sign / marking recognition results, and road surface condition) and corresponding perception confidence levels. High-precision map static information specifically includes road geometric attributes (such as road centerline coordinates, lane width, road curvature, slope, and slope length) and road semantic attributes (such as road type, number of lanes, lane function division, fixed location and meaning of traffic signs / markings, intersection type, and bridge / tunnel markings). Due to the general... Probability forecasts use gridded data, vehicle perception uses point cloud / image data, and high-precision maps use vector data. This eliminates the heterogeneity of the original formats of these three types of data by uniformly converting them into a raster format, enabling compatibility and correlation among them, and providing a unified carrier for subsequent alignment. Spatiotemporal alignment: Based on the standardized raster carrier, using the vehicle as the origin and a unified spatiotemporal reference of UTC time + WGS-84 / local plane coordinates, accurate matching of timestamps and spatial rasters for the three types of data is achieved (ensuring that data from the same time and location can be correlated). Validity verification: The standardized and spatiotemporally aligned data undergoes quality screening, eliminating invalid data. Probability forecasts are verified for confidence interval integrity, vehicle perception results with confidence levels <0.7 are eliminated, and high-precision maps are verified for the integrity of key road elements, ensuring the reliability of the data input to the fusion process.

[0089] The credibility of the three types of data to be fused is quantified into initial values ​​of 0-1, with the assignment logic being "the higher the credibility, the larger the initial value". The specific value of the environmental element probability confidence level (the credibility corresponding to the probability forecast) is obtained through two steps: "uncertainty quantification calculation + scene adaptive calibration". The complete calculation chain and the assignment logic for other data are as follows: 1. Probability forecast (environmental element probability confidence level): First step, basic confidence level calculation (based on uncertainty quantification results): using variance and confidence interval as the core calculation indicators, first... The variance of environmental elements at each spatiotemporal point is obtained, and then the basic confidence level is determined by combining it with the span of the three-level confidence interval. Preset calculation rules: variance ≤ 50m² and 95% confidence interval span ≤ 30m → basic confidence level 0.9-1.0; variance 50-100m² and 95% confidence interval span 30-50m → basic confidence level 0.7-0.8; variance 100-200m² and 95% confidence interval span 50-80m → basic confidence level 0.5-0.6; variance > 200m² and 95% confidence interval span > 80m → Base confidence level 0.2-0.4; Step 2, Scene Adaptive Calibration: Fine-tune the base confidence level based on the driving scenario, increasing it by 0.05-0.1 for highway straight sections and decreasing it by 0.05-0.1 for complex urban intersections, finally obtaining the specific values ​​of the probability confidence levels of environmental elements (range 0.2-1.0); Example: Visibility forecast variance 45 m², 95% confidence interval span 25 m, base confidence level 0.95, calibrated to 1.0 in highway scenarios; 2. Vehicle Perception Results: Combined with equipment precision... For the degree of interference and real-time environmental interference, first set the equipment baseline value (0.9 for LiDAR, 0.8 for camera). Maintain the baseline value when there is no interference. Adjust the value by 0.1-0.2 for general interference and 0.3-0.5 for severe interference. Example: Assign a value of 0.9 for camera vehicle recognition in clear weather and adjust it to 0.5 in dense fog. 3. High-precision map: Assign a value of 0.9-1.0 for normal scenes and adjust it to 0.4-0.6 when the deviation from real-time perception is >1m. Example: Assign a value of 0.95 for normal road scenes and 0.5 for construction sections with excessive deviation.

[0090] Based on the current driving scenario, the initial confidence value is corrected. For highway straight sections, the probability forecast weight is increased (to improve the priority of pre-decision) and the vehicle perception weight is slightly reduced. For complex urban intersections, the vehicle perception weight is increased (to ensure real-time performance) and the probability forecast weight is decreased. The weight of high-precision map remains unchanged. Weight normalization: the weight coefficient is calculated using the formulas w1'=w1 / (w1+w2+w3), w2'=w2 / (w1+w2+w3), and w3'=w3 / (w1+w2+w3) (where w1 is the confidence value after the probability forecast adjustment, w2 is the vehicle perception, and w3 is the high-precision map) to ensure that the weight sum is 1.

[0091] Raster-level fusion and result integration: 1. Weighted fusion: Fusion of three types of data attributes grid-by-grid according to normalized weights, formula: Environmental dynamic information Eij =w1'×E1 ij +w2'×E2 ij (E1) ij E2 is the median of the probability forecast. ij (Perceived value), speed V of traffic participants ij =w2'×V2 ij +w1'×V1 ij (V2) ij To sense speed, V1 ij For forecast correction speed), road static information R ij =w3'×R3 ij +w2'×R2 ij (R3) ij For map values, R2 ij 1. Perceived value); 2. Conflict resolution: Adopt the principle of "credibility priority + physical rule verification" to resolve data attribute conflicts, and combine physical rules such as road width to judge when the weights are close; 3. Information integration: Add credibility identifier (mark the contribution ratio of each data) and integrate it into comprehensive environmental-road information (i.e., the integrated information after fusion, including environmental dynamics, traffic participants, road static attributes and credibility identifier).

[0092] By analyzing comprehensive environmental and road information, three core elements are extracted: Time dimension: continuous time slice data for the next 15-30 minutes; Spatial dimension: lane-level grid information 500 meters ahead of the vehicle; Element dimension: environmental variables such as visibility, wind speed, and road surface adhesion coefficient, and their corresponding confidence levels, as well as the speed / position of traffic participants, road type / number of lanes, etc., to obtain a multi-dimensional driving mapping element set (i.e., a standardized driving parameter mapping input element set). The multi-dimensional driving mapping element set includes various environmental elements of different dimensions and their corresponding confidence levels. By using a pre-defined environmental element-driving parameter mapping rule base and combining it with a multi-dimensional driving mapping element set (including environmental variables and confidence levels) of real-time analysis of comprehensive environmental and road information, the corresponding control parameter adjustment scheme is matched. If the confidence level of environmental elements in the probability forecast results is ≥80%, precise parameter adjustment is performed; if the confidence level is 80%~60%, conservative parameter adjustment is performed; if the confidence level is <60%, no adjustment is performed.

[0093] The specific operation and mapping rules are as follows, covering control needs across all scenarios: 1. Longitudinal Control Mapping: Establish a mapping rule for "road adhesion coefficient - following distance / braking force", and match parameters according to the predicted road adhesion coefficient value; Rule example: Adhesion coefficient ≥ 0.8 (dry road surface, confidence level ≥ 80%) → following distance 2s, braking force gradient 5m / s²; Adhesion coefficient 0.4~0.8 (wet road surface, confidence level ≥ 80%) → following distance 3s, braking force gradient 3m / s²; Adhesion coefficient < 0.4 (icy and snowy road surface, confidence level ≥ 80%) → following distance 5s, braking force gradient 2m / s²; Operation: Based on the analyzed predicted road adhesion coefficient value, automatically match the corresponding following distance and braking force gradient to achieve smooth adjustment.

[0094] 2. Lateral Control Mapping: Establish a mapping rule for "crosswind level - steering assist / EPS compensation parameters", matching parameters according to the crosswind forecast wind speed and direction; Rule example: Crosswind speed < 5m / s (no impact, confidence ≥ 80%) → default steering assist, EPS compensation 0 N·m; Wind speed 5~15m / s (slight crosswind, confidence ≥ 80%) → steering assist increased by 10%, EPS compensation 0.5~1 N·m; Wind speed ≥ 15m / s (strong crosswind, confidence ≥ 80%) → steering assist increased by 20%, EPS compensation 1~2 N·m (fine-tuning compensation direction according to wind direction); Operation: After parsing the crosswind forecast information, match the corresponding steering assist curve and EPS compensation parameters to suppress vehicle lateral deviation.

[0095] 3. Lighting Visibility Control Mapping: Establish a mapping rule for "visibility / precipitation intensity - lighting mode / wiper speed" and match it according to the corresponding environmental element thresholds; Rule examples: Visibility ≥ 200m (normal visibility, confidence ≥ 80%) → low beam headlights, wipers off; Visibility 100~200m (light fog, confidence ≥ 80%) → low beam headlights + side marker lights, wipers on intermittent setting (when precipitation intensity ≥ 2mm / h); Visibility < 100m (dense fog, confidence ≥ 80%) → fog lights + low beam headlights + side marker lights, wipers on slow setting; Precipitation intensity ≥ 5mm / h (moderate to heavy rain, confidence ≥ 80%) → wipers on fast setting; Operation: Automatically switch lighting modes and adjust wiper speeds based on the analyzed visibility and precipitation intensity forecast values.

[0096] 4. Chassis Power Control Mapping: Establish a mapping rule for "Road Condition - Chassis Mode / Power Output" and match parameters according to road condition forecasts; Rule example: Road condition is dry paved road (confidence ≥ 80%) → Sport mode, power output 100%; Road condition is bumpy road (confidence ≥ 80%) → Comfort mode, power output 80%; Road condition is soft road (sand / mud, confidence ≥ 80%) → Off-road mode, power output 60% (limited torque); Operation: Based on the analyzed road condition forecast value, switch the corresponding chassis mode and limit the upper limit of power output through the ECU.

[0097] 5. Strategy-level control mapping: Establish a mapping rule for "high-confidence environmental events - strategy suggestions", which is triggered only when the confidence level of the environmental event is ≥90%; Rule example: Forecast dense fog on the road ahead in the next 10 minutes (visibility <50m, confidence level ≥90%) → output "leave the current road section and enter the nearest service area" strategy; Forecast road construction on the road ahead in the next 15 minutes (confidence level ≥90%) → output "change lanes to the left lane in advance" strategy; Operation: After identifying a high-confidence environmental event, automatically output the corresponding lane change, leave, and other strategy suggestions, and simultaneously match auxiliary control parameters (such as steering angle and vehicle speed stability range when changing lanes).

[0098] In this embodiment, based on the targeted mapping of the analyzed elements, the control requirements of the entire scenario are covered: longitudinal control: adjusting the following distance and braking force gradient according to the road adhesion coefficient forecast; lateral control: fine-tuning the steering assist curve or EPS compensation parameters according to the crosswind forecast; headlight visibility control: switching the headlight mode and adjusting the wiper speed according to the visibility / precipitation intensity forecast; chassis dynamics control: adjusting the chassis mode and limiting power output according to the road condition forecast; strategy-level control: outputting lane change and departure strategy suggestions under high confidence (≥90%) environmental forecasts, obtaining multi-dimensional driving control parameters and driving strategy suggestions (including vehicle speed, steering angle, braking force, etc.).

[0099] It should also be noted that, based on the confidence level of probability forecasts, priority sorting and reliability screening are performed according to the logic of "element matching + threshold verification". Specifically, the driving control parameters and strategy suggestions are first matched one by one with the confidence levels of the probability forecasts of the corresponding environmental elements (e.g., "turn on low beam headlights" corresponds to the confidence level of the "visibility" element, and "adjust following distance" corresponds to the confidence level of the "road adhesion coefficient" element). Then, priorities are assigned according to the confidence level (the higher the confidence level, the higher the priority of the control parameters and strategies). At the same time, three levels of threshold screening are set: confidence level ≥ 80% is high priority and is directly included in the final control command; confidence level 60%~80% is medium priority, conservative adjustments are performed and included in the command; confidence level < 60% is low priority, information is only recorded and continuously monitored, and no active adjustments are performed. Finally, the high and medium priority control parameters are integrated to generate driving control commands that the vehicle actuators can understand.

[0100] In this embodiment, multi-source raw data within a preset range of the currently driven vehicle is first collected in real time and preprocessed to obtain the environmental observation vector field at the current moment. Then, the environmental observation vector field at the current moment is assimilated with the latest short-term forecast field using 4D-Var to obtain the current environmental analysis field. The current environmental analysis field is then perturbed by EnKF to obtain the EnKF perturbation analysis set. Subsequently, the EnKF perturbation analysis set and the SPDE-INLA physical field are input into the HSTAN model group to output the environmental state forecast field within the prediction time range. The HSTAN model group includes multiple HSTAN sub-models constructed based on the physical constraint U-RNN module. Finally, the uncertainty of the environmental state forecast field is quantified to generate the environmental probability forecast result. The environmental probability forecast result, vehicle perception result, and map static information are fused, and the fused comprehensive information is mapped to driving parameters to achieve intelligent driving control. This embodiment constructs a closed-loop architecture of perception-prediction-decision-execution. It achieves environmental state prediction within a future time range through ultra-fast data assimilation, EnKF perturbation, and HSTAN model set. Then, by combining vehicle perception results and static map information, it realizes accurate and orderly transformation from environmental information to control parameters, realizes intelligent driving control, and thus improves the reliability of intelligent driving.

[0101] Reference Figure 3 , Figure 3 This is a structural block diagram of the first embodiment of the intelligent driving control system based on environmental perception of the present invention.

[0102] like Figure 3 As shown, the intelligent driving control system based on environmental perception proposed in this embodiment of the invention includes: The data processing module 3001 is used to collect multi-source raw data within a preset range of the currently driven vehicle in real time, and to preprocess the multi-source raw data to obtain the environmental observation vector field at the current moment. The 4D-Var assimilation module 3002 is used to perform 4D-Var assimilation on the current environmental observation vector field and the latest short-term forecast field to obtain the current environmental analysis field; EnKF perturbation module 3003 is used to perturb the current environmental analysis field with EnKF to obtain an EnKF perturbation analysis set; HSTAN inference module 3004 is used to input the EnKF perturbation analysis set and SPDE-INLA physical field into the HSTAN model group and output the environmental state forecast field within the prediction time range. The HSTAN model group includes multiple HSTAN sub-models constructed based on the physical constraint U-RNN module. Uncertainty quantification module 3005 is used to perform uncertainty quantification processing on the environmental state prediction field and generate environmental probability prediction results; The data processing module 3001 is also used to fuse the environmental probability prediction results, vehicle perception results and map static information. The driving control module 3006 is used to map driving parameters to the fused integrated information in order to achieve intelligent driving control.

[0103] Other embodiments or specific implementations of the intelligent driving control system based on environmental perception of the present invention can be referred to the above-described method embodiments, and will not be repeated here.

[0104] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0105] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0106] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0107] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for intelligent driving control based on environment perception, characterized in that, The method comprises the following steps: Real-time acquisition of multi-source original data within a preset range of the current driving vehicle, and preprocessing of the multi-source original data to obtain an environment observation vector field at the current time; 4D-Var assimilation of the environment observation vector field at the current time and the latest short-term prediction field to obtain a current environment analysis field; EnKF perturbation of the current environment analysis field to obtain an EnKF perturbed analysis set; Inputting the EnKF perturbed analysis set and the SPDE-INLA physical field into an HSTAN model group to output an environment state prediction field within a prediction time range, wherein the HSTAN model group comprises a plurality of HSTAN sub-models constructed based on a physically constrained U-RNN module; Uncertainty quantification of the environment state prediction field to generate an environment probability prediction result; Data fusion of the environment probability prediction result, vehicle perception result and map static information, and driving parameter mapping of the fused comprehensive information to realize intelligent driving control.

2. The method of claim 1, wherein, The 4D-Var assimilation of the environment observation vector field at the current time and the latest short-term prediction field to obtain a current environment analysis field comprises: Taking the latest short-term prediction field as a background field, constructing a target function based on the environment observation vector field at the current time and the background field; The target function is: wherein is the environmental analysis variable to be solved, is the background field variable, is the background error covariance matrix, is the observation vector, is the observation operator, is the observation error covariance matrix; Solving the environment analysis variable by minimizing the target function to obtain the current environment analysis field.

3. The method of claim 1, wherein, The EnKF perturbation of the current environment analysis field to obtain an EnKF perturbed analysis set comprises: Based on the EnKF algorithm, adding a Gaussian distributed perturbation term to the current environment analysis field to generate a plurality of perturbed analysis samples; The EnKF algorithm comprises: wherein is the ith perturbation analysis sample, is the ith random perturbation vector, is the total number of perturbation analysis samples; Constructing an EnKF perturbed analysis set according to the plurality of perturbed analysis samples.

4. The method of claim 1, wherein, The inputting of the EnKF perturbed analysis set and the SPDE-INLA physical field into the HSTAN model group to output an environment state prediction field within a prediction time range comprises: Inputting the EnKF perturbed analysis set and the SPDE-INLA physical field into the HSTAN model group; Transmitting each EnKF perturbed analysis sample in the EnKF perturbed analysis set to each HSTAN sub-model; Performing time series prediction and feature fusion on the corresponding EnKF perturbed analysis sample and the SPDE-INLA physical field by the physically constrained U-RNN module in each HSTAN sub-model to obtain a single-sample environment state prediction result; Average fusion of the single-sample environment state prediction results output by all HSTAN sub-models to output an environment state prediction field within a prediction time range.

5. The method of claim 1, wherein, The uncertainty quantification of the environment state prediction field to generate an environment probability prediction result comprises: Temporal and spatial dimension alignment of the environment state prediction field and each sample environment state prediction result; Based on the environment state prediction field, extracting the predicted value of each time-space position from each sample environment state prediction result according to the mapping relationship of the environment variable of each time-space position within the prediction time range; Calculating the variance and quantile based on the predicted value of each time-space position; According to the quantile, a confidence interval and an event probability are determined; The variance, the quantile, the confidence interval and the event probability are data integrated to obtain an environmental probability prediction result.

6. The method of claim 1, wherein, The data fusion of the environmental probability prediction result, the vehicle perception result and the map static information comprises: The data preprocessing of the environmental probability prediction result, the vehicle perception result and the map static information is performed; The environmental probability prediction result, the vehicle perception result and the map static information after the preprocessing are subjected to basic credibility evaluation to obtain a probability prediction initial credibility, a vehicle perception initial credibility and a map scene initial credibility; Based on current driving scene information, the probability prediction initial credibility, the vehicle perception initial credibility and the map scene initial credibility are used to determine a probability prediction weight coefficient, a vehicle perception weight coefficient and a map scene weight coefficient; The data fusion of the environmental probability prediction result, the vehicle perception result and the map static information after the preprocessing is performed according to the probability prediction weight coefficient, the vehicle perception weight coefficient and the map scene weight coefficient.

7. The method of claim 1, wherein, The driving parameter mapping of the fused comprehensive information is performed to realize intelligent driving control, which comprises: The fused comprehensive information is analyzed based on time dimension, space dimension and element dimension to obtain a multi-dimensional driving mapping element set, wherein the multi-dimensional driving mapping element set comprises environmental elements of multiple different dimensions and corresponding credibilities; According to the environmental elements of multiple different dimensions and the corresponding credibilities, multi-dimensional driving control parameters and driving strategy suggestions are obtained through a preset environmental element-driving parameter mapping rule library; Based on the environmental probability prediction result, driving instructions are generated according to the multi-dimensional driving control parameters and the driving strategy suggestions to realize intelligent driving control.

8. An environment perception based intelligent driving control system, characterized in that, The system comprises: A data processing module is configured to collect multi-source original data within a preset range of a current driving vehicle in real time, and to preprocess the multi-source original data to obtain an environment observation vector field at a current time; A 4D-Var assimilation module is configured to perform 4D-Var assimilation on the environment observation vector field at the current time and a latest short-term prediction field to obtain a current environment analysis field; An EnKF perturbation module is configured to perform EnKF perturbation on the current environment analysis field to obtain an EnKF perturbed analysis set; An HSTAN inference module is configured to input the EnKF perturbed analysis set and a SPDE-INLA physical field into an HSTAN model group to output an environmental state prediction field within a prediction time range, wherein the HSTAN model group comprises a plurality of HSTAN sub-models constructed based on a physical constraint U-RNN module; An uncertainty quantification module is configured to perform uncertainty quantification processing on the environmental state prediction field to generate an environmental probability prediction result; The data processing module is further configured to perform data fusion on the environmental probability prediction result, a vehicle perception result and a map static information; A driving control module is configured to perform driving parameter mapping on the fused comprehensive information to realize intelligent driving control.

9. An environment perception based intelligent driving control device, characterized by, The device comprises a memory, a processor, and an environment-aware intelligent driving control program stored on the memory and executable on the processor, and the environment-aware intelligent driving control program is configured to implement the steps of the environment-aware intelligent driving control method according to any one of claims 1 to 7.

10. A storage medium, characterized by The storage medium stores an environment-aware intelligent driving control program, and the environment-aware intelligent driving control program, when executed by a processor, implements the steps of the environment-aware intelligent driving control method according to any one of claims 1 to 7.

Images