Automatic driving lane changing decision-making method based on risk assessment

Abstract

The invention 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 steps of fusing multi-dimensional features according to real-time data of a vehicle surrounding environment, generating spatial-temporal feature representation containing position, speed and environment information, and realizing comprehensive perception of a complex traffic environment. And inputting the spatial-temporal feature representation into a risk assessment model, carrying out probabilistic modeling on an abnormal risk in a lane changing process, outputting a risk feature vector with confidence, and providing a refined risk reference for decision making. And then constructing a risk distribution field based on the risk feature vector, predicting a propagation path of the risk distribution field through a sequence prediction model, and prospectively perceiving the dynamic change of the risk. And finally, aiming at a prediction result, constructing a self-adaptive decision threshold generation model, and according to a comparison result of a real-time risk feature vector and a multi-level decision threshold, realizing an accurate and flexible automatic driving lane changing decision, and improving the safety and adaptability of a lane changing process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to an automatic driving lane changing decision method based on risk assessment. BACKGROUND

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

[0003] In existing automatic driving lane changing decision methods, some rely on a single sensor to obtain environmental data, which is difficult to fully capture the dynamic changes around the vehicle. For example, relying only on visual sensors may cause perception deviation in adverse weather conditions, and relying only on radar sensors may not accurately identify target categories, resulting in insufficient completeness and accuracy of environmental information. At the same time, in terms of feature processing, traditional methods often lack effective fusion of time and spatial dimensions, and cannot form coherent spatio-temporal feature representations, making it difficult to accurately judge the movement trends of surrounding vehicles.

[0004] In the risk assessment stage, existing technologies mostly use empirical rules or simple threshold judgments, lacking probabilistic modeling of abnormal risks in the lane changing process. This results in rough risk assessment results, making it difficult to quantify the occurrence probability and impact of different risks, and unable to provide fine-grained risk basis for decision-making. In addition, the risk in the lane changing process is not static, but dynamically propagates over time and space. Existing methods lack effective prediction of the dynamic propagation law of risk, and often can only deal with the risk state at the current time, making it difficult to avoid potential risk evolution in advance.

[0005] The setting of decision threshold is a key link in lane changing decision, and existing methods mostly use fixed thresholds or thresholds preset based on limited scenarios, which cannot be dynamically adjusted according to real-time driving state 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 problems make existing lane changing decision methods still have room for improvement in adaptability, accuracy and safety, making it difficult to meet the needs of automatic driving in diverse traffic environments. SUMMARY

[0006] The present application aims to provide an automatic driving lane changing decision method based on risk assessment to solve the problems raised in the background.

[0007] To achieve the above object, the application provides an automatic driving lane-changing decision-making method based on risk assessment, which comprises the following steps: According to real-time data of the vehicle surrounding environment, multi-dimensional features are fused to generate a space-time feature representation containing position, speed and environmental information; The space-time feature representation is input into a risk assessment model to probabilistically model abnormal risks in the lane-changing process and output a risk feature vector with confidence; According to 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; For the prediction result of the risk distribution field, an adaptive decision threshold generation model is constructed, a multi-level decision threshold that changes with the driving state is generated through the dynamic game of a normal driving generator and a real-time discriminator, and automatic driving lane-changing decision-making is realized according to the comparison result of the real-time risk feature vector and the multi-level decision threshold.

[0008] Preferably, the real-time data of the vehicle surrounding environment is fused to generate a space-time feature representation containing position, speed and environmental information, which comprises the following steps: According to the real-time data collected by the multi-source sensor, timestamp alignment and missing value filling processing are performed, and a dynamic time warping algorithm is used to eliminate the sampling frequency difference of the sensor to obtain a time-synchronized multi-source data sequence, wherein the real-time data includes vehicle position, speed and environmental data; The multi-source data sequence is input into a feature fusion network, wherein the spatial dimension uses convolution operation to capture road structure features, and the time dimension uses sliding window to extract time series features to obtain preliminary space-time features; The preliminary space-time features are weighted and fused according to the correlation weight between different sensor data calculated by the attention mechanism, and the contribution of key sensor data is highlighted to obtain weighted space-time features; The weighted space-time features are input into a dimension reduction model to remove redundant information through nonlinear transformation and retain key features to generate a low-dimensional space-time feature representation containing position, speed and environmental information.

[0009] Preferably, the space-time feature representation is input into a risk assessment model to probabilistically model abnormal risks in the lane-changing process and output a risk feature vector with confidence, which comprises the following steps: The space-time feature representation is input into a multi-layer network to be mapped to a high-dimensional latent space through nonlinear transformation to obtain a deep feature representation; For the deep feature representation, a probabilistic regression model is constructed to capture the nonlinear relationship and periodic change of the driving state by using kernel function combination to obtain a probabilistic feature representation; According to the probabilistic feature representation, a predicted mean and variance of each feature point are calculated, and a probabilistic feature distribution with a confidence interval is obtained by quantifying model uncertainty through an inference algorithm; For the probabilistic feature distribution, a distance detection algorithm is used to calculate the distance of each feature point from the normal driving distribution, and abnormal feature points are filtered according to a pre-set confidence threshold to generate a risk feature vector with confidence.

[0010] Preferably, according to the risk feature vector, a risk distribution field is constructed, and a propagation path of the risk distribution field is predicted based on a sequence prediction model, including: According to the risk feature vector, each risk feature point is mapped into a two-dimensional coordinate by combining road topological information, and an interpolation algorithm is used to perform spatial interpolation on discrete risk feature points to generate a preliminary risk distribution field; According to the preliminary risk distribution field, a spatiotemporal interpolation algorithm is used to dynamically correct the risk distribution field by combining time dimension information to obtain a risk distribution field that changes over time; The risk distribution field that changes over time is input into a sequence prediction model, a road node graph model is constructed, and an attention mechanism is used to capture the risk propagation dependency relationship between nodes to obtain an initial prediction of risk propagation; According to the initial prediction of risk propagation, the risk propagation path and diffusion trend at a future time step are predicted by combining historical risk propagation data to generate a propagation prediction result of the dynamic risk distribution field.

[0011] Preferably, for the prediction result of the risk distribution field, an adaptive decision threshold generation model is constructed, and a multi-level decision threshold that changes with driving state is generated through a dynamic game between a normal driving generator and a real-time discriminator, including: According to historical normal driving data, a normal driving generator based on a generative adversarial network is trained to generate simulated data conforming to safe driving features; A real-time discriminator is constructed, the prediction result of the risk distribution field and the simulated data generated by the normal driving generator are input, the normal and abnormal state boundary features are learned through a dynamic game, and a discrimination result and its confidence are output; According to the discrimination result, a clustering algorithm is used to divide the discrimination result into multiple levels, and a multi-level decision threshold that changes with driving state is generated by combining the dynamic changes of the risk distribution field, including low-risk, medium-risk and high-risk thresholds.

[0012] Preferably, according to the comparison result of the real-time risk feature vector and the multi-level decision threshold, automatic lane changing decision of autonomous driving is realized, including: The real-time risk feature vector is compared with the multi-level decision threshold, and a decision model is used to dynamically adjust the decision level according to the confidence of the risk feature and the deviation degree of the threshold; Based on the decision level, a lane changing instruction or a current lane maintaining instruction is generated to realize automatic driving lane changing decision.

[0013] Preferably, the method further comprises: When the deviation between real-time data and predicted data lasts for more than a set number of times, a phase delay feature of the feature point is extracted; According to the phase delay feature, the weight proportion of parameters in the risk assessment model is adjusted; The corrected model parameters are stored in a historical database as a reference value for the next model initialization.

[0014] Preferably, the method further comprises: A multi-dimensional decision parameter set containing speed fluctuation entropy, environmental interference sensitivity and road friction coefficient is constructed; When a single parameter exceeds a first threshold value, a state prompt is triggered, and when the combined effect of at least two parameters exceeds a second threshold value, an emergency avoidance instruction is triggered.

[0015] Preferably, the calculation of the speed fluctuation entropy comprises: The speed data in a specified time window is subjected to frequency domain transformation to extract the energy distribution feature of a preset frequency band; According to the energy distribution feature, the speed fluctuation entropy is calculated, and when the proportion of low-frequency energy exceeds a preset proportion, the traffic flow risk mode is determined.

[0016] Preferably, the method further comprises: When the traffic flow risk mode is identified, the acceleration data of surrounding vehicles is synchronously collected; The acceleration data and the low-frequency energy are subjected to correlation analysis, and if the correlation coefficient is greater than a preset correlation value, a path adjustment task is added in the lane changing decision.

[0017] Compared with the prior art, the present application has the following beneficial effects: The automatic driving lane changing decision method based on risk assessment significantly improves the adaptability and accuracy of lane changing decision through multi-dimensional technical optimization. In the environment perception and feature processing stage, the method performs multi-dimensional feature fusion according to the real-time data of the vehicle's surrounding environment, which can integrate the position, speed and environmental information obtained by different sensors to form a comprehensive spatio-temporal feature representation. This fusion method breaks through the limitations of single sensor perception, not only preserving the relative position relationship between the vehicle and the surrounding targets in the spatial dimension, but also capturing the motion trend changes in the time dimension, making the system's perception of complex traffic environment more comprehensive and coherent, providing a solid foundation for subsequent risk assessment.

[0018] In the risk assessment link, the spatio-temporal feature representation is input into the risk assessment model for probabilistic modeling of abnormal risk, and a risk feature vector with confidence is output. This probabilistic modeling method can quantitatively describe various abnormal risks that may occur during the lane changing process, and the confidence index reflects the reliability of the risk assessment result, avoiding the roughness of traditional experience rules or simple threshold judgment. The output of the risk feature vector enables different types and degrees of risk to be clearly distinguished, providing a refined risk reference for the decision system and helping to more accurately identify potential hazards.

[0019] Based on the risk feature vector, a risk distribution field is constructed, and its propagation path is predicted through a sequence prediction model, realizing the forward-looking perception of dynamic changes in risk. The risk in the lane changing process is not constant, but evolves with the movement of the vehicle and the behavior of the surrounding targets. The risk distribution field can intuitively present the distribution state of risk in space, and the propagation path prediction can reveal the development trend of risk in the time dimension in advance. This enables the system to go beyond passive response to the current risk state, and to predict the risk changes in the future for a certain period of time, leaving enough reaction time for decision-making, and enhancing the initiative and predictability of lane changing decision-making.

[0020] An adaptive decision threshold generation model is constructed for the risk distribution field prediction result, and a multi-level decision threshold is generated through the dynamic game between the normal driving generator and the real-time discriminator, solving the problem of insufficient adaptability of traditional fixed thresholds. The dynamic game process can dynamically adjust the decision threshold by combining the current driving state, environmental features and risk prediction results in real time: in a low-risk scenario, the threshold can be appropriately relaxed to improve lane changing efficiency; in a high-risk scenario, the threshold is tightened accordingly to ensure safety. This multi-level decision threshold that changes with the driving state enables the lane changing decision to adapt flexibly to different traffic environments, achieving a better balance between safety and efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A working principle diagram of the automatic driving lane changing decision-making method based on risk assessment described in the present application; Figure 2 A flowchart for generating a spatio-temporal feature representation by multi-dimensional feature fusion; Figure 3 A flowchart for outputting a risk feature vector with confidence by a risk assessment model; Figure 4 A flowchart for constructing a risk distribution field and predicting a propagation path. DETAILED DESCRIPTION

[0022] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described, obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0023] Please refer to Figure 1 The present application provides a risk assessment-based automatic driving lane-changing decision method, which comprises.

[0024] By integrating real-time data of the vehicle's surrounding environment, a spatiotemporal feature representation is generated through multi-dimensional feature fusion, which covers position, speed and environmental information. Subsequently, the spatiotemporal feature representation is input into a risk assessment model, which assesses the abnormal risks in the lane-changing process through probabilistic modeling and outputs a risk feature vector with confidence. Based on the risk feature vector, a risk distribution field is constructed, and a sequence prediction model is used to predict the propagation path of the risk distribution field. For the prediction results of the risk distribution field, a dynamic game between the normal driving generator and the real-time discriminator is performed to generate multi-level decision thresholds that change with driving states. Finally, by comparing the real-time risk feature vector with the multi-level decision thresholds, adaptive control of automatic driving lane-changing decision is realized. This process ensures timely response to risk changes in dynamic driving environments and optimizes the safety of lane-changing behavior.

[0025] Embodiment 1: refer to Figure 2 A multi-source sensor system collects real-time vehicle position, speed and environmental data. Position data is derived from a GPS and inertial measurement unit combined positioning system; speed data integrates wheel speed sensors and longitudinal / lateral acceleration information transmitted by the vehicle bus; environmental data includes millimeter wave radar target list, laser radar point cloud, camera semantic segmentation results and V2X communication data. These raw data are first processed by timestamp alignment, which matches the time reference values of each sensor through a hardware clock synchronization mechanism, and generates filling values for missing data caused by transmission delay using a cubic spline interpolation method at the corresponding timestamp. The dynamic time warping algorithm processes the sampling frequency difference: the data streams of laser radar data with a sampling rate of 20Hz and camera with a sampling rate of 30Hz are time axis aligned through the construction of a time series distance matrix to search for the minimum cost path to realize the time axis alignment of non-uniform sampling points, generating time-synchronized multi-source data sequences.

[0026] The multi-source data sequence input feature fusion network is processed. The spatial dimension processing unit loads the road grid map as the initial value of the convolution kernel weight, and extracts the spatial structure features through three layers of convolution operation: the first layer uses a 5x5 convolution kernel to perform grid processing on the vehicle within a 50-meter range, and each grid stores the existence probability of environmental objects; the second layer uses a 3x3 hollow convolution to expand the receptive field and identify the lane line topological correlation features; the third layer applies a 1x1 convolution to compress the channel dimension and generate a spatial feature map with a resolution of 256x256. The time dimension processing adopts a double sliding window mechanism: a 50ms short window calculates the speed change rate and acceleration statistics, and a 3s long window performs autoregressive integral sliding average modeling to extract the trend fluctuation features. In the spatio-temporal cross processing stage, the spatial feature map and the time series statistics are input into the feature splicing layer according to the time step to generate a preliminary spatio-temporal feature tensor containing the kinematic state of the vehicle and the position of the static obstacle.

[0027] The attention mechanism constructs a sensor weight distribution system on the feature tensor. A trainable feature correlation matrix is created, which has a dimension corresponding to the number of sensors, and the initial weight is generated by calculating the cosine similarity of the sensor feature vector. The weight is input into the gated recurrent unit, which is dynamically modulated according to the current vehicle speed and road curvature: when the GPS positioning accuracy factor is lower than the threshold, the weight of the millimeter wave radar is increased; in rainy weather, the weight of the camera is attenuated and the weight of the laser radar is multiplied. The modulated weight matrix is normalized by Softmax, and the multi-source sensor input is weighted and fused. In specific operations, the environmental data feature vector is multiplied by the corresponding weight and then added element by element, and the kinematic feature uses a weighted average strategy.

[0028] The dimension reduction model realizes feature compression based on the variational autoencoder architecture. The encoder part is composed of three fully connected layers: the first layer projects the 2048-dimensional weighted feature to 1024-dimensional, using the LeakyReLU activation function; the second layer compresses to 512-dimensional, applying batch normalization to prevent gradient deviation; the third layer generates a 256-dimensional mean vector and a logarithmic variance vector. The hidden variable space performs a reparameterization operation and samples a 128-dimensional feature vector. The decoder reconstructs the original feature through a symmetric structure to minimize the reconstruction loss function and constrain the information integrity. In the key feature extraction stage, the hidden variable vector is input into the feature selection module, the mutual information value of each dimension and the vehicle position change is calculated, and the first 64-dimensional features with mutual information value higher than 0.15 are retained, and the remaining dimensions are set to zero. The final output low-dimensional spatio-temporal feature representation includes 32-dimensional position features (XY coordinates, heading angle), 24-dimensional velocity features (vectorized components), and 8-dimensional environmental features (obstacle density, visible distance), with a total dimension of 64, which is lower than the original data.

[0029] The data processing chain configures a real-time feedback mechanism. The output feature representation inputs an anomaly detector, which calculates a reconstruction error value. When the error of five consecutive frames exceeds the historical mean by three standard deviations, a sensor recalibration signal is triggered. This signal controls the dynamic time warping algorithm to reinitialize the time warping path and reset the convolution kernel weights in the feature fusion network. The decoder output of the dimension reduction model synchronously inputs a quality assessment module. When the position feature reconstruction bias exceeds 0.3 meters, a redundant dimension retention strategy is activated, temporarily expanding the feature dimension to 96 dimensions until the bias returns to the normal range. The entire processing flow is executed in a 100-millisecond cycle, with a processing delay strictly controlled within 20 milliseconds.

[0030] The persistence storage of the spatiotemporal feature representation uses an incremental update mechanism. Every 100 frames of feature data generated triggers a clustering analysis, which identifies the feature distribution pattern cluster center through the K-means algorithm, replacing the reference feature vector in the historical database. In online operation, the real-time feature vector is compared with the database pattern for similarity. When an unknown pattern appears, the storage area is automatically expanded and the new driving scenario is marked. This database connects the parameter adjustment channel of the risk assessment model, and the statistical characteristics of the feature distribution directly affect the variance calculation parameters in the model.

[0031] The multi-source sensor fusion system is configured with a fault protection logic. When the millimeter wave radar does not detect obstacles for 10 consecutive frames, the camera feature weight is automatically increased by 20%; if the laser radar point cloud density drops sharply, the V2X data replacement strategy is enabled. The feature fusion network has a built-in self-checking mechanism that generates virtual data periodically to verify the running state of each channel, and the abnormal channel is automatically switched to the backup computing node. The output data is finally packaged into a standardized data structure, including a timestamp checksum, feature dimension units, and a completeness marker field, and is transmitted to the risk assessment module through shared memory.

[0032] Embodiment 2: refer to Figure 3 , the spatiotemporal feature representation inputs a multi-layer network processing system. The network uses a residual connection structure, including five fully connected layers and three gate units. The first layer receives a 64-dimensional feature vector, expands the feature dimension through 512 neurons, and applies an ELU activation function to handle nonlinear relationships. The output features are normalized and input into the second layer, which has 256 neurons and uses a hyperbolic tangent activation function to modulate the feature amplitude. The third layer introduces a gated linear unit that dynamically adjusts the feature gating weight based on the vehicle's current acceleration value: when the acceleration absolute value exceeds 0.3g, the time dimension weight is increased by 25%. The fourth layer performs feature cross operation to generate the interaction product of position and speed features, enhancing the motion state coupling relationship. The fifth layer outputs a 256-dimensional deep feature representation, with an information entropy 60% higher than the input feature.

[0033] The probability regression model constructs a multi-dimensional probability distribution on the deep feature representation. The kernel function combination system integrates Gaussian radial basis kernel and periodic kernel two types of core processors. The Gaussian kernel configuration adjustable bandwidth parameter, based on the feature vector standard deviation automatic setting initial value, processing acceleration mutation and other transient nonlinear relationship. The periodic kernel is preset with three basic wavelengths (2 seconds, 5 seconds, 10 seconds), corresponding to the time period of common traffic scenes, and the optimal wavelength combination is activated after analyzing the main frequency components of the input features by fast Fourier transform. The kernel function output executes weighted fusion, and the weight coefficient is obtained by training historical data and fixed in read-only memory to generate a probabilistic feature representation containing vehicle position displacement probability. The representation is stored in the form of a tensor, and the three-dimensional structure corresponds to the spatial position, time step and probability density value respectively.

[0034] The confidence interval calculation module realizes uncertainty quantification on the probabilistic feature representation. The module integrates mean prediction unit and variance calculation unit: the mean prediction adopts a three-layer perceptron structure, which outputs the position expectation value and speed expectation value at each time point; the variance calculation executes a three-step processing procedure, the first step uses Monte Carlo sampling to generate 1000 feature perturbations, the second step calculates the posterior distribution parameters by Bayesian inference algorithm, and the third step applies variational inference optimizer to reduce the confidence interval width. The inference result is converted into a confidence level parameter to generate a confidence interval enclosing the true value at 95% confidence level. The feature point confidence degree label is expressed in the form of a three-dimensional color cloud map of the probability feature distribution, with red area representing high uncertainty (confidence level < 85%) and blue area representing low uncertainty.

[0035] 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 distribution of three basic scenes: sunny, rainy and foggy. When the detection process is activated, the corresponding benchmark distribution is loaded first according to the real-time weather data, and the mixed distribution mode is enabled for unmatched scenes. The distance calculator performs double-path parallel operation: the first path uses an improved Mahalanobis distance formula, which introduces feature correlation matrix inverse operation to eliminate the influence of dimension coupling; the second path calculates the Euclidean distance between the feature vector and the benchmark center point. The double-path results are input into the anomaly score synthesizer, and if the score exceeds the preset threshold, an abnormality flag is triggered. The confidence threshold system divides the response standard into five levels: when the anomaly score is in the range of 0.05 < p < 0.1, a yellow warning is set, and when the score p < 0.01, a red alarm is activated. The confidence parameter is transmitted through an independent channel, and the original calculation accuracy is always preserved.

[0036] The risk feature vector generation system implements a quality check mechanism. The vector structure is designed as a 128-dimensional floating-point array, with the first 64 dimensions storing position risk parameters, the middle 48 dimensions recording speed risk components, and the last 16 dimensions reserved for environmental risk factors. The checker performs three checks before output: parameter value range verification, confidence validity verification, and feature logic self-consistency detection. Abnormal output triggers a reprocessing mechanism, returning to the probability regression model stage for recalculation. The final risk feature vector is packaged as a fixed-structure data packet, with a timestamp and a check code attached, and transmitted through a fiber channel.

[0037] The hardware processing unit integrates a failsafe logic. The probability regression module sets a calculation timeout threshold of 300 milliseconds, and automatically switches to a simplified kernel function mode when the timeout occurs. The confidence calculation unit is configured with a triple-redundancy system, which starts an arbitration mechanism when the deviation between the master and slave units exceeds 10%.

[0038] The dynamic parameter library supports online evolution of the model. Core algorithm parameters are stored in a writable memory, which is automatically refreshed every 24 hours. The refresh mechanism is based on the distribution characteristics of the latest 2000 frames of data, dynamically adjusting the center point coordinates of the normal driving distribution. When the system detects a new driving mode, it automatically expands the reference distribution library and adds a timestamp marker. The confidence threshold parameter is adaptively adjusted according to the intensity of the ambient light, with the system automatically relaxing the confidence requirement by 5 percentage points at night, while enhancing the audit intensity of high-confidence outputs.

[0039] The computing resource management unit implements multi-dimensional optimization. Probability processing tasks are allocated computing resources based on confidence levels: high-confidence demand tasks call GPU acceleration cores, and low-confidence demand tasks use CPU general-purpose cores. The data caching strategy implements dynamic grading, with feature points having a confidence level below 80% retaining a copy of the original input data in memory. This module implements an energy efficiency control mechanism that automatically shuts down the third level of optimization of the variance calculation unit when the vehicle's power supply voltage is below 12V, ensuring real-time performance of critical path processing.

[0040] The vector transmission system establishes a feedback control channel. The receiving module monitors the risk feature vector processing efficiency in real time, and when the delay exceeds 50 milliseconds for 30 consecutive frames, it automatically sends a dimension reduction control signal. This signal triggers the dimension compression mechanism of the front-end dimension reduction model, temporarily reducing the feature dimension to 48. The receiving module simultaneously records the data packet loss rate, and when it exceeds 1%, it automatically starts a redundant transmission mode, duplicating each data frame twice and transmitting it through different channels. The output port is equipped with a data format converter that converts the internal binary format to the AUTOSAR standard data format for recognition and processing by the vehicle's bus system.

[0041] Example 3: see Figure 4, the risk feature vector input space mapping engine for coordinate conversion processing. The road topology database stores high-precision digital maps, including lane centerline three-dimensional coordinates and topological connection relationships. The mapping engine performs coordinate transformation operations to convert vehicle relative position information into absolute positions in the global Cartesian coordinate system, and the conversion matrix is dynamically updated based on the GPS positioning origin. Each risk feature point is assigned a two-dimensional plane coordinate (x, y) and a risk intensity value z, forming a discrete point set in three-dimensional space. These discrete points form an irregular distribution pattern within a perception range of 150 meters, and spatial interpolation is required to fill in data gaps. The Kriging interpolation algorithm is activated to construct a prediction model using spatial autocorrelation characteristics. This algorithm relies on the semi-variogram to quantify spatial dependence, and its parameters are automatically calibrated through maximum likelihood estimation:

[0042] wherein: represents the semi-variogram value (dimensionless) when the spatial position distance is h, h represents the Euclidean distance (unit: meters) between the predicted point and the known point, represents the nugget effect constant (typical value 0.2), represents the structural variance (range 0.5-1.5), a is the range parameter (default value 25 meters). The model automatically optimizes the parameters based on the real-time collection of 80 neighboring points, generating a preliminary risk distribution field with a resolution of 0.5 meters x 0.5 meters. During the spatial interpolation process, constraints are set: when the risk value in the lane division line area exceeds the threshold, the interpolation boundary constraints are automatically strengthened to prevent risk from spreading across lanes.

[0043] The time dimension processing system loads the preliminary risk distribution field at consecutive time points. The spatio-temporal interpolation module uses a four-dimensional data cube structure (x, y, z, t), where t represents the time axis index. The interpolation algorithm performs Kalman filter correction in the time dimension: by establishing a state transition equation to describe the risk propagation law, and an observation equation to associate the decay effect of historical data. The dynamic correction process includes three links: spatial compensation coefficient calculation, time decay factor setting, and boundary condition resetting. Finally, 30 frames of spatio-temporal coupled risk fields per second are output, with a data format of three-channel tensor.

[0044] The sequence prediction model constructs a multi-layer processing architecture. The risk distribution field is converted into a graph structure: the road plane is divided into 0.5m x 0.5m grid cells, each cell serving as a graph node; the connection relationship between nodes is defined according to the road network topology, and the adjacent cells in the direction of travel automatically establish a directed edge. Each node stores the current risk value and the historical 20-second risk sequence. In the initialization stage of the graph neural network, four layers of graph convolution processing are configured: the first layer aggregates the features of the nodes within a 1-meter radius; the second layer extends to a 3-meter radius to capture regional features; the third layer adds lane topology constraints to realize cross-lane information transmission; the fourth layer filters out low-correlation nodes through a gating mechanism. The network outputs a 512-dimensional node state vector as the initial prediction benchmark.

[0045] The attention mechanism implements dynamic weight allocation in the graph structure. An eight-head attention structure is used to process different scales of spatial dependence in parallel. Each attention head calculates three groups of parameters: the query vector Q is generated according to the speed direction of the current node; the key vector K extracts the risk gradient features of adjacent nodes; and the value vector V encodes the historical risk change pattern. The attention weight calculation formula introduces a distance penalty factor, and the weight of nodes 50 meters away decays to 30%. After attention weighted fusion, a risk propagation prediction feature vector is formed, which is input into a double-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 every frame, labeling high-risk diffusion areas (probability > 70%) and potential propagation directions (azimuth quantized to 8 partitions).

[0046] The propagation result post-processing system performs three optimizations. The accuracy compensation unit detects the conflict between the predicted path and the actual road network: when the predicted risk propagates to the isolation area, an 80% decay coefficient is automatically applied; when the deviation between the predicted direction and the road curvature exceeds 15 degrees, the direction correction algorithm is started. The trend extrapolation module identifies risk vortex points based on the analysis of the velocity field vorticity and corrects the diffusion pattern. The result fusion system integrates the current time prediction and the historical 10 times prediction results, and generates the final propagation prediction through a weighted voting mechanism. The output data packet includes three sub-modules: path trajectory point sequence (50Hz sampling rate), risk intensity evolution curve (10 sampling points per second), and spatial distribution heat map (0.1m precision).

[0047] The system features a real-time calibration process. After every five predictions, it automatically cross-validates with subsequent perception results. A model retraining signal is triggered when the root mean square (RMS) position deviation exceeds 0.8 meters. The historical risk propagation database utilizes a ring buffer structure, retaining the last 72 hours of propagation pattern data. Prediction model parameter updates utilize an incremental learning strategy, with only 20% of the network's weight parameters updated during each training session. Computing resources are dynamically allocated at the hardware level: path prediction tasks are assigned to a dedicated AI accelerator, while spatial interpolation tasks are processed in parallel by a multi-core CPU. A dual-buffered data pipeline ensures continuous processing of 30 frames per second.

[0048] The error handling unit integrates multiple safeguards. The spatial interpolation phase has a maximum iteration limit (300), switching to fast linear interpolation mode upon timeout. Isolated nodes (connectivity < 3) are detected during graph network construction, and the graph structure is automatically reconstructed. When the prediction confidence falls below 65%, a historical data comparison mode is activated: historical cases with a similarity > 85% are retrieved from the database to replace the model output. All output data is assigned quality flags: A (no correction), B (manual intervention parameters < 15%), and C (manual review required).

[0049] A layered compression strategy is employed for the storage and transmission of prediction results. The Douglas-Peucker algorithm is used to compress key points of trajectory data, with a compression error threshold set at 0.2 meters. The risk heat map uses JPEG2000 encoding technology, maintaining a peak signal-to-noise ratio above 45dB at a 50% compression rate. The transmission protocol incorporates a data integrity check field: a 16-bit CRC checksum is embedded in each packet header, and the receiving end's timeout retransmission mechanism implements a 150-millisecond response window. The final propagation prediction results are written to a specific address segment in the shared memory area for subsequent use by the decision threshold generation module.

[0050] Example 4: The historical normal driving database stores rigorously screened 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 utilizes a hierarchical storage structure: the raw data layer stores raw 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, to produce a training set containing 1.2 million frames of standard driving data.

[0051] The generative adversarial network architecture is designed as a dual-channel structure. The generator input layer receives a 128-dimensional random noise vector, which is expanded to 1024 dimensions through a fully connected network. The middle layer uses a gated recurrent unit to process the time sequence features, and the output layer is configured with a Tanh activation function to generate simulated trajectory data. The discriminator uses a three-dimensional convolutional neural network, inputs a mixed batch of real driving data and generated data, and the output layer uses a Sigmoid function to calculate the authenticity probability. The training process implements dynamic learning rate adjustment: the initial value is set to 0.0002, and it is attenuated by 15% every 10 epochs. The loss function introduces the Wasserstein distance metric, combined with a gradient penalty term to control the stability of training. After 300 iterations, the simulated data generated by the generator has a difference of less than 8% in trajectory curvature, acceleration distribution and other indicators compared with real data.

[0052] The specific training example shows the evolution process of the generator. In the initial stage (epoch 1-50), the generated trajectory has obvious jitter, with a maximum lateral acceleration deviation of 0.4g; in the middle stage (epoch 150), the output trajectory smoothness is improved, but the lane change timing error still maintains at 1.2 seconds; in the mature stage (epoch 300), the generated lane change trajectory parameters enter the reasonable range, and Table 1 shows the comparison statistics of the final generated data and real data.

[0053] Table 1: Comparison statistics of the final generated data and real data.

[0054]

[0055] The real-time discriminator is built as a five-layer convolutional neural network, which receives two types of data: risk distribution field prediction results and simulated data generated by the generator. The first layer of the network performs spatial feature extraction, using a 5x5 convolution kernel to capture local risk patterns; the second layer implements a channel attention mechanism to highlight key sensor features; the third layer performs spatio-temporal feature fusion to handle the dynamic changes of consecutive three frames of data; the fourth layer compresses the feature dimension to 256 dimensions; and the fifth layer outputs the discrimination result and confidence score. The discriminator updates its parameters every 200 milliseconds to maintain its ability to recognize new risk patterns.

[0056] The dynamic game process implements a three-step coordination mechanism. The first step performs data confrontation: the generator tries to generate data closer to the risk field features, and the discriminator simultaneously improves its recognition accuracy; the second step performs parameter tuning: when the discriminator accuracy exceeds 85% for 10 consecutive times, the generator learning rate is increased by 20%; the third step implements strategy balancing: the game controller monitors the diversity indicators of the generated data, and injects random noise stimulation when the indicators are below the threshold of 0.6. The entire game cycle is completed within 15 seconds, ensuring the real-time response of decision-making.

[0057] The multi-level threshold generation module adopts a spectral clustering algorithm. The input data is a 500-dimensional feature vector output by the discriminator. In the preprocessing stage, a whitening operation is performed to eliminate feature correlation. The similarity matrix is constructed using an improved cosine similarity measure, and the risk propagation direction is introduced as a weight factor. The number of clusters is dynamically adjusted to three levels (low, medium, and high risk), and the cluster center coordinates are determined through iterative optimization. The threshold boundary setting considers two types of constraints: a hard constraint that requires at least a 15% safety margin between adjacent risk levels; and a soft constraint that optimizes the cluster profile coefficient to above 0.5. The final three-level decision threshold is dynamically updated through a sliding window mechanism, and the window size is adaptively adjusted according to the vehicle speed (range of 50-120 frames).

[0058] Specific road scene examples demonstrate the threshold generation process. In an urban expressway scenario, the system detects a continuous 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 the historical data, the discriminator outputs three-level thresholds: low risk threshold 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 that it is currently in a medium risk state and triggers a lane change preparation instruction but does not execute an active lane change. At this time, a construction cone appears 200 meters ahead, and the risk distribution field prediction shows that the risk value will rise to 0.72 after 3 seconds. The system proactively adjusts the high risk threshold to 0.7, realizing preventive decision adjustment.

[0059] The system implements a triple fault-tolerant protection mechanism. The first protection monitors the distribution deviation of generated data, and automatically rolls back to the last stable version when the KL divergence exceeds 0.3. The second protection checks the consistency of the discriminator output, and triggers an artificial review request when the fluctuation of consecutive 5 discrimination results exceeds 40%. The third protection limits the threshold adjustment range, and does not allow a single update to exceed ±20% of the previous period value. All protection mechanisms are implemented through independent hardware modules, with a response delay controlled within 50 milliseconds.

[0060] The model update system adopts an incremental learning strategy. Newly added normal driving data is filtered and input into the generator fine-tuning process, with a parameter variation limited to within 5% each time. The discriminator performs full update every week, and retains 20% of historical difficult samples during retraining. The cluster center points of the threshold generation module are automatically calibrated every 8 hours, considering factors such as day-night traffic flow differences. All update operations record detailed version information, supporting quick rollback to any historical version in case of failure.

[0061] The computing resource allocation implements dynamic scheduling. The GAN training task is only started when the system is idle, occupying no more than 30% of the GPU resources; the real-time discrimination task is set to the highest priority to ensure that the processing is completed within 100 milliseconds; and the threshold calculation task is decomposed into multiple sub-tasks for parallel processing. The object pool technology is used for memory management, and a special cache area of 200 MB is pre-allocated to store the intermediate calculation results.

[0062] The data interface designs a standardized communication protocol. The generator output data format complies with the OpenDRIVE specification, including a sequence of trajectory points, timestamps, and confidence markers. The discrimination result is packaged into a fixed structure, containing an 8-byte risk level code and a 4-byte checksum. The threshold parameters are broadcasted through the CAN bus, with a complete parameter set sent every 100 milliseconds. All communication data is appended with a sequence number to prevent out-of-order processing of data packets.

[0063] In embodiment 5, a real-time risk feature vector input decision comparator module is implemented, which is equipped with a three-level cache structure for data alignment processing. The feature vector contains 128-dimensional data points, which are divided into three logical segments according to risk type: position risk (first 64 dimensions), speed 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 vector , medium-risk , and high-risk , with each level of threshold corresponding to 128-dimensional boundary values. The comparator implements a parallel processing architecture: the position risk unit uses vector dot product to calculate similarity, the speed risk unit performs Euclidean distance measurement, and the environmental risk unit applies Hamming distance analysis. The distance calculation results are input into a standardization converter to generate a deviation score in the 0-1 interval :

[0064] wherein: represents the real-time feature vector sub-segment, is the corresponding threshold sub-segment, represents the historical standard deviation of that dimension. 、 、 is the dynamic weight coefficient (initial value 0.4, 0.3, 0.3), and its symbol meaning is as follows: : position risk feature vector (64 dimensions), : position risk threshold vector (64 dimensions), : position 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), : Environmental risk historical standard deviation (scalar), : Position risk weight coefficient (dynamic range 0.3-0.5), : Speed risk weight coefficient (dynamic range 0.25-0.35), : Environmental risk weight coefficient (dynamic range 0.2-0.4), : Composite deviation score (0-1 scalar).

[0065] The confidence fusion unit receives a 128-dimensional confidence vector from the risk assessment model , calculates the global confidence level by weighted average (range 0-1). The decision adjustment logic implements a five-stage state machine: when and , maintain the current decision level; activate the level pre-adjustment mode when in the 0.3-0.5 interval; trigger the emergency review process. The adjustment range is calculated based on a linear piecewise function: for every 0.1 unit increase, the decision level is raised by 0.5 levels (maximum 3-level adjustment). The output decision label is mapped to an operation instruction: level 0 generates a lane maintenance code (0xF0), level 1 outputs a lane change preparation instruction (0xF1), level 2 sends an active lane change request (0xF2), and level 3 triggers emergency braking (0xF3).

[0066] Specific scenario demonstration decision-making process: under highway working conditions, the real-time position 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. The initial calculation (medium risk), . The system activates the first-level adjustment, and the decision level rises from L1 to L2. At this time, the left lane radar detects a rapidly approaching vehicle, and the environmental risk submodule feature value jumps by 30% in 0.2 seconds, increases to 0.61, and the decision level rises to L3. After the vehicle control system receives the 0xF2 instruction, it starts the lane change program, and the braking system pre-charges.

[0067] The deviation monitoring system operates continuously after the decision instruction is output. The anomaly detector deployed on the data bus records the difference between the real-time sensor stream and the prediction model. When the same category deviation (such as position prediction error) appears continuously for a certain number of times (the default is 5 times), the phase analysis module is started. The phase delay feature The calculation process includes: selecting a feature point time series, performing Hilbert transform to obtain an analytical signal, calculating an instantaneous phase angle, and comparing a reference phase curve to obtain a delay amount. The delay amount is quantized to millisecond precision and stored as a 32-bit floating-point array.

[0068] The model correction engine updates the weights of the convolution kernel in the risk assessment model according to the following formula: and performs parameter recombination. The weight adjustment factor is calculated according to the following formula: wherein is the maximum allowed delay under the current working condition (default 300 ms). The convolution kernel weight matrix in the risk assessment model is updated as follows: wherein is the historical reference weight. The full connection layer bias item is adjusted synchronously: the bias vector is updated as follows: . A revision log is recorded for each correction operation, indicating the modification time, original parameter signature, and adjustment amplitude.

[0069] The parameter storage system implements version management. The corrected model parameters are stored in a specific partition of the historical database after SHA-256 hash calculation. The storage structure is a four-dimensional tensor: dimension one indexes the model components (convolution layer / full connection layer, etc.), dimension two records the parameter type (weight / bias), dimension three stores the timestamp, and dimension four saves the parameter data block. The database retains the last 50 correction records and uses LRU algorithm to eliminate old data. When a new model is initialized, the parameter snapshot closest to the current scenario in the database is retrieved, and the matching basis is the cosine similarity of the scenario feature vector.

[0070] The hardware interaction layer configures a protection mechanism. Before the decision instruction is output, cross-validation is required: the main decision module generates the instruction, and a simplified version of the model (retaining 30% of the parameters) performs parallel reasoning at the same time. When the difference between the two instructions exceeds one level, a safety lock is triggered, the instruction sending 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 vehicle stationary or low-speed states (<20 km / h), and the update process uses double buffering technology to avoid service interruption.

[0071] The running monitoring unit implements closed-loop verification. After each decision execution, the vehicle's actual response data (steering angle change rate, deceleration value, etc.) is compared with the expected model. The deviation persistence rate is calculated according to the following formula:

[0072] wherein: is the abnormal event count, is the total number of decisions. When ​If this continues for 10 minutes, the model parameters are forcibly reset to the baseline state before the last three revisions. Monitoring logs generate summary reports every minute, recording the decision success rate, number of model revisions, and storage usage.

[0073] The communication protocol incorporates an anti-collision mechanism. Decision commands are transmitted via independent CAN channels, with message IDs 0x5A0-0x5A3 (corresponding to level 4 commands). Parameter update messages are transmitted using the enhanced FlexRay protocol, with each frame containing a 128-byte parameter block and a 16-byte checksum. The historical database is accessed via a dedicated memory-mapped interface with a physical address range of 0x500000-0x5FFFFF. Access conflicts are resolved by a hardware arbiter. All communication transactions are timestamped with 100 nanosecond accuracy.

[0074] Example 6: The process of constructing a multidimensional decision parameter set begins with the data acquisition system. The speed fluctuation entropy calculation module is connected to the vehicle bus to obtain the longitudinal speed signal, and the sampling frequency is set to 100Hz. The raw speed data is first processed by a Butterworth low-pass filter with a cutoff frequency set to 5Hz to eliminate high-frequency noise. The preprocessed data stream enters the analysis window, and the window length is dynamically adjusted: a 10-second window is used for urban road conditions and a 20-second window is extended for highway conditions. The data within each window is subjected to a fast Fourier transform, and the spectrum analysis range is limited to the 0-10Hz frequency band. Energy distribution feature extraction focuses on the 0-1Hz low-frequency component and calculates the percentage of energy in this frequency band to the total energy. The entropy value calculation uses an improved Shannon entropy formula, dividing the spectrum into 8 equal-width sub-bands, and the energy proportion of each sub-band is used as a probability input. When the low-frequency band energy proportion exceeds 75%, the system marks the current traffic flow risk mode, and the status flag is stored in the shared memory area.

[0075] Environmental interference sensitivity parameters are derived from multi-sensor consistency analysis. The detection results of the millimeter-wave radar, lidar, and camera for the same target are input into a difference calculation unit, which calculates the standard deviation of each sensor's position and velocity measurements. The sensitivity score is calculated using a three-layer neural network: the input layer receives the difference statistics of each sensor, the hidden layer contains 16 neurons, and the output layer generates a normalized score between 0 and 1. The road friction coefficient is obtained from the on-board road surface recognition system. The system estimates the current road adhesion coefficient in real time based on the relationship between tire slip rate and brake pressure, updating the value every 200 milliseconds. The three types of parameters are encapsulated as a structured data packet and transmitted to the decision center via a message queue.

[0076] The parameter monitoring system implements a hierarchical response strategy. The speed fluctuation entropy value sets two threshold values: the first threshold value is 0.65, and when it is exceeded, the head-up display warning icon is triggered; the second threshold value is 0.8, and the voice prompt is activated to suggest reducing speed. The environmental disturbance sensitivity and road friction coefficient combination evaluation uses a fuzzy logic system: when the sensitivity is >0.7 and the friction coefficient is <0.3, the system determines it as a high-risk combination; when the sensitivity is >0.5 and the friction coefficient is <0.5, it is evaluated as medium risk. Risk level conversion is achieved through a state machine, which includes a 5-second state retention time to prevent frequent jumps. After the emergency avoidance command is generated, the system takes over the control of the steering wheel and brake system, and executes the pre-programmed avoidance trajectory.

[0077] Traffic flow risk pattern recognition triggers a cooperative perception mechanism. The system obtains the motion state data of surrounding vehicles through V2X communication, focusing on collecting acceleration information. The data synchronization module aligns the time base of the host vehicle and surrounding vehicles, with a maximum allowed deviation of 50 milliseconds. The correlation analysis unit calculates the Pearson correlation coefficient of the host vehicle's low-frequency energy and the neighbor's acceleration, and the analysis window is consistent with the entropy value calculation window. When the correlation coefficient exceeds the preset value of 0.7, the path planning module receives adjustment instructions: the curvature radius of the desired trajectory increases by 20%, and the lane changing acceleration limit decreases by 30%. The adjusted path task is re-parameterized through a cubic spline curve to ensure smooth motion.

[0078] Hardware resource allocation follows the principle of dynamic priority. Parameter calculation tasks run on a dedicated digital signal processor, occupying no more than 40% of the computing resources. The emergency avoidance command generation channel is configured with a hardware watchdog timer, which automatically resets the processing unit if no heartbeat signal is received within 500 milliseconds. Data communication uses a dual-channel redundant design, with the main channel using CANFD bus and the backup channel using Ethernet transmission. Memory management reserves a dedicated buffer area for key parameters and implements a write protection mechanism to prevent accidental modification.

[0079] The fault handling system establishes a multi-layer defense system. The speed signal anomaly detector monitors data validity, and switches to a backup sensor when 10 consecutive sampling points exceed the reasonable range. The environmental parameter calculation unit is equipped with a result rationality checker that discards the current frame data when the difference value of three sensors exceeds the physical possible range. The road identification system sets a confidence threshold, and estimates the friction coefficient below 0.6 without updating. All fault events are recorded with detailed logs, including timestamp, fault code, and recovery measures.

[0080] The system maintenance interface supports remote configuration updates. Threshold parameters are stored in programmable read-only memory and can be modified online through the diagnostic interface. The algorithm module employs a plug-in architecture, with critical processing units supporting hot replacement without interrupting service. A version control system records the complete trajectory of each parameter adjustment, supporting rollback to any historical version. Maintenance operations require dual authentication: physical interface connection verification plus digital certificate validation.

[0081] A real-time debug information output channel transmits detailed operational status. The data stream contains current parameter values, risk level markers, and system load status. Debug information is encapsulated as fixed-format message packets, output to external analysis equipment through a USB interface. Message filtering mechanisms are dynamically configurable, allowing selection of a key parameter subset for transmission on demand.

[0082] Baseline data established during vehicle integration testing serves as the initial reference. Parameter distribution ranges collected during 500 km of real-world road testing are used to calibrate initial threshold values. After the system goes online, operational data is continuously collected, and parameter distribution statistical analysis is performed once a month to automatically adjust threshold boundaries to match changes in the actual traffic environment. The calibration process takes seasonal differences into account, with winter data and summer data establishing reference models separately.

[0083] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one from another entity or action without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0084] While the embodiments of the application have been illustrated and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and alterations can be made therein without departing from the spirit and scope of the application, which is defined by the following claims and their equivalents.

Claims

1. A risk assessment-based automated driving lane change decision-making method, comprising: Based on the real-time data of the vehicle's surrounding environment, multi-dimensional features are integrated to generate a spatiotemporal feature representation containing position, speed, and environmental information; Inputting the spatiotemporal feature representation into a risk assessment model, probabilistically modeling the abnormal risk during the lane change process, and outputting a risk feature vector with confidence; Constructing a risk distribution field based on the risk feature vector, and predicting a propagation path of the risk distribution field based on a sequence prediction model; 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 of the real-time risk feature vector and the multi-level decision threshold, the automatic driving lane change decision is realized.

2. The risk assessment-based automatic driving lane change decision method according to claim 1, characterized in that: The method of fusing multi-dimensional features based on real-time data of the vehicle's surrounding environment to generate a spatiotemporal feature representation containing position, speed, and environmental information includes: Based on the real-time data collected by multiple sensors, timestamp alignment and missing value filling are performed, and a dynamic time warping algorithm is used to eliminate sensor sampling frequency differences to obtain a time-synchronized multi-source data sequence, where the real-time data includes vehicle position, speed, and environmental data; The multi-source data sequence is input into a feature fusion network, wherein a convolution operation is used in the spatial dimension to capture the road structure features, and a sliding window is used in the temporal dimension to extract the time series features, thereby obtaining preliminary spatiotemporal features; For the preliminary spatiotemporal features, the attention mechanism is used to calculate the correlation weights between different sensor data, and the features are weightedly fused according to the weights to highlight the contribution of key sensor data and obtain weighted spatiotemporal features; The weighted spatiotemporal features are input into a dimensionality reduction model, and redundant information is removed through nonlinear transformation, while retaining key features to generate a low-dimensional spatiotemporal feature representation containing position, speed and environmental information.

3. The risk assessment-based automatic driving lane change decision method according to claim 2, characterized in that: The input of the spatiotemporal feature representation into the risk assessment model, probabilistic modeling of the abnormal risk during the lane change process, and output of a risk feature vector with confidence level include: Inputting the spatiotemporal feature representation into a multi-layer network and mapping it to a high-dimensional latent space through nonlinear transformation to obtain a deep feature representation; Based on the deep feature representation, a probabilistic regression model is constructed, and a kernel function combination is used to capture the nonlinear relationship and periodic changes of the driving state to obtain 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 through an inference algorithm to obtain a probabilistic feature distribution with a confidence interval; For the probabilistic feature distribution, a distance detection algorithm is used to calculate the distance between each feature point and the normal driving distribution, and abnormal feature points are screened according to a preset confidence threshold to generate a risk feature vector with confidence.

4. The risk assessment-based automatic driving lane change decision method according to claim 3, characterized in that: The step of constructing a risk distribution field according to the risk feature vector and predicting a propagation path of the risk distribution field based on a sequence prediction model includes: Based on the risk feature vector and combined with road topology information, each risk feature point is mapped to a two-dimensional coordinate, and an interpolation algorithm is used to perform spatial interpolation on the discrete risk feature points to generate a preliminary risk distribution field; Based on the preliminary risk distribution field and combined with the time dimension information, a spatiotemporal interpolation algorithm is used to dynamically modify the risk distribution field to obtain a risk distribution field that changes over time; The time-varying risk distribution field is input into a sequence prediction model to construct a road node graph model, and an attention mechanism is used to capture the risk propagation dependency between nodes to obtain an 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 of future time steps are predicted to generate the propagation prediction results of the dynamic risk distribution field.

5. The risk assessment-based automatic driving lane change decision method according to claim 4, characterized in that: 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, including: Based on historical normal driving data, a normal driving generator based on a generative adversarial network is trained to generate simulated data that meets safe driving characteristics; Build 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 characteristics 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, and combined with the dynamic changes of 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.

6. The risk assessment-based automatic driving lane change decision method according to claim 5, characterized in that: The method of implementing an automatic driving lane change decision based on a comparison result between a real-time risk feature vector and a multi-level decision threshold includes: Comparing the real-time risk feature vector with a multi-level decision threshold, and using a decision model to dynamically adjust the decision level based on the confidence level of the risk feature and the degree of deviation from the threshold; Based on the decision level, a lane change instruction or a current lane maintenance instruction is generated to implement automatic driving lane change decision.

7. The risk assessment-based automatic driving lane change decision method according to claim 6, characterized in that: The method further comprises: When the deviation between real-time data and predicted data exceeds the set number of times, the phase delay feature of the feature point is extracted; Adjusting the parameter weight ratio in the risk assessment model according to the phase delay characteristics; The modified model parameters are stored in the historical database as the benchmark values ​​for the next model initialization.

8. The risk assessment-based automatic driving lane change decision method according to claim 7, characterized in that: The method further comprises: Construct a multidimensional decision parameter set including speed fluctuation entropy, environmental disturbance sensitivity and road friction coefficient; When a single parameter exceeds the first-level threshold, a status prompt is triggered, and when the combined effect of at least two parameters exceeds the second-level threshold, an emergency avoidance instruction is triggered.

9. The risk assessment-based automatic driving lane change decision method according to claim 8, characterized in that: The calculation of the speed fluctuation entropy value includes: Perform frequency domain transformation on the velocity data within the specified time window to extract the energy distribution characteristics of the preset frequency band; The speed fluctuation entropy value is calculated based on the energy distribution characteristics, and when the proportion of low-frequency energy exceeds a preset proportion, it is determined to be a traffic flow risk mode.

10. The risk assessment-based automatic driving lane change decision method according to claim 9, characterized in that: The method further comprises: When a traffic flow risk pattern is identified, the acceleration data of surrounding vehicles is collected simultaneously; A correlation analysis is performed on the acceleration data and the low-frequency energy. If the correlation coefficient is greater than a preset correlation value, a path adjustment task is added to the lane change decision.

Images