Automatic driving lane changing decision method based on risk assessment

Abstract

The application relates to the technical field of automatic driving and discloses an automatic driving lane-changing decision-making method based on risk assessment. The method comprises the following steps: according to real-time data of a surrounding environment of a vehicle, fusing multi-dimensional features to generate a space-time feature representation containing position, speed and environmental information, and comprehensively perceiving a complex traffic environment. Then, the space-time feature representation is input into a risk assessment model, the abnormal risk in the lane-changing process is probabilistically modeled, and a risk feature vector with a confidence degree is output, thereby providing a fine risk reference for decision-making. Then, a risk distribution field is constructed based on the risk feature vector, and a sequence prediction model is used to predict a propagation path of the risk distribution field, thereby prospectively perceiving dynamic changes of the risk. Finally, according to the prediction result, an adaptive decision threshold generation model is constructed, the real-time risk feature vector is compared with multi-level decision thresholds, and accurate and flexible automatic driving lane-changing decision-making is realized, thereby improving the safety and adaptability of the lane-changing process.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to an autonomous driving lane-changing decision-making method based on risk assessment. Background Technology

[0002] With the rapid development of intelligent transportation technology, autonomous vehicles have gradually become a hot topic in research and application. Lane-changing decisions, as one of the key functions of autonomous driving systems, directly affect the safety, efficiency, and comfort of vehicle operation. In complex and ever-changing traffic environments, vehicles need to perceive surrounding environmental information in real time, accurately assess potential risks, and make reasonable lane-changing decisions.

[0003] Existing autonomous driving lane-changing decision-making methods partially rely on a single sensor to acquire environmental data, making it difficult to comprehensively capture dynamic changes around the vehicle. For example, relying solely on visual sensors may lead to perception biases in adverse weather conditions, while relying solely on radar sensors struggles to accurately identify target categories, resulting in insufficient completeness and accuracy of environmental information. Furthermore, in terms of feature processing, traditional methods often lack effective fusion of temporal and spatial features, failing to form a coherent spatiotemporal feature representation, thus making the judgment of surrounding vehicle movement trends inaccurate.

[0004] In the risk assessment phase, existing technologies mostly rely on empirical rules or simple threshold judgments, lacking probabilistic modeling of abnormal risks during lane changes. This results in relatively coarse risk assessment results, making it difficult to quantify the probability of occurrence and the degree of impact of different risks, and failing to provide refined risk basis for decision-making. Furthermore, risks during lane changes are not static but dynamically propagate over time and space. Existing methods lack effective prediction of the dynamic propagation patterns of risks, often only addressing the risk state at the current moment, and failing to proactively avoid potential risk evolution.

[0005] Setting the decision threshold is a crucial step in lane-changing decisions. Existing methods often use fixed thresholds or thresholds preset based on limited scenarios, failing to dynamically adjust according to real-time driving conditions and environmental changes. In complex traffic scenarios, fixed thresholds can easily lead to overly conservative or aggressive decisions: conservative thresholds may miss reasonable lane-changing opportunities, reducing driving efficiency; aggressive thresholds may ignore potential risks, increasing the risk of collisions. These issues mean that existing lane-changing decision-making methods still have room for improvement in terms of adaptability, accuracy, and safety, making it difficult to meet the needs of autonomous driving in diverse traffic environments. Summary of the Invention

[0006] The purpose of this invention is to provide a risk assessment-based lane-changing decision-making method for autonomous driving, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a lane-changing decision-making method for autonomous driving based on risk assessment, the method comprising:

[0008] Based on real-time data of the vehicle's surrounding environment, multi-dimensional features are fused to generate a spatiotemporal feature representation that includes location, speed, and environmental information.

[0009] The spatiotemporal features are input into the risk assessment model to probabilistically model the abnormal risks during lane changing and output a risk feature vector with confidence.

[0010] Based on the risk feature vector, a risk distribution field is constructed, and the propagation path of the risk distribution field is predicted based on a sequence prediction model;

[0011] Based on the prediction results of the risk distribution field, an adaptive decision threshold generation model is constructed. Through the dynamic game between the normal driving generator and the real-time discriminator, a multi-level decision threshold that changes with the driving state is generated. Based on the comparison results between the real-time risk feature vector and the multi-level decision threshold, the autonomous driving lane-changing decision is realized.

[0012] Preferably, the step of fusing multi-dimensional features based on real-time data of the vehicle's surrounding environment to generate a spatiotemporal feature representation including location, speed, and environmental information includes:

[0013] Based on real-time data collected by multiple sensors, timestamp alignment and missing value filling are performed. A dynamic time warping algorithm is used to eliminate differences in sensor sampling frequencies to obtain a time-synchronized multi-source data sequence. The real-time data includes vehicle position, speed, and environmental data.

[0014] The multi-source data sequence is input into the feature fusion network, where convolutional operations are used to capture road structure features in the spatial dimension and sliding windows are used to extract time series features in the temporal dimension to obtain preliminary spatiotemporal features.

[0015] For the initial spatiotemporal features, the correlation weights between different sensor data are calculated using an attention mechanism. The features are then weighted and fused according to the weights to highlight the contribution of key sensor data, resulting in weighted spatiotemporal features.

[0016] The weighted spatiotemporal features are input into the dimensionality reduction model, and redundant information is removed through nonlinear transformation, while key features are retained to generate a low-dimensional spatiotemporal feature representation containing position, velocity, and environmental information.

[0017] Preferably, the step of inputting the spatiotemporal feature representation into the risk assessment model to probabilistically model the abnormal risks during lane changing and outputting a risk feature vector with confidence includes:

[0018] The spatiotemporal feature representation is input into a multilayer network and mapped to a high-dimensional latent space through nonlinear transformation to obtain a deep feature representation;

[0019] For the aforementioned deep feature representation, a probabilistic regression model is constructed, and a combination of kernel functions is used to capture the nonlinear relationship and periodic changes in driving state, resulting in a probabilistic feature representation;

[0020] Based on the probabilistic feature representation, the predicted mean and variance of each feature point are calculated, and the model uncertainty is quantified by the inference algorithm to obtain the probabilistic feature distribution with confidence intervals.

[0021] For the aforementioned probability feature distribution, a distance detection algorithm is used to calculate the distance between each feature point and the normal driving distribution. Abnormal feature points are filtered according to a preset confidence threshold to generate a risk feature vector with confidence.

[0022] Preferably, the step of constructing a risk distribution field based on the risk feature vector and predicting the propagation path of the risk distribution field based on a sequence prediction model includes:

[0023] Based on the risk feature vector and combined with road topology information, each risk feature point is mapped to a two-dimensional coordinate system. An interpolation algorithm is then used to spatially interpolate the discrete risk feature points to generate a preliminary risk distribution field.

[0024] Based on the preliminary risk distribution field, combined with time dimension information, a spatiotemporal interpolation algorithm is used to dynamically correct the risk distribution field, resulting in a risk distribution field that changes over time.

[0025] The time-varying risk distribution field is input into the sequence prediction model to construct a road node graph model. An attention mechanism is used to capture the risk propagation dependency between nodes to obtain the initial risk propagation prediction.

[0026] Based on the initial risk propagation prediction and combined with historical risk propagation data, the risk propagation path and diffusion trend at future time steps are predicted, generating a propagation prediction result for a dynamic risk distribution field.

[0027] Preferably, based on the prediction results of the risk distribution field, an adaptive decision threshold generation model is constructed. Through a dynamic game between the normal driving generator and the real-time discriminator, a multi-level decision threshold that changes with the driving state is generated, including:

[0028] Based on historical normal driving data, a normal driving generator based on generative adversarial networks is trained to generate simulated data that conforms to the characteristics of safe driving.

[0029] Construct a real-time discriminator, input the predicted results of the risk distribution field and the simulated data generated by the normal driving generator, learn the boundary features of normal and abnormal states through dynamic game theory, and output the discrimination result and its confidence level.

[0030] Based on the discrimination results, a clustering algorithm is used to divide the discrimination results into multiple levels. Combined with the dynamic changes in the risk distribution field, a multi-level decision threshold that changes with the driving state is generated. The multi-level decision threshold includes low-risk, medium-risk, and high-risk thresholds.

[0031] Preferably, the step of making autonomous driving lane-changing decisions based on the comparison results of real-time risk feature vectors and multi-level decision thresholds includes:

[0032] The real-time risk feature vector is compared with a multi-level decision threshold, and the decision level is dynamically adjusted by the decision model based on the confidence level of the risk features and the degree of deviation from the threshold.

[0033] Based on the decision level, a lane-changing instruction or a command to maintain the current lane is generated to realize autonomous driving lane-changing decision-making.

[0034] Preferably, the method further includes:

[0035] When the deviation between real-time data and predicted data continues to exceed a set number of times, the phase delay feature of the feature points is extracted.

[0036] Adjust the parameter weight ratios in the risk assessment model according to the phase delay characteristics;

[0037] The corrected model parameters are stored in the historical database as the baseline values ​​for the next model initialization.

[0038] Preferably, the method further includes:

[0039] Construct a multi-dimensional decision parameter set that includes speed fluctuation entropy, environmental disturbance sensitivity, and road friction coefficient;

[0040] A status alert is triggered when a single parameter exceeds the first-level threshold, and an emergency evasion command is triggered when the combined effect of at least two parameters exceeds the second-level threshold.

[0041] Preferably, the calculation of the velocity fluctuation entropy value includes:

[0042] Perform frequency domain transformation on velocity data within a specified time window to extract energy distribution characteristics of a preset frequency band;

[0043] The speed fluctuation entropy value is calculated based on the energy distribution characteristics. When the proportion of low-frequency energy exceeds the preset proportion, it is determined to be a traffic flow risk mode.

[0044] Preferably, the method further includes:

[0045] When a traffic flow risk pattern is identified, acceleration data of surrounding vehicles is collected simultaneously.

[0046] The acceleration data is correlated with the low-frequency energy. If the correlation coefficient is greater than the preset correlation value, a path adjustment task is added to the lane change decision.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] This risk assessment-based autonomous driving lane-changing decision-making method significantly improves the adaptability and accuracy of lane-changing decisions through multi-dimensional technical optimization. In the environmental perception and feature processing stage, the method performs multi-dimensional feature fusion based on real-time data of the vehicle's surrounding environment. This integrates position, speed, and environmental information acquired from different sensors to form a comprehensive spatiotemporal feature representation. This fusion approach overcomes the limitations of single-sensor perception, preserving both the relative positional relationships between the vehicle and surrounding targets in the spatial dimension and capturing changes in motion trends in the temporal dimension. This makes the system's perception of complex traffic environments more comprehensive and coherent, providing a solid foundation for subsequent risk assessment.

[0049] In the risk assessment phase, spatiotemporal features are input into the risk assessment model for probabilistic modeling of abnormal risks, outputting a risk feature vector with confidence levels. This probabilistic modeling approach can quantitatively describe various abnormal risks that may occur during lane changes, reflecting the reliability of the risk assessment results through confidence indices, thus avoiding the crudeness of traditional empirical rules or simple threshold judgments. The output of the risk feature vectors allows for clear differentiation of different types and levels of risks, providing a refined risk reference for the decision-making system and helping to more accurately identify potential hazards.

[0050] By constructing a risk distribution field based on risk feature vectors and predicting its propagation path through a sequence prediction model, the system achieves a forward-looking perception of dynamic risk changes. The risks during lane changing are not static but evolve continuously with vehicle movement and the behavior of surrounding targets. The risk distribution field visually presents the spatial distribution of risks, while propagation path prediction reveals the development trend of risks over time. This allows the system to move beyond passively responding to the current risk state and anticipate risk changes over a future period, providing sufficient reaction time for decision-making and enhancing the proactivity and predictability of lane-changing decisions.

[0051] An adaptive decision threshold generation model, constructed based on risk distribution field prediction results, generates multi-level decision thresholds through a dynamic game between a normal driving generator and a real-time discriminator, thus addressing the insufficient adaptability of traditional fixed thresholds. The dynamic game process dynamically adjusts the decision thresholds in real time, incorporating current driving status, environmental characteristics, and risk prediction results: in low-risk scenarios, the threshold can be appropriately relaxed to improve lane-changing efficiency; in high-risk scenarios, the threshold is tightened accordingly to ensure safety. This multi-level decision threshold, which changes with driving status, allows lane-changing decisions to flexibly adapt to different traffic environments, achieving a better balance between safety and efficiency. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the working principle of the autonomous driving lane-changing decision-making method based on risk assessment described in this invention.

[0053] Figure 2 A flowchart for generating spatiotemporal feature representations through multi-dimensional feature fusion;

[0054] Figure 3 A flowchart for outputting risk feature vectors with confidence scores for a risk assessment model;

[0055] Figure 4 A flowchart for constructing a risk distribution field and predicting propagation paths. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Please see Figure 1 The present invention provides a lane-changing decision-making method for autonomous driving based on risk assessment, the method comprising:

[0058] By integrating real-time data from the vehicle's surrounding environment, a multi-dimensional feature fusion is performed to generate a spatiotemporal feature representation encompassing position, speed, and environmental information. This spatiotemporal feature representation is then input into a risk assessment model, which uses probabilistic modeling to evaluate abnormal risks during lane changes, outputting a risk feature vector with confidence levels. Based on this risk feature vector, a risk distribution field is constructed, and a sequence prediction model is used to predict the propagation path of this risk distribution field. For the predicted risk distribution field, a dynamic game between a normal driving generator and a real-time discriminator generates multi-level decision thresholds that change with the driving state. Finally, by combining the comparison results of the real-time risk feature vector and the multi-level decision thresholds, adaptive control of lane-changing decisions for autonomous driving is achieved. This process ensures timely response to risk changes in dynamic driving environments, optimizing the safety of lane-changing behavior.

[0059] Example 1: See Figure 2 The multi-source sensor system collects vehicle position, speed, and environmental data in real time. Position data comes from a GPS and inertial measurement unit (IMU) combined positioning system; speed data integrates longitudinal / lateral acceleration information transmitted from wheel speed sensors and the vehicle bus; environmental data includes a millimeter-wave radar target list, LiDAR point cloud, camera semantic segmentation results, and V2X communication data. This raw data first undergoes timestamp alignment processing. A hardware clock synchronization mechanism matches the time reference values ​​of each sensor, and cubic spline interpolation is used to fill in missing data due to transmission delays at the corresponding timestamps. A dynamic time warping algorithm handles sampling frequency differences: the LiDAR data sampling rate of 20Hz and the camera data stream of 30Hz are aligned along the time axis of non-uniform sampling points by constructing a time-series distance matrix and searching for the minimum-cost path, generating a time-synchronized multi-source data sequence.

[0060] Multi-source data sequences are input into a feature fusion network for processing. The spatial dimension processing unit loads a road grid map as the initial weights for the convolutional kernels and extracts spatial structure features through three layers of convolutional operations: the first layer uses a 5×5 convolutional kernel to rasterize a 50-meter radius around the vehicle, with each grid storing the probability of environmental objects; the second layer uses a 3×3 dilated convolution to expand the receptive field and identify lane line topological association features; the third layer applies a 1×1 convolution to compress the channel dimension, generating a 256×256 resolution spatial feature map. The temporal dimension processing employs a dual sliding window mechanism: a 50-millisecond short window calculates the rate of change of velocity and acceleration statistics, while a 3-second long window performs autoregressive integral moving average modeling to extract trend fluctuation features. In the spatiotemporal cross-processing stage, the spatial feature map and time-series statistics are input into the feature concatenation layer step-by-step to generate a preliminary spatiotemporal feature tensor containing the vehicle's kinematic state and the positions of static obstacles.

[0061] The attention mechanism constructs a sensor weight allocation system on this feature tensor. A trainable feature correlation matrix is ​​created, with dimensions matching the number of sensors. Initial weights are generated by calculating the cosine similarity of the sensor feature vectors. These weights are input into a gated recurrent unit and dynamically modulated based on the current vehicle speed and road curvature: when the GPS positioning accuracy factor is below a threshold, the millimeter-wave radar weight is increased; in rainy scenarios, the camera weight decreases while the LiDAR weight increases. The modulated weight matrix is ​​then normalized using Softmax and weighted fusion is performed on the multi-source sensor inputs. Specifically, environmental data feature vectors are multiplied by their corresponding weights and then summed element-wise, while kinematic features are weighted averaged.

[0062] The dimensionality reduction model achieves feature compression based on a variational autoencoder architecture. The encoder consists of three fully connected layers: the first layer projects 2048-dimensional weighted features to 1024 dimensions using the LeakyReLU activation function; the second layer compresses to 512 dimensions and applies batch normalization to prevent gradient shift; the third layer generates a 256-dimensional mean vector and log-variance vector. Reparameterization is performed on the latent variable space, sampling to obtain a 128-dimensional feature vector. The decoder reconstructs the original features using a symmetric structure, minimizing the reconstruction loss function to constrain information integrity. In the key feature extraction stage, the latent variable vector is input into the feature selection module, which calculates the mutual information value between each dimension and the vehicle's position change, retaining the first 64 features with mutual information values ​​higher than 0.15, and setting the remaining dimensions to zero. The final output low-dimensional spatiotemporal feature representation includes 32-dimensional position features (XY coordinates, heading angle), 24-dimensional velocity features (vectorized components), and 8-dimensional environmental features (obstacle density, visibility distance), for a total of 64 dimensions, a reduction from the original data.

[0063] The data processing chain is configured with a real-time feedback mechanism. The output feature represents the input anomaly detector, calculates the reconstruction error value, and triggers a sensor recalibration signal when the error exceeds three standard deviations from the historical mean for five consecutive frames. This signal controls the dynamic time warping algorithm to reinitialize the time warping path and reset the convolutional kernel weights in the feature fusion network. The decoder output of the dimensionality reduction model synchronously inputs the quality assessment module. When the positional feature reconstruction deviation exceeds 0.3 meters, a redundant dimension preservation strategy is activated, temporarily expanding the feature dimension to 96 dimensions until the deviation returns to the normal range. The entire processing flow executes cyclically with a 100-millisecond cycle, with processing latency strictly controlled within 20 milliseconds.

[0064] The persistent storage of spatiotemporal feature representations employs an incremental update mechanism. Cluster analysis is triggered every 100 frames of feature data generated, using the K-means algorithm to identify cluster centers of feature distribution patterns and replacing the reference feature vectors in the historical database. During online runtime, real-time similarity comparisons are performed between the feature vectors and database patterns. When unknown patterns are encountered, the storage area is automatically expanded and new driving scenarios are labeled. This database is connected to the parameter adjustment channel of the risk assessment model; the statistical characteristics of the feature distribution directly affect the variance calculation parameters in the model.

[0065] The multi-source sensor fusion system is equipped with fault protection logic. When the millimeter-wave radar fails to detect an obstacle for 10 consecutive frames, the camera feature weight is automatically increased by 20%; if the lidar point cloud density drops sharply, a V2X data replacement strategy is activated. The feature fusion network has a built-in self-checking mechanism that periodically generates virtual data to verify the operating status of each channel, and abnormal channels are automatically switched to backup computing nodes. The output data is finally encapsulated into a standardized data structure, including timestamp checksums, feature dimension units, and completeness marker fields, and is transmitted to the risk assessment module via shared memory.

[0066] Example 2: See Figure 3 The spatiotemporal feature representation is input into a multi-layer network processing system. This network employs a residual connection structure, containing five fully connected layers and three gating units. The first layer receives a 64-dimensional feature vector, expands the feature dimension through 512 neurons, and applies the ELU activation function to handle nonlinear relationships. The output features are normalized and then input into the second layer, where the number of neurons is reduced to 256, and the hyperbolic tangent activation function is used to modulate the feature amplitude. The third layer introduces a gating linear unit, dynamically adjusting the feature gating weights based on the vehicle's current acceleration value: when the absolute acceleration value exceeds 0.3g, the time dimension weight is increased by 25%. The fourth layer performs feature cross-operation, generating an interactive product term of position and velocity features to enhance the coupling relationship of motion states. The fifth layer outputs a 256-dimensional deep feature representation, whose information entropy is increased by 60% compared to the input features.

[0067] The probabilistic regression model constructs a multidimensional probability distribution on deep feature representation. The kernel function combination system integrates two types of core processors: Gaussian radial basis kernels and periodic kernels. The Gaussian kernel is configured with adjustable bandwidth parameters and automatically sets initial values ​​based on the standard deviation of the feature vectors, handling transient nonlinear relationships such as abrupt acceleration changes. The periodic kernel is pre-set with three basic wavelengths (2 seconds, 5 seconds, and 10 seconds), corresponding to the time periods of common traffic scenarios. After analyzing the dominant frequency components of the input features using Fast Fourier Transform, it activates the optimal wavelength combination. The kernel function output performs weighted fusion, with weight coefficients obtained from historical data training and stored in read-only memory, generating a probabilistic feature representation containing the probability of vehicle position shift. This representation is stored in tensor form, with a three-dimensional structure corresponding to spatial location, time step, and probability density value, respectively.

[0068] The confidence interval calculation module implements uncertainty quantification on the probabilistic feature representation. This module integrates a mean prediction unit and a variance calculation unit: The mean prediction adopts a three-layer perceptron structure, and outputs the expected values of position and velocity at each time point; The variance calculation performs a three-step processing flow. The first step uses Monte Carlo sampling to generate 1000 groups of feature perturbations. The second step calculates the posterior distribution parameters through the Bayesian inference algorithm. The third step applies a variational inference optimizer to reduce the width of the confidence interval. The inference result is converted into a confidence level parameter, and a confidence interval enclosing the true value is generated at a 95% confidence level. The confidence marking of feature points is expressed by a three-dimensional color cloud map of the probability feature distribution. The red area represents high uncertainty (confidence level < 85%), and the blue represents low uncertainty.

[0069] The distance detection algorithm is equipped with a multi-mode anomaly recognition mechanism. The normal driving distribution model is stored in a distributed database, including the benchmark distributions of three basic scenarios: sunny, rainy, and foggy days. When the detection process is activated, the corresponding benchmark distribution is first loaded according to the real-time weather data. If the scenario does not match, the mixed distribution mode is enabled. The distance calculator performs a two-path parallel operation: The first path uses an improved Mahalanobis distance formula, introducing the inverse operation of the feature correlation matrix to eliminate the influence of dimensional coupling; The second path calculates the Euclidean geometric distance between the feature vector and the benchmark center point. The results of the two paths are input into the anomaly score synthesizer, and an anomaly flag is triggered when the score exceeds the preset threshold. The confidence threshold system divides five-level response criteria: When the anomaly score is in the range of 0.05 < p < 0.1, a yellow warning is marked, and when the score p < 0.01, a red alarm is activated. The confidence parameter is transmitted through an independent channel, always retaining the original calculation accuracy.

[0070] The risk feature vector generation system implements a quality verification mechanism. The vector structure is designed as a 128-dimensional floating-point array. The first 64 dimensions store the position risk parameters, the middle 48 dimensions record the velocity risk components, and the last 16 dimensions retain the environmental risk factors. The verifier performs three checks before output: parameter value range verification, confidence validity verification, and feature logical self-consistency detection. An abnormal output triggers a reprocessing mechanism, returning to the probability regression model stage for recalculation. The final risk feature vector is encapsulated into a fixed-structure data packet, and is transmitted through an optical fiber channel with a timestamp and a check code attached.

[0071] The hardware processing unit integrates fail-safe logic. The probability regression module sets a calculation timeout threshold of 300 milliseconds, and automatically switches to the simplified kernel function mode when the timeout occurs. The confidence calculation unit is configured with a triple redundant system, and an arbitration mechanism is started when the result deviation between the master and slave units exceeds 10%.

[0072] The dynamic parameter library supports online model evolution. Core algorithm parameters are stored in erasable memory and are automatically refreshed every 24 hours. The refresh mechanism dynamically adjusts the center point coordinates of the normal driving distribution based on the distribution characteristics of the latest 2000 frames of data. When the system detects a new driving mode, it automatically expands the baseline distribution library and adds a timestamp. The confidence threshold parameter is adaptively adjusted according to the ambient light intensity; during nighttime hours, the system automatically relaxes the confidence requirement by 5 percentage points while enhancing the review strength of high-confidence outputs.

[0073] The computing resource management unit implements multi-dimensional optimization. Probabilistic processing tasks allocate computing resources based on confidence level: high-confidence tasks utilize GPU acceleration cores, while low-confidence tasks use general-purpose CPU cores. The data caching strategy employs dynamic tiering, retaining copies of the original input data in memory for feature points with confidence levels below 80%. This module implements an energy efficiency control mechanism, automatically shutting down the third-level optimization stage of the variance calculation unit when the onboard power supply voltage drops below 12V, ensuring real-time performance of critical path processing.

[0074] The vector transmission system establishes a feedback control path. The receiving module monitors the processing efficiency of risk feature vectors in real time. When the delay exceeds 50 milliseconds for 30 consecutive frames, it automatically sends a dimensionality reduction control signal. This signal triggers the dimensionality compression mechanism of the front-end dimensionality reduction model, temporarily reducing the feature dimension to 48 dimensions. The receiving module synchronously records the data packet loss rate. When it exceeds 1%, it automatically activates the redundant transmission mode, copying each data frame twice and transmitting it through different channels. The output port is equipped with a data format converter to convert the internal binary format to the AUTOSAR standard data format for recognition and processing by the vehicle bus system.

[0075] Example 3: See Figure 4 The risk feature vector is input into the spatial mapping engine for coordinate transformation. The road topology database stores a high-precision digital map, including the three-dimensional coordinates of lane centerlines and topological connections. The mapping engine performs coordinate transformation operations, converting the relative position information of vehicles into absolute positions in a global Cartesian coordinate system. The transformation matrix is ​​dynamically updated based on the GPS positioning origin. Each risk feature point is assigned two-dimensional plane coordinates (x, y) and a risk intensity value z, forming a discrete set of points in three-dimensional space. These discrete points form an irregular distribution pattern within a 150-meter radius sensing range, requiring spatial interpolation to fill data gaps. The Kriging interpolation algorithm is activated, utilizing spatial autocorrelation characteristics to construct a prediction model. This algorithm relies on a semi-variogram to quantify spatial dependencies, and its parameters are automatically calibrated through maximum likelihood estimation.

[0076]

[0077] in: This represents the semivariogram (dimensionless) when the spatial distance is h, where h represents the Euclidean distance (in meters) between the point to be predicted and the known point. This represents the nugget effect constant (typical value 0.2). The structural variance is represented (range 0.5-1.5), and 'a' is the range parameter (default 25 meters). The model automatically optimizes parameters based on 80 nearby points collected in real time, generating a rasterized preliminary risk distribution field with a resolution of 0.5 meters × 0.5 meters. Constraints are set during spatial interpolation: when the risk value in the lane dividing line area exceeds the threshold, the interpolation boundary constraints are automatically strengthened to prevent the risk from spreading across lanes.

[0078] The time-dimensional processing system loads the initial risk distribution field at consecutive time points. The spatiotemporal interpolation module adopts a four-dimensional data cube structure (x, y, z, t), where t represents the time axis index. The interpolation algorithm performs Kalman filtering correction in the time dimension: it establishes a state transition equation to describe the risk propagation law and observes the attenuation effect of the equation in relation to historical data. The dynamic correction process includes three stages: spatial compensation coefficient calculation, time attenuation factor setting, and boundary condition reset. The final output is a spatiotemporally coupled risk field at 30 frames per second, in a three-channel tensor format.

[0079] The sequence prediction model employs a multi-layered processing architecture. The risk distribution field is transformed into a graph structure: the road surface is divided into 0.5m × 0.5m grid cells, each cell serving as a graph node; the connections between nodes are defined according to the road network topology, with directed edges automatically established between adjacent cells in the direction of travel. Each node stores the current risk value and a historical 20-second risk sequence. During the graph neural network initialization phase, four layers of graph convolutional processing are configured: the first layer aggregates features from nodes within a 1-meter radius; the second layer expands to a 3-meter radius to capture regional features; the third layer incorporates lane topology constraints to achieve cross-lane information transfer; and the fourth layer filters out low-association nodes through a gating mechanism. The network outputs a 512-dimensional node state vector as the initial prediction baseline.

[0080] The attention mechanism implements dynamic weight allocation in the graph structure. An eight-head attention structure is used to process spatial dependencies at different scales in parallel. Each attention head calculates three sets of parameters: the query vector Q is generated based on the velocity direction of the current node; the key vector K extracts the risk gradient features of neighboring nodes; and the value vector V encodes historical risk change patterns. A distance penalty factor is introduced into the attention weight calculation formula, with the weight of nodes beyond 50 meters decaying to 30%. After attention weighted fusion, a risk propagation prediction feature vector is formed, which is input into a dual-channel LSTM network: the forward channel predicts the propagation path in the next 5 seconds, and the backward channel optimizes the prediction accuracy at the current time step. The model outputs a probability distribution map of the predicted path in each frame, marking high-risk propagation areas (probability > 70%) and potential propagation directions (azimuth quantized to 8 partitions).

[0081] The propagation result post-processing system performs three optimizations. The accuracy compensation unit detects conflicts between the predicted path and the actual road network: when the predicted risk propagates to the median strip area, an 80% attenuation coefficient is automatically applied; when the predicted direction deviates from the road curvature by more than 15 degrees, a direction correction algorithm is activated. The trend extrapolation module, based on velocity field curl analysis, identifies risk vortex points and corrects the diffusion pattern. The result fusion system integrates the current prediction with the results of 10 previous predictions, generating the final propagation prediction through a weighted voting mechanism. The output data packet contains three sub-modules: a path trajectory point sequence (50Hz sampling rate), a risk intensity evolution curve (10 sampling points per second), and a spatial distribution heatmap (0.1-meter accuracy).

[0082] The system employs a real-time calibration process. After every five predictions, it automatically cross-validates with subsequent sensing results. A model retraining signal is triggered when the root mean square deviation of the location exceeds 0.8 meters. The historical risk propagation database uses a circular buffer structure, retaining propagation pattern data from the most recent 72 hours. Predictive model parameter updates utilize an incremental learning strategy, updating only 20% of the network's weight parameters in each training iteration. At the hardware level, computing resources are dynamically allocated: path prediction tasks are assigned to dedicated AI accelerators, while spatial interpolation tasks are processed in parallel by multi-core CPUs. The data pipeline features a dual-buffer structure, ensuring a continuous data throughput of 30 frames per second.

[0083] The error handling unit integrates multiple safeguards. A maximum iteration limit (300 times) is set during the spatial interpolation stage; upon timeout, it switches to fast linear interpolation mode. Isolated nodes (connectivity < 3) are detected during graph network construction, and graph structure reconstruction is automatically performed. When the prediction confidence is below 65%, a historical data comparison mode is activated: historical cases with similarity > 85% are retrieved from the database to replace the model output. All output data are labeled with quality tags: Grade A (no correction), Grade B (human intervention parameters < 15%), and Grade C (requires manual review).

[0084] The storage and transmission of prediction results employ a layered compression strategy. Key points in the path trajectory data are compressed using the Douglas-Pock algorithm, with a compression error threshold set at 0.2 meters. The risk heatmap uses JPEG2000 encoding, maintaining a peak signal-to-noise ratio above 45dB at a compression rate of 50%. The transmission protocol includes a data integrity verification field: a 16-bit CRC checksum is embedded in the header of each data packet, and the receiver's timeout retransmission mechanism has a 150-millisecond response window. Finally, the propagation prediction results are written to a specific address range in the shared storage area for subsequent use by the decision threshold generation module.

[0085] Example 4: The historical normal driving database stores rigorously selected driving scenario data, covering six typical environments including urban roads and highways. Data collection vehicles are equipped with seven sensors, including millimeter-wave radar and lidar, with a recording period of 18 months and a total effective mileage of 250,000 kilometers. The database adopts a hierarchical storage structure: the raw data layer stores the original sensor information, the feature extraction layer stores preprocessed motion trajectory features, and the scene annotation layer contains 5,600 manually labeled standard lane-change events. The data cleaning process performs seven steps, including outlier removal and trajectory smoothing, ultimately forming a training set containing 1.2 million frames of standard driving data.

[0086] The generative adversarial network (GAN) architecture is designed as a dual-channel structure. The generator's input layer receives a 128-dimensional random noise vector, which is expanded to 1024 dimensions through a fully connected network. The intermediate layers use gated recurrent units to process temporal features, and the output layer uses a Tanh activation function to generate simulated trajectory data. The discriminator employs a 3D convolutional neural network, taking a mixed batch of real driving data and generated data as input, and using a Sigmoid function to calculate the probability of authenticity in the output layer. Dynamic learning rate adjustment is implemented during training: the initial value is set to 0.0002, decreasing by 15% every 10 epochs. The loss function incorporates the Wasserstein distance metric, along with a gradient penalty term to control training stability. After 300 iterations, the simulated data output by the generator shows less than 8% difference from the real data in terms of trajectory curvature, acceleration distribution, and other metrics.

[0087] Specific training examples illustrate the evolution of the generator. In the initial stage (epoch 1-50), the generated trajectories exhibit significant jitter, with a maximum lateral acceleration deviation of 0.4g. In the mid-stage (epoch 150), the smoothness of the output trajectory improves, but the error in judging lane change timing remains at 1.2 seconds. In the mature stage (epoch 300), the parameters of the generated lane change trajectories enter a reasonable range. Table 1 shows the statistical comparison between the final generated data and the real data.

[0088] Table 1: Comparison statistics between the final generated data and the actual data.

[0089]

[0090] The real-time discriminator is constructed as a five-layer convolutional neural network. The input channels receive two types of data: predicted risk distribution field results and simulated data output from the generator. The first layer performs spatial feature extraction, using 5×5 convolutional kernels to capture local risk patterns. The second layer implements a channel attention mechanism to highlight key sensor features. The third layer performs spatiotemporal feature fusion, handling the dynamic changes of three consecutive frames of data. The fourth layer compresses the feature dimension to 256 dimensions. The fifth layer outputs the discrimination result and confidence score. The discriminator updates its parameters every 200 milliseconds to maintain its ability to recognize novel risk patterns.

[0091] The dynamic game process employs a three-step coordination mechanism. The first step involves data adversarial processing: the generator attempts to generate data that more closely resembles the characteristics of the risk field, while the discriminator simultaneously improves its recognition accuracy. The second step involves parameter tuning: when the discriminator's accuracy exceeds 85% for 10 consecutive times, the generator's learning rate increases by 20%. The third step involves strategy balancing: the game controller monitors the diversity index of the generated data, injecting random noise stimulation when it falls below a threshold of 0.6. The entire game cycle is completed within 15 seconds, ensuring real-time decision-making and response.

[0092] The multi-level threshold generation module employs a spectral clustering algorithm. The input data is a 500-dimensional feature vector output by the discriminator; a whitening operation is performed during preprocessing to eliminate feature correlation. The similarity matrix is ​​constructed using an improved cosine similarity metric, incorporating the risk propagation direction as a weighting factor. The number of clusters is dynamically adjusted to three levels (low / medium / high risk), and the cluster center coordinates are determined through iterative optimization. Threshold boundary settings consider two types of constraints: a hard constraint requiring at least a 15% safety margin between adjacent risk levels; and a soft constraint optimizing the cluster silhouette coefficient to above 0.5. The final generated three-level decision thresholds are dynamically updated using a sliding window mechanism, with the window size adaptively adjusted based on vehicle speed (range 50-120 frames).

[0093] A specific road scenario example demonstrates the threshold generation process. On an urban expressway, the system detects a sustained medium-risk state (risk value 0.4-0.6) in the left lane and intermittent high-risk peaks (risk value > 0.7) in the right lane. After analyzing historical data, the discriminator outputs three threshold levels: low risk 0.35, medium risk 0.55, and high risk 0.75. When the real-time risk feature vector value is 0.58, the system determines the current state is medium-risk, triggering a lane-change preparation command but not immediately executing an active lane change. At this moment, construction cones appear 200 meters ahead. The risk distribution field prediction shows the risk value will rise to 0.72 in 3 seconds. The system proactively lowers the high-risk threshold to 0.7, achieving a preventative decision adjustment.

[0094] The system implements a triple fault-tolerant protection mechanism. The first layer monitors the distribution deviation of generated data, automatically rolling back to the previous stable version when the KL divergence exceeds 0.3. The second layer verifies the consistency of the discriminator output, triggering a manual review request when the discriminator results fluctuate by more than 40% for five consecutive times. The third layer limits the threshold adjustment range, allowing a single update to not exceed ±20% of the previous cycle value. All protection mechanisms are implemented through independent hardware modules, with response latency controlled within 50 milliseconds.

[0095] The model update system employs an incremental learning strategy. Daily new normal driving data is filtered and then input into the generator for fine-tuning, with each update limiting parameter changes to within 5%. The discriminator performs a full update weekly, retaining 20% ​​of historically difficult samples during retraining. The cluster centers of the threshold generation module are automatically calibrated every 8 hours, taking into account day-night traffic flow differences. Detailed version information is recorded for all update operations, supporting rapid rollback to any historical version in case of failure.

[0096] Computational resource allocation is dynamically scheduled. Generative Adversarial Network (GAN) training tasks are only initiated when the system is idle, consuming no more than 30% of GPU resources; real-time discrimination tasks are set to the highest priority, ensuring completion within 100 milliseconds; threshold calculation tasks are decomposed into multiple subtasks for parallel processing. Memory management employs object pooling technology, with a pre-allocated 200MB dedicated cache for storing intermediate computation results.

[0097] The data interface is designed with a standardized communication protocol. The generator output data format conforms to the OpenDRIVE specification, including a sequence of trajectory points, timestamps, and confidence level markers. The discrimination result is encapsulated into a fixed structure containing an 8-byte risk level code and a 4-byte checksum. Threshold parameters are broadcast via the CAN bus, with a complete parameter set sent every 100 milliseconds. All communication data is appended with a sequence number identifier to prevent out-of-order packet processing.

[0098] Example 5: Real-time risk feature vector input decision comparator module, equipped with a three-level cache structure for data alignment. The feature vector contains 128 data points, divided into three logical segments according to risk type: location risk (first 64 dimensions), velocity risk (middle 48 dimensions), and environmental risk (last 16 dimensions). Multi-level decision thresholds are stored in a dual-port memory, including low-risk threshold vectors. Medium risk High risk Each threshold level corresponds to a 128-dimensional boundary value. The comparator implements a parallel processing architecture: the location risk unit calculates similarity using vector dot product, the velocity risk unit performs Euclidean distance measurement, and the environment risk unit applies Hamming distance analysis. The distance calculation results are input into a normalization transformer to generate deviation scores in the 0-1 interval. :

[0099]

[0100] in: Represents a real-time feature vector segment. For the corresponding threshold sub-segment, This represents the historical standard deviation of this dimension. , , These are dynamic weighting coefficients (initial values ​​0.4, 0.3, 0.3), and their symbols have the following meanings: Location risk feature vector (64 dimensions). Location risk threshold vector (64 dimensions). Location risk: historical standard deviation (scalar). Speed ​​risk feature vector (48 dimensions) : Speed ​​risk threshold vector (48 dimensions) Speed ​​risk historical standard deviation (scalar) Environmental risk feature vector (16 dimensions). : Environmental risk threshold vector (16 dimensions). Historical standard deviation of environmental risk (scalar) Location risk weighting coefficient (dynamic range 0.3-0.5). Speed ​​risk weighting coefficient (dynamic range 0.25-0.35). Environmental risk weighting coefficient (dynamic range 0.2-0.4). Overall deviation score (0-1 scalar).

[0101] The confidence fusion unit receives a 128-dimensional confidence vector from the risk assessment model. The global confidence level is calculated by weighted average. (Range 0-1). The decision adjustment logic implements a fifth-order state machine: when and Maintain the current decision-making level; When the value is in the 0.3-0.5 range, the level pre-adjustment mode is activated; Triggering emergency review process. Adjustment range. The calculation is based on a linear piecewise function: For every 0.1 unit increase, the decision level increases by 0.5 levels (maximum adjustment of 3 levels). The output decision label is mapped to the operation command: Level 0 generates a lane maintenance code (0xF0), Level 1 outputs a lane change preparation command (0xF1), Level 2 sends an active lane change request (0xF2), and Level 3 triggers emergency braking (0xF3).

[0102] Demonstrating the decision-making process in a specific scenario: Under highway conditions, the real-time location risk vector value suddenly increases to 1.8 times the normal value, the speed risk change reaches 2.3 times the historical standard deviation, and the environmental risk remains stable. Initial calculations. (Medium risk) The system activates Level 1 adjustment, raising the decision level from L1 to L2. At this point, the left lane radar detects a rapidly approaching vehicle, and the environmental risk submodule characteristic value jumps by 30% within 0.2 seconds. The value increases to 0.61, raising the decision level to L3. Upon receiving the 0xF2 command, the vehicle control system initiates a lane-changing procedure and simultaneously pre-charges the braking system.

[0103] The deviation monitoring system operates continuously after the decision command is output. Anomaly detectors deployed on the data bus record the differences between the real-time sensor stream and the prediction model. When a counter for the consecutive occurrences of the same type of deviation (such as position prediction error) reaches a set value... (Default 5 times), the phase analysis module is activated. Phase delay characteristics. The calculation process includes: selecting feature point time series, performing Hilbert transform to obtain analytic signals, calculating the instantaneous phase angle, and comparing with a reference phase curve to obtain the delay. This delay... Quantized to millisecond precision and stored as a 32-bit floating-point array.

[0104] Model correction engine based on Execution parameter reorganization. Weight adjustment factor. The calculation formula is ,in This represents the maximum permissible latency under the current operating conditions (default 300ms). The convolutional kernel weight matrix in the risk assessment model. Updated to ,in Historical baseline weights. Synchronous adjustment of bias terms in fully connected layers: bias vector. Updated to Each correction operation is recorded in the revision log, noting the modification time, the original parameter signature, and the adjustment range.

[0105] The parameter storage system implements version control. Corrected model parameters are stored in a specific partition of the historical database after SHA-256 hashing. The storage structure is a four-dimensional tensor: dimension one indexes model components (convolutional layers / fully connected layers, etc.), dimension two records parameter types (weights / biases), dimension three stores timestamps, and dimension four stores parameter data blocks. The database retains the 50 most recent correction records and uses an LRU algorithm to evict older data. When initializing a new model, the database is searched for the parameter snapshot closest to the current scene, with matching based on the cosine similarity of scene feature vectors.

[0106] The hardware interaction layer is configured with a protection mechanism. Before outputting decision commands, cross-validation is required: while the main decision module generates commands, a simplified model (retaining 30% of parameters) performs parallel inference. If the difference between the two commands exceeds one level, a safety lock is triggered, command transmission is suspended, and the diagnostic program is activated. During phase analysis, the system maintains the original data processing path to ensure real-time response capability. Model update operations are limited to when the vehicle is stationary or at low speeds (<20km / h), and double buffering technology is used during the update process to avoid service interruption.

[0107] The monitoring unit performs closed-loop verification. After each decision is executed, the vehicle's actual response data (steering angle change rate, deceleration value, etc.) is compared with the expected model. Deviation persistence rate. The calculation formula is

[0108]

[0109] in: For the count of abnormal events, The total number of decisions. When After 10 minutes, the model parameters are forcibly reset to the baseline state before the three most recent corrections. The monitoring log generates a summary report every minute, recording the decision success rate, the number of model corrections, and the storage area usage status.

[0110] The communication protocol incorporates an anti-collision mechanism. Decision commands are transmitted via an independent CAN channel, with message IDs ranging from 0x5A0 to 0x5A3 (corresponding to level 4 commands). Parameter update messages utilize the enhanced FlexRay protocol, with each frame containing a 128-byte parameter block and a 16-byte checksum. Historical database access employs a dedicated memory-mapped interface, with a physical address range of 0x500000-0x5FFFFF; access conflicts are resolved via a hardware arbitrator. All communication transactions are timestamped with an accuracy of 100 nanoseconds.

[0111] Example 6: The construction process of the multidimensional decision parameter set begins with the data acquisition system. The speed fluctuation entropy calculation module connects to the vehicle bus to acquire longitudinal speed signals, with the sampling frequency set to 100Hz. The raw speed data is first processed by a Butterworth low-pass filter, with the cutoff frequency set to 5Hz to eliminate high-frequency noise. The preprocessed data stream enters the analysis window, with the window length dynamically adjusted: a 10-second window is used for urban road conditions, and the window is extended to 20 seconds for highway conditions. The data within each window undergoes a Fast Fourier Transform, and the spectrum analysis range is limited to the 0-10Hz frequency band. Energy distribution feature extraction focuses on the low-frequency components of 0-1Hz, calculating the percentage of energy in this frequency band relative to the total energy. Entropy calculation uses a modified Shannon entropy formula, dividing the spectrum into 8 equal-width sub-bands, with the energy percentage of each sub-band serving as a probability input. When the low-frequency energy percentage exceeds 75%, the system marks the current traffic flow risk mode, and this status flag is stored in the shared memory area.

[0112] Environmental interference sensitivity parameters are derived from multi-sensor consistency analysis. Detection results from millimeter-wave radar, lidar, and cameras for the same target are input into a difference calculation unit, which calculates the standard deviation of each sensor's position and velocity measurements. Sensitivity scores are calculated using a three-layer neural network: the input layer receives the difference statistics from each sensor, the hidden layer contains 16 neurons, and the output layer generates a normalized score between 0 and 1. Road friction coefficients are obtained from an onboard road surface recognition system. The system estimates the current road surface adhesion coefficient in real time based on the relationship between tire slip ratio and braking pressure, updating the value every 200 milliseconds. These three types of parameters are encapsulated into a structured data packet and transmitted to the decision center via a message queue.

[0113] The parameter monitoring system implements a tiered response strategy. Two threshold levels are set for speed fluctuation entropy: the first threshold is 0.65, triggering a warning icon on the head-up display when exceeded; the second threshold is 0.8, activating a voice prompt suggesting a reduction in speed. The combined assessment of environmental disturbance sensitivity and road friction coefficient uses a fuzzy logic system: when sensitivity > 0.7 and friction coefficient < 0.3, the system classifies it as a high-risk combination; when sensitivity > 0.5 and friction coefficient < 0.5, it is assessed as medium-risk. Risk level transitions are implemented through a state machine, including a 5-second state hold time to prevent frequent jumps. After an emergency avoidance command is generated, the system takes over control of the steering wheel and braking system, executing the pre-programmed avoidance trajectory.

[0114] Traffic flow risk pattern recognition triggers a collaborative perception mechanism. The system acquires motion state data of surrounding vehicles via V2X communication, focusing on acceleration information. The data synchronization module aligns the time reference of the current vehicle with that of surrounding vehicles, with a maximum allowable deviation of 50 milliseconds. The correlation analysis unit calculates the Pearson correlation coefficient between the low-frequency energy of the current vehicle's speed and the acceleration of neighboring vehicles, keeping the analysis window consistent with the entropy calculation window. When the correlation coefficient exceeds a preset value of 0.7, the path planning module receives an adjustment instruction: the radius of curvature of the desired trajectory increases by 20%, and the lane-changing acceleration limit decreases by 30%. The adjusted path task is reparameterized using a cubic spline curve to ensure smooth motion.

[0115] Hardware resource allocation follows a dynamic priority principle. Parameter calculation tasks run on a dedicated digital signal processor, consuming no more than 40% of the computing resources. The emergency avoidance instruction generation channel is equipped with a hardware watchdog timer; if no heartbeat signal is received within 500 milliseconds, the processing unit is automatically reset. Data communication employs a dual-channel redundancy design: the primary channel uses the CANFD bus, and the backup channel transmits via Ethernet. Memory management reserves a dedicated buffer for critical parameters and implements a write protection mechanism to prevent accidental modification.

[0116] The fault handling system establishes a multi-layered defense system. A speed signal anomaly detector monitors data validity, switching to a backup sensor when 10 consecutive sampling points exceed a reasonable range. The environmental parameter calculation unit is equipped with a result reasonableness checker, discarding the current frame data when the difference between the three sensors exceeds the physically possible range. The road surface recognition system sets a confidence threshold; estimates below 0.6 do not update the friction coefficient. All fault events are logged in detail, including timestamps, fault codes, and recovery measures.

[0117] The system maintenance interface supports remote configuration updates. Threshold parameters are stored in a programmable read-only memory and can be modified online via the diagnostic interface. The algorithm module adopts a plug-in architecture, and key processing units support hot replacement without service interruption. The version control system records the complete trajectory of each parameter adjustment and supports rollback to any historical version. Maintenance operations require dual authentication: physical interface connection verification plus digital certificate verification.

[0118] The real-time debug information output channel transmits detailed operational status. The data stream includes current parameter values, risk level markers, and system load status. Debug information is encapsulated into fixed-format message packets and output to external analysis devices via a USB interface. The message filtering mechanism is dynamically configurable, selecting a subset of key parameters for transmission as needed.

[0119] Baseline data established during the vehicle integration testing phase serves as an initial reference. Parameter distribution ranges collected from 500 kilometers of actual road testing are used to calibrate initial threshold values. After the system goes live, operational data is continuously collected, and parameter distribution statistical analysis is performed monthly to automatically adjust threshold boundaries to match changes in the actual traffic environment. The calibration process considers seasonal differences, with winter and summer data used to establish separate reference models.

[0120] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0121] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art 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 appended claims and their equivalents.

Claims

1. A risk assessment-based lane-changing decision-making method for autonomous driving, comprising: Based on real-time data of the vehicle's surrounding environment, multi-dimensional features are fused to generate a spatiotemporal feature representation that includes location, speed, and environmental information, specifically including: Based on real-time data collected by multiple sensors, timestamp alignment and missing value filling are performed. A dynamic time warping algorithm is used to eliminate differences in sensor sampling frequencies to obtain a time-synchronized multi-source data sequence. The real-time data includes vehicle position, speed, and environmental data. The multi-source data sequence is input into the feature fusion network. Convolutional operations are used to capture road structure features in the spatial dimension, and sliding windows are used to extract time series features in the temporal dimension to obtain preliminary spatiotemporal features. For the initial spatiotemporal features, the correlation weights between different sensor data are calculated using an attention mechanism. The features are then weighted and fused according to the weights to highlight the contribution of key sensor data, resulting in weighted spatiotemporal features. The weighted spatiotemporal features are input into the dimensionality reduction model. Redundant information is removed by nonlinear transformation, key features are retained, and a low-dimensional spatiotemporal feature representation containing position, velocity and environmental information is generated. The spatiotemporal features are input into the risk assessment model to probabilistically model the abnormal risks during lane changing, and output a risk feature vector with confidence, specifically including: The spatiotemporal feature representation is input into a multilayer network and mapped to a high-dimensional latent space through nonlinear transformation to obtain a deep feature representation; For the aforementioned deep feature representation, a probabilistic regression model is constructed, and a combination of kernel functions is used to capture the nonlinear relationship and periodic changes in driving state, resulting in a probabilistic feature representation; Based on the probabilistic feature representation, the predicted mean and variance of each feature point are calculated, and the model uncertainty is quantified by the inference algorithm to obtain the probabilistic feature distribution with confidence intervals. For the aforementioned probability feature distribution, a distance detection algorithm is used to calculate the distance between each feature point and the normal driving distribution. Abnormal feature points are filtered according to a preset confidence threshold to generate a risk feature vector with confidence. Based on the risk feature vector, a risk distribution field is constructed, and the propagation path of the risk distribution field is predicted based on a sequence prediction model, specifically including: Based on the risk feature vector and combined with road topology information, each risk feature point is mapped to a two-dimensional coordinate system. An interpolation algorithm is then used to spatially interpolate the discrete risk feature points to generate a preliminary risk distribution field. Based on the preliminary risk distribution field, combined with time dimension information, a spatiotemporal interpolation algorithm is used to dynamically correct the risk distribution field, resulting in a risk distribution field that changes over time. The time-varying risk distribution field is input into the sequence prediction model to construct a road node graph model. An attention mechanism is used to capture the risk propagation dependency between nodes to obtain the initial risk propagation prediction. Based on the initial risk propagation prediction and combined with historical risk propagation data, the risk propagation path and diffusion trend at future time steps are predicted, generating a propagation prediction result for a dynamic risk distribution field. Based on the predicted results of the risk distribution field, an adaptive decision threshold generation model is constructed. Through a dynamic game between a normal driving generator and a real-time discriminator, a multi-level decision threshold that varies with driving status is generated, specifically including: Based on historical normal driving data, a normal driving generator based on generative adversarial networks is trained to generate simulated data that conforms to the characteristics of safe driving. Construct a real-time discriminator, input the predicted results of the risk distribution field and the simulated data generated by the normal driving generator, learn the boundary features of normal and abnormal states through dynamic game, and output the discrimination result and its confidence level; Based on the discrimination results, a clustering algorithm is used to divide the discrimination results into multiple levels. Combined with the dynamic changes in the risk distribution field, a multi-level decision threshold that changes with the driving state is generated, including a low-risk threshold, a medium-risk threshold, and a high-risk threshold. Based on the comparison results between real-time risk feature vectors and multi-level decision thresholds, autonomous driving lane-changing decisions are made, specifically including: The real-time risk feature vector is compared with a multi-level decision threshold, and the decision level is dynamically adjusted based on the confidence level of the risk features and the degree of deviation from the threshold using a decision model. Based on the decision level, a lane-changing instruction or a command to maintain the current lane is generated to realize autonomous driving lane-changing decision-making.

2. The autonomous driving lane-changing decision-making method based on risk assessment according to claim 1, characterized in that, The method further includes: When the deviation between real-time data and predicted data continues to exceed a set number of times, the phase delay feature of the feature points is extracted. Adjust the parameter weight ratios in the risk assessment model according to the phase delay characteristics; The corrected model parameters are stored in the historical database as the baseline values ​​for the next model initialization.

3. The autonomous driving lane-changing decision-making method based on risk assessment according to claim 2, characterized in that, The method further includes: Construct a multi-dimensional decision parameter set that includes speed fluctuation entropy, environmental disturbance sensitivity, and road friction coefficient; A status alert is triggered when a single parameter exceeds the first-level threshold, and an emergency evasion command is triggered when the combined effect of at least two parameters exceeds the second-level threshold.

4. The autonomous driving lane-changing decision-making method based on risk assessment according to claim 3, characterized in that, The calculation of the velocity fluctuation entropy value includes: Perform frequency domain transformation on velocity data within a specified time window to extract energy distribution characteristics of a preset frequency band; The speed fluctuation entropy value is calculated based on the energy distribution characteristics. When the proportion of low-frequency energy exceeds the preset proportion, it is determined to be a traffic flow risk mode.

5. The autonomous driving lane-changing decision-making method based on risk assessment according to claim 4, characterized in that, The method further includes: When a traffic flow risk pattern is identified, acceleration data of surrounding vehicles is collected simultaneously. The acceleration data is correlated with the low-frequency energy. If the correlation coefficient is greater than the preset correlation value, a path adjustment task is added to the lane change decision.

Images