Trajectory prediction-based ice and snow road vehicle lane-changing risk dynamic assessment system

Abstract

This invention relates to the field of intelligent driving safety technology, specifically to a dynamic risk assessment system for vehicle lane-changing on icy and snowy roads based on trajectory prediction. The system includes: a prediction sampling module, which acquires the probability density distribution of the target vehicle's future position and extracts trajectory sampling points using a Markov chain Monte Carlo method; a risk field construction module, which calculates the risk field intensity of each sampling point based on the vehicle's equivalent mass, a human occupancy model, and a future time discounting weight function, and aggregates them to form a vehicle risk field; a binary conflict calculation module, which matches the vehicle risk fields of the vehicle itself and surrounding vehicles, calculates the sum of the intensity products within the overlapping area, and obtains a binary conflict field; and a comprehensive risk assessment module, which overlays all binary conflict fields to generate a real-time total conflict field. This system achieves a refined and quantitative assessment of dynamic, multi-vehicle interaction risks during lane-changing on icy and snowy roads.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving safety technology, and in particular to a dynamic assessment system for vehicle lane-changing risks on icy and snowy roads based on trajectory prediction. Background Technology

[0002] Existing methods for assessing vehicle lane-changing risks often employ deterministic models or models based on simplified probability distributions in the trajectory prediction stage. These methods output one or more most probable trajectories or use simple models such as Gaussian distributions to describe positional uncertainties, failing to fully utilize the complex, non-Gaussian probability distribution characteristics caused by driving behavior and road conditions. This makes the prediction results unable to adequately represent the true probability structure of the space that the vehicle may occupy in the future, resulting in inherent biases in the inputs for subsequent risk assessments.

[0003] In the risk assessment phase, conventional techniques often rely on models of vehicle geometry or physical collision energy to construct risk indicators. These methods tend to construct static risk fields or only consider immediate physical conditions, neglecting the dynamic impact of driver behavior and intentions on risk distribution. They also lack mechanisms for reasonably discounting long-term risk events. This method of risk representation is insufficient in reflecting drivers' personalized risk decisions and the cognitive characteristic of risk decay over time.

[0004] Due to the low coefficient of adhesion on icy and snowy roads, vehicles are prone to changes in dynamic characteristics such as sideslip, increased braking distance, and delayed handling response. At the same time, drivers tend to be more conservative or hesitant in icy and snowy conditions, further exacerbating the uncertainty of vehicle trajectory. Existing deterministic or simplified probabilistic models for trajectory prediction, as well as static risk assessment methods, will have significantly increased prediction deviations for future vehicle trajectories under the special condition of icy and snowy roads. The accuracy and real-time performance of risk assessment are even more difficult to meet the requirements of lane-changing safety in intelligent driving. The limitations of such technologies are particularly prominent in lane-changing scenarios on low-adhesion icy and snowy roads. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a dynamic assessment system for vehicle lane-changing risks on icy and snowy roads based on trajectory prediction.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a dynamic assessment system for vehicle lane-changing risks on icy and snowy roads based on trajectory prediction, comprising:

[0007] The prediction sampling module obtains the probability density distribution of the target vehicle's position at future times, and extracts a series of future trajectory sampling points from the probability density distribution of the position at future times using the Markov chain Monte Carlo method.

[0008] The risk field construction module calculates the risk field intensity value generated in space for each of the future trajectory sampling points based on the vehicle's equivalent mass, human occupation model, and future time discounting weight function, and aggregates them to form a vehicle risk field that characterizes the intensity of the risk that the vehicle will encounter at various points in space in the future.

[0009] The binary conflict calculation module, when assessing the risk interaction between the vehicle and surrounding vehicles, spatially matches the vehicle risk field of the vehicle with the vehicle risk fields of the surrounding vehicles, calculates the sum of the products of the vehicle risk field intensity value of the vehicle and the vehicle risk field intensity value of the surrounding vehicles in the overlapping spatial region, and obtains the binary conflict field characterizing the conflict intensity between the vehicle and a single surrounding vehicle.

[0010] The comprehensive risk assessment module calculates the binary conflict field between the vehicle and each of the surrounding vehicles in a scenario where the vehicle interacts with multiple surrounding vehicles. It then overlays all the binary conflict fields to generate a real-time total conflict field that reflects the comprehensive risks faced by the vehicle during the entire lane-changing process.

[0011] As a further aspect of the present invention, obtaining the probability density distribution of the target vehicle's future position includes:

[0012] The vehicle motion features and vehicle interaction features are obtained from continuous historical moments of the vehicle and surrounding vehicles. The vehicle motion features include instantaneous velocity values, instantaneous acceleration values, planar position coordinates, and vehicle sideslip angle. The vehicle interaction features include the relative velocity and relative position of the vehicle and each surrounding vehicle.

[0013] The vehicle motion features and vehicle interaction features at the continuous historical moments are time-aligned and dimension-stitched to form a time series feature input matrix containing multi-dimensional features of multiple vehicles at continuous time steps.

[0014] The time-series feature input matrix is ​​processed by an environmental attention network to extract the temporal dependencies, spatial dependencies, and interactive attention relationships between vehicles. Based on these temporal dependencies, spatial dependencies, and interactive attention relationships, the probability density distribution of the target vehicle's future position is predicted.

[0015] As a further aspect of the present invention, the step of processing the time-series feature input matrix through an environmental attention network to extract temporal dependencies, spatial dependencies, and interactive attention relationships between vehicles, and predicting the probability density distribution of the target vehicle's future position based on these temporal dependencies, spatial dependencies, and interactive attention relationships, includes:

[0016] The time series feature input matrix is ​​processed using a long short-term memory encoder. The long short-term memory encoder encodes the historical trajectory sequence of each vehicle, extracts and outputs the encoded feature vector of each vehicle at each historical moment, and the encoded feature vector contains the temporal dependency relationship of the vehicle.

[0017] The encoded feature vectors of all vehicles at the same time are used as nodes to construct a graph structure based on the spatial distance relationship between vehicles. The graph structure is then input into a graph attention network. The graph attention network adaptively calculates the attention weights between nodes to generate an interaction attention weight matrix that reflects the relative importance between vehicles and outputs an interaction feature vector that contains the interaction attention relationship between vehicles.

[0018] Based on the interaction feature vector, a three-dimensional social tensor centered on the target vehicle is constructed. The two planar dimensions of the three-dimensional social tensor represent a spatial grid, and the channel dimension is filled with the interaction feature vectors of other vehicles in the corresponding spatial grid.

[0019] The three-dimensional social tensor is input into a convolutional social pooling network with a squeeze excitation structure. The convolutional social pooling network with a squeeze excitation structure extracts the local spatial patterns around the target vehicle through convolution operations and uses a channel attention mechanism to weight different spatial feature channels, outputting spatial context features that contain the spatial dependency relationship between the target vehicle and the surrounding environment.

[0020] The final hidden state corresponding to the temporal dependency, the interaction feature vector corresponding to the interaction attention relationship, and the spatial context feature corresponding to the spatial dependency are fused and input into the combined structure of the long short-term memory decoder and the hybrid density network. The long short-term memory decoder is responsible for decoding the future time sequence, and the hybrid density network processes the output of the decoder at each future time to estimate the two-dimensional Gaussian mixture distribution parameters of the target vehicle's planar position at the future time. The two-dimensional Gaussian mixture distributions of all future times together constitute the position probability density distribution of the future time.

[0021] As a further aspect of the present invention, a series of future trajectory sampling points are extracted from the position probability density distribution at the future time using the Markov chain Monte Carlo method, including:

[0022] Initialize a Markov chain whose initial state is a sequence of future trajectory points randomly drawn from the position probability density distribution of the future time;

[0023] Based on the parameters of the two-dimensional Gaussian mixture distribution, a proposal distribution is defined to generate a new sequence of candidate future trajectory points from the current Markov chain state;

[0024] Calculate the joint probability density of the candidate future trajectory point sequence under the position probability density distribution at the current future time, and compare it with the joint probability density of the current state. Based on a preset acceptance criterion, decide whether to accept the candidate future trajectory point sequence as a new state of the Markov chain.

[0025] Repeat the steps from defining the proposed distribution to deciding whether to accept the candidate state until the state transition of the Markov chain reaches a preset number of samples, which ensures that the chain reaches a stable distribution;

[0026] From the Markov chain that has reached a stationary distribution, states are extracted at preset intervals. Each extracted state is used as a complete sampling point of the future trajectory, and finally a set of samples is obtained to approximate the position probability density distribution of the future time.

[0027] As a further aspect of the present invention, the step of calculating the risk field intensity value generated in space for each of the future trajectory sampling points based on the vehicle's equivalent mass, the human occupancy model, and the future time discounting weight function includes:

[0028] The equivalent mass of the target vehicle is obtained by looking up the corresponding physical parameters in a table. The equivalent mass of the vehicle combines the actual mass and dynamic characteristics of the vehicle.

[0029] Based on the artificial occupancy model, the vehicle planar position represented by the future trajectory sampling point is expanded into a polygonal spatial region occupied by the vehicle in the planar position. The artificial occupancy model defines the safety boundary of the vehicle outline expansion.

[0030] A future time discount weight is defined for each predicted future moment. The future time discount weight decreases monotonically as the predicted moment is delayed, and the more distant the future moment is, the smaller the weight of its impact on the current risk.

[0031] For each spatial grid point within the polygonal spatial region, the equivalent mass of the vehicle, the future time discount weight, and the probability value of the future trajectory sampling point itself are multiplied to obtain the basic value of the risk field intensity contributed by the future trajectory sampling point to the spatial grid point.

[0032] Traverse all sampling points in the future trajectory sampling point sample set. For the same spatial grid point, accumulate the basic value of the risk field intensity contributed by all future trajectory sampling points to it to obtain the final risk field intensity value of the spatial grid point. The risk field intensity values ​​of all spatial grid points together constitute a continuous risk intensity spatial distribution, that is, the vehicle risk field.

[0033] As a further aspect of the present invention, the sum of the products of the vehicle risk field intensity value of the self-vehicle and the vehicle risk field intensity values ​​of the surrounding vehicles within the overlapping spatial region, to obtain a binary conflict field characterizing the conflict intensity between the self-vehicle and a single surrounding vehicle, includes:

[0034] The vehicle risk field of the vehicle itself and the vehicle risk fields of the surrounding vehicles are mapped onto the same high-resolution spatial discrete grid, so that the two vehicle risk fields have the same spatial coordinate definition and grid cell division.

[0035] Traverse each grid cell of the spatial discrete grid, read the intensity value of the grid cell in the vehicle risk field of the vehicle itself, and read the intensity value of the grid cell in the vehicle risk field of the surrounding vehicles.

[0036] The risk field intensity value of the self-vehicle is multiplied by the risk field intensity values ​​of the surrounding vehicles to obtain the conflict intensity contribution value on the grid cell.

[0037] Determine whether the conflict intensity contribution value on the grid cell is greater than zero. If it is greater than zero, mark the grid cell as a risk field overlap region.

[0038] The conflict intensity contribution values ​​of all grid cells marked as overlapping areas of the risk field are summed to obtain the scalar value of the binary conflict field, which quantitatively describes the overall conflict risk between the vehicle and surrounding vehicles in terms of space occupation.

[0039] As a further aspect of the present invention, the step of spatially matching the vehicle risk field of the self-vehicle with the vehicle risk fields of surrounding vehicles, and calculating the sum of the products of the vehicle risk field intensity values ​​of the self-vehicle and the vehicle risk field intensity values ​​of surrounding vehicles in the overlapping spatial region to obtain a binary conflict field characterizing the conflict intensity between the self-vehicle and a single surrounding vehicle, further includes the step of approximating continuous spatial integration through Monte Carlo integration:

[0040] Within the spatial region where the risk fields of the vehicle and surrounding vehicles overlap, a specified number of sampling points are randomly generated, and the spatial coordinates of the sampling points follow a uniform distribution within the spatial region.

[0041] For each randomly generated sampling point, the intensity value of the sampling point in the vehicle risk field of its own vehicle and the intensity value of the sampling point in the vehicle risk field of the surrounding vehicles are calculated by bilinear interpolation.

[0042] Multiply the calculated risk field intensity value of the vehicle by the risk field intensity values ​​of the surrounding vehicles to obtain the contribution of the sampling point to the conflict intensity.

[0043] Calculate the average contribution of all random sampling points, and multiply the average by the area of ​​the overlapping spatial region of the risk field to obtain an approximate estimate of the continuous spatial integral. The approximate estimate is the scalar value of the binary conflict field.

[0044] As a further aspect of the present invention, the step of calculating the binary conflict fields between the vehicle and each surrounding vehicle, and superimposing all the binary conflict fields to generate a real-time total conflict field reflecting the comprehensive risks faced by the vehicle throughout the lane-changing process includes:

[0045] For each surrounding vehicle within the current perception range of the vehicle, a binary conflict field scalar value is calculated using the product summation method or the Monte Carlo integration method based on the vehicle risk field of the vehicle and the vehicle risk fields of the surrounding vehicles.

[0046] The calculated binary conflict field scalar values ​​of the vehicle and each surrounding vehicle are stored according to the vehicle identifier.

[0047] Perform an arithmetic summation operation on all stored binary collision field scalar values ​​at the current moment, and the sum is the real-time total collision field scalar value at the current moment;

[0048] As time progresses and the scenario is updated, the calculation and accumulation operations are repeatedly performed to generate a real-time total conflict field scalar value sequence that changes over time. This real-time total conflict field scalar value sequence constitutes the real-time total conflict field that reflects the dynamic evolution of risk.

[0049] As a further aspect of the present invention, it also includes:

[0050] The dynamic risk update module repeatedly generates a real-time total conflict field at preset fixed time intervals, dynamically updating and outputting the risk status of continuously changing lane-changing scenarios. Specifically, this includes:

[0051] Set a system update cycle, and at the beginning of each update cycle, synchronously acquire the latest sensor data of the vehicle and all surrounding vehicles;

[0052] The vehicle motion features and vehicle interaction features are updated based on the latest sensor data, and the time series feature input matrix is ​​refreshed accordingly.

[0053] Using the refreshed time-series feature input matrix, the environmental attention network is driven to re-execute trajectory prediction, generating an updated position probability density distribution for future moments.

[0054] Based on the updated location probability density distribution for future moments, the entire process of Markov chain Monte Carlo sampling, vehicle risk field construction, binary conflict field calculation, and real-time total conflict field generation is re-executed.

[0055] Output the real-time total conflict field scalar value calculated in the current update cycle, as well as the binary conflict field scalar value between the vehicle and each key surrounding vehicle.

[0056] As a further aspect of the present invention, the instantaneous acceleration values ​​in the vehicle motion characteristics are longitudinal and lateral accelerations corrected for the adhesion coefficient of icy and snowy road surfaces. The correction process includes:

[0057] The estimated ice and snow adhesion coefficient of the current road surface is obtained through vehicle-mounted sensors or road condition recognition modules;

[0058] Read the raw longitudinal acceleration request and raw lateral acceleration request output by the vehicle dynamics controller;

[0059] Multiply the original longitudinal acceleration request by the estimated ice and snow adhesion coefficient to obtain the upper limit of safe longitudinal acceleration under icy and snowy road surface. If the original request exceeds the upper limit of safe longitudinal acceleration, the upper limit of safe longitudinal acceleration is used as the instantaneous value of longitudinal acceleration actually used.

[0060] Multiply the original lateral acceleration request by the estimated ice and snow adhesion coefficient to obtain the upper limit of safe lateral acceleration under icy and snowy road surface. If the original request exceeds the upper limit of safe lateral acceleration, the upper limit of safe lateral acceleration is used as the instantaneous value of lateral acceleration actually used.

[0061] The safe longitudinal acceleration upper limit value and the safe lateral acceleration upper limit value, after being processed by the upper limit constraint, are used as the final instantaneous acceleration values ​​and input into the time series feature input matrix.

[0062] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0063] This system addresses the low adhesion characteristics of icy and snowy roads by correcting the vehicle acceleration for ice and snow adhesion coefficients during the trajectory prediction stage. This makes the probability distribution of trajectory prediction more closely match the actual dynamic performance of vehicles on icy and snowy roads. At the same time, the entire process of risk field construction and conflict assessment takes into account the uncertainty of vehicle motion on icy and snowy roads, enabling more accurate capture of dynamic risk interactions between vehicles during lane changes on icy and snowy roads, thus meeting the safety assessment needs of intelligent driving on icy and snowy roads.

[0064] A Markov chain Monte Carlo method is employed to sample future trajectories from the location probability density distribution. This method effectively handles complex probability distributions jointly generated by vehicle dynamics and driver behavior, especially considering the non-Gaussian and multimodal characteristics that may occur on icy and snowy roads. By directly sampling the probability density function, a series of trajectory points and their probability weights are obtained, fully preserving the original statistical characteristics of prediction uncertainty. This avoids the loss of probabilistic information caused by traditional methods that assume distribution forms or select a few typical trajectories. The sampling results provide a more reliable probabilistic basis for subsequent risk calculations, enabling risk estimation to more accurately reflect the various possibilities of the vehicle's future state and improving the authenticity of the input data for risk assessment.

[0065] A vehicle risk field is constructed that integrates a manual occupancy model and future time-discounted weights. The strength of the risk field is jointly determined by the vehicle's equivalent mass, the output value of the manual occupancy model, and the future time-discounted weights. The manual occupancy model quantifies the driver's personalized behavioral characteristics as a spatial probability distribution, enabling the risk field to reflect individualized behavioral risks. The future time-discounted weight function applies exponential decay based on the time of the risk event, simulating the different psychological levels of importance that human drivers place on immediate dangers and long-term threats. This construction method makes the risk field a dynamically changing field, whose spatial distribution changes not only with the predicted vehicle location but also with the driver's behavioral characteristics and time perception. It achieves a fine-grained characterization of risk, integrating objective physical risk and subjective behavioral perception risk within a unified mathematical framework. Attached Figure Description

[0066] Figure 1 This is a time-series diagram of the dynamic assessment system for vehicle lane-changing risks on icy and snowy roads based on trajectory prediction, as described in this invention.

[0067] Figure 2 A flowchart for obtaining the probability density distribution of the future location of the target vehicle;

[0068] Figure 3 A flowchart illustrating the processing of time-series feature input matrices for an environmental attention network;

[0069] Figure 4 Heat map of the spatial distribution of risk areas for autonomous vehicles;

[0070] Figure 5 This is a line graph showing the evolution of the intensity of a binary conflict between the vehicle and multiple vehicles behind it over time. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0072] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0073] See Figure 1 The system obtains the probability density distribution of the target vehicle's future position through a prediction sampling module and extracts a series of future trajectory sampling points. These sampling points represent the possible future planar positions of the target vehicle and their probabilities. The risk field construction module calculates the spatial risk field intensity value generated by each future trajectory sampling point based on the vehicle's equivalent mass, a human occupancy model, and a future time discounting weight function, and aggregates them to form a vehicle risk field representing the intensity of risk the vehicle will face at various points in space. When assessing the risk interaction between the vehicle and surrounding vehicles, the binary conflict calculation module spatially matches the vehicle's risk field with the risk fields of surrounding vehicles, calculates the sum of the products of their intensity values ​​in the overlapping spatial region, and obtains a binary conflict field representing the conflict intensity between the vehicle and a single surrounding vehicle. In complex scenarios where the vehicle interacts with multiple surrounding vehicles, the comprehensive risk assessment module calculates the binary conflict field between the vehicle and each surrounding vehicle separately, and superimposes all binary conflict fields to finally generate a real-time total conflict field reflecting the comprehensive risks faced by the vehicle throughout the lane-changing process.

[0074] See Figure 2In one embodiment of the present invention, the system acquires vehicle motion characteristics and vehicle interaction characteristics from sensors of the vehicle and surrounding vehicles at continuous historical moments. The vehicle motion characteristics include instantaneous vehicle speed, planar position coordinates, vehicle sideslip angle, and instantaneous longitudinal and lateral acceleration values ​​corrected for the icy / snowy road surface adhesion coefficient. The vehicle interaction characteristics include the relative speed and relative position of the vehicle and each surrounding vehicle. The correction process for the instantaneous acceleration values ​​is based on the icy / snowy road surface adhesion coefficient, specifically obtained through onboard sensors or a separate road surface condition recognition module to obtain an estimated icy / snowy road surface adhesion coefficient for the current road condition. The original longitudinal acceleration request and the original lateral acceleration request output by the vehicle dynamics controller are read. The original longitudinal acceleration request is multiplied by the estimated icy / snowy road surface adhesion coefficient, and the product is used as the upper limit of safe longitudinal acceleration under icy / snowy road conditions. When the absolute value of the original longitudinal acceleration request exceeds this upper limit of safe longitudinal acceleration, the signed upper limit of safe longitudinal acceleration is used as the actual instantaneous longitudinal acceleration value. The original lateral acceleration request is multiplied by the estimated ice and snow adhesion coefficient. The product is used as the upper limit of safe lateral acceleration on icy and snowy roads. When the absolute value of the original lateral acceleration request exceeds this upper limit, the signed upper limit of safe lateral acceleration is used as the actual instantaneous lateral acceleration value. The upper limits of safe longitudinal acceleration and safe lateral acceleration, after upper limit constraint processing, are determined as the final instantaneous acceleration values ​​and prepared for input. The low adhesion characteristics of icy and snowy roads are the core consideration for acceleration correction. The corrected instantaneous acceleration value perfectly matches the physical limits of vehicle acceleration, braking, and steering on icy and snowy roads, avoiding trajectory prediction deviations caused by insufficient road adhesion, and making subsequent trajectory predictions based on this feature more consistent with the actual driving scenario on icy and snowy roads.

[0075] In practice, these modified vehicle motion and interaction features are organized to form a time-series feature input matrix. The system continuously collects feature data of the vehicle and multiple surrounding vehicles at multiple moments within a historical time window. For each tracked vehicle, its vehicle motion and interaction features constitute a feature vector at each corresponding historical moment. The feature vectors of the vehicle and all surrounding vehicles at the same moment are aligned and concatenated to form a global feature vector that comprehensively describes the overall traffic situation at that moment. Then, the global feature vectors from multiple consecutive historical moments are stacked in chronological order to finally form a three-dimensional time-series feature input matrix, the dimension of which is the time step multiplied by the number of vehicles multiplied by the total feature dimension. This time-series feature input matrix provides a structured input data foundation for the subsequent environmental attention network.

[0076] In some embodiments, the environmental attention network processes a time-series feature input matrix to extract complex relationships. The environmental attention network performs multi-level encoding and attention computation on the input time-series feature input matrix. First, the network processes the temporal information in the input matrix, extracting the temporal dependencies inherent in the vehicle's historical motion patterns. Next, the network analyzes the relationships between feature vectors of different vehicles at the same time, extracting interactive attention relationships by calculating the spatial correlations between vehicles. Simultaneously, the network models the spatial environment centered on the target vehicle, extracting the spatial dependencies between the target vehicle and its surrounding physical environment. Finally, based on the extracted temporal dependencies, spatial dependencies, and interactive attention relationships, the environmental attention network decodes and estimates probability distribution parameters to output the probability density distribution of the target vehicle's position at multiple predicted future times. This probability distribution describes the probabilistic characteristics of the space the target vehicle may occupy in the future.

[0077] It can be understood that the entire feature processing and trajectory prediction process is an end-to-end system. It begins with acquiring the original motion and interaction features, including the corrected acceleration, then constructs a time-series feature input matrix, followed by deep feature extraction and relationship modeling via an environmental attention network, ultimately generating a probability distribution describing the uncertainty of the future trajectory. Correcting the acceleration features is a necessary step to adapt to the low-adhesion conditions of icy and snowy roads, and its mathematical constraints can be expressed as follows:

[0078]

[0079] in: This indicates the actual instantaneous value of safe acceleration (longitudinal or lateral) after correction. This indicates the raw acceleration request value (longitudinal or lateral) output by the vehicle dynamics controller. This represents the current road surface estimated ice and snow adhesion coefficient obtained from sensors or recognition modules. This is a sign function used to maintain the direction of acceleration. The function ensures that the output value does not exceed the upper limit of the ice and snow adhesion coefficient. This processing directly integrates physical constraints into the feature level, enabling subsequent trajectory prediction to be performed under conditions that conform to the dynamic laws of icy and snowy roads. The various stages of feature extraction, time series construction, and network processing are closely integrated, jointly achieving quantitative prediction of the uncertainty of vehicle future behavior under the special and dangerous conditions of icy and snowy roads, providing probabilistic trajectory input for subsequent risk field construction and conflict assessment.

[0080] See Figure 3In one embodiment of the present invention, a Long Short-Term Memory (LSTM) encoder processes a time-series feature input matrix to extract the temporal dependencies of vehicles. The input to the LSM encoder is an organized time-series feature input matrix containing feature vectors of the vehicle and multiple surrounding vehicles at consecutive historical moments. The LSM encoder constructs a dedicated encoder instance for each independently tracked vehicle in the scene. Each LSM encoder instance receives the historical feature sequence of the corresponding vehicle and iteratively processes it along the time steps. At each historical moment, the LSM encoder unit receives the feature vector of the current moment and the hidden state and cell state from the previous moment, updating the hidden state and cell state of the current moment through an internal gating mechanism. After processing the complete historical time series, the LSM encoder outputs the encoded feature vector of the vehicle at each historical moment. This encoded feature vector is a compressed representation of all historical motion and state information of the vehicle up to that moment, implying the temporal dependencies of the vehicle's trajectory.

[0081] In some embodiments, the graph attention network processes encoded feature vectors to extract interactive attention relationships between vehicles. At each historical moment, the encoded feature vectors of all vehicles are used as node features of the graph attention network. The connections between nodes, i.e., the edges of the graph, are constructed based on the spatial distance relationships between vehicles. It can be understood that the connection strategy can employ the K-nearest neighbor method or the distance threshold method to ensure that each vehicle establishes a connection with its spatially neighboring vehicles. The graph attention network performs multi-layer information transfer and aggregation on these nodes, and its core mechanism lies in calculating the attention weights between nodes. For the target node, the graph attention network calculates its correlation with the feature vectors of all neighboring nodes and normalizes these correlations into attention weights using the Softmax function. These attention weights adaptively reflect the relative importance of different neighboring vehicles in predicting the future behavior of the target vehicle. After processing by the graph attention network, the encoded feature vector of each vehicle is updated to an interactive feature vector containing the interactive influence of surrounding vehicles, while simultaneously generating an interactive attention weight matrix reflecting the relative importance between vehicles.

[0082] In practice, a 3D social tensor centered on the target vehicle is constructed to capture spatial dependencies. Interaction feature vectors serve as the foundational data for constructing the 3D social tensor. A 2D spatial grid map centered on the target vehicle's current location is defined, covering a certain area around the target vehicle. The two planar dimensions of the 3D social tensor correspond to the row and column indices of this spatial grid map, respectively. For each spatial grid cell of the 3D social tensor, it is checked whether other vehicles are located within or projected onto that grid cell. If other vehicles exist within a grid cell, their interaction feature vectors are filled into the channel dimension of the 3D social tensor corresponding to that grid cell's position. If multiple vehicles exist within a grid cell, their interaction feature vectors can be concatenated or pooled along the channel dimension. For grid cells not occupied by any vehicles, their channel dimension is filled with zero vectors. The resulting 3D social tensor is a structured data block whose spatial dimension encodes the distribution of vehicles in the physical world, and whose channel dimension encodes the interaction feature information of these vehicles.

[0083] Optionally, a convolutional social pooling network with a squeeze-excitation structure processes the 3D social tensor to generate spatial context features. This network receives the 3D social tensor as input. The network first operates on the 3D social tensor through a series of convolutional layers. The convolutional kernels slide along the spatial dimension, extracting pattern features of the local region surrounding the target vehicle, such as the distribution shape and density gradient of the vehicle group. The core component of the convolutional social pooling network with a squeeze-excitation structure is the squeeze-excitation module. This module processes the feature map obtained from the convolutional operations. First, it performs global average pooling in the spatial dimension, compressing the 2D information of each feature channel into a scalar; this operation is called "squeezing." Then, it learns the weights of each channel through a bottleneck structure composed of fully connected layers and normalizes the weights to between 0 and 1 using the sigmoid function; this operation is called "excitation." The learned channel weights represent the importance of different feature channels. Finally, the normalized channel weights are multiplied channel-by-channel with the original feature map to recalibrate the feature channels. The output of a convolutional social pooling network with a squeezed excitation structure is a spatial context feature weighted by channel attention, which implies the complex spatial dependencies between the target vehicle and its surrounding environment.

[0084] It is understandable that feature fusion and probability distribution generation are the final steps in trajectory prediction. The final hidden state obtained by the Long Short-Term Memory (LSTM) encoder after processing the entire historical sequence represents a summary of temporal dependencies. The interaction feature vector output by the graph attention network represents a summary of interaction attention relationships. The spatial context features output by the convolutional social pooling network with a squeezed excitation structure represent a summary of spatial dependencies. These three sources of features are concatenated or combined through other fusion layers into a unified context vector. This unified context vector is input into a LSM decoder, which progressively generates a sequence of hidden states for future time steps in an autoregressive manner. At each future prediction time step, the output hidden state of the LSM decoder is fed into a hybrid density network. The hybrid density network consists of fully connected layers, and its output parameters define a two-dimensional Gaussian mixture distribution. The hybrid density network outputs multiple sets of Gaussian distribution parameters for each prediction time step, including the weights, mean vector, and covariance matrix of each Gaussian component. The two-dimensional Gaussian mixture distributions of all predicted future times together constitute the probability density distribution of the target vehicle's future position. Its mathematical expression involves a linear combination of multiple Gaussian components. For the predicted time t, the position... probability density It can be represented as:

[0085]

[0086] in: It is the preset number of Gaussian mixture components. It is the mixing weight of the k-th Gaussian component at prediction time t, satisfying , It is the two-dimensional mean vector of the k-th Gaussian component at prediction time t. It is the two-dimensional covariance matrix of the k-th Gaussian component at prediction time t. Indicates For the mean, Let be the two-dimensional Gaussian probability density function of the covariance matrix.

[0087] In one embodiment of the present invention, the Markov chain Monte Carlo method extracts future trajectory sampling points from the position probability density distribution at future time steps. The extraction process revolves around the construction and state transition of a Markov chain. The initial state of the Markov chain is a sequence of future trajectory points randomly extracted from the position probability density distribution at future time steps. This sequence contains the planar position coordinates of the target vehicle at all predicted time steps. Subsequently, a proposal distribution based on the current state is defined, which is used to generate a new candidate sequence of future trajectory points. The specific form of the proposal distribution needs to be compatible with the characteristics of the position probability density distribution. A typical approach is to construct a multidimensional Gaussian distribution with small variance centered on each position point of the current sequence as a perturbation source, and generate candidate sequences by adding random noise conforming to this Gaussian distribution to the current sequence. Considering the non-Gaussian, multimodal probability distribution characteristics of vehicle trajectories on icy and snowy roads, the Markov chain Monte Carlo method can accurately capture such trajectory uncertainties caused by icy and snowy roads during sampling, making it more suitable for trajectory sampling requirements on icy and snowy roads compared to traditional sampling methods.

[0088] In some embodiments, the acceptance criterion for Markov chain state transitions is calculated based on probability density comparison. The joint probability density of the candidate future trajectory point sequence under the position probability density distributions at future time points is calculated. The joint probability density is the product of the probability densities of the candidate sequence at all prediction time points, or other joint forms. Simultaneously, the joint probability density of the current Markov chain state sequence under the same position probability density distribution is calculated. The joint probability density of the candidate sequence is compared with the joint probability density of the current state sequence. Based on a preset acceptance criterion, such as the Metropolis-Hastings criterion, a decision is made on whether to accept the candidate future trajectory point sequence as the new state of the Markov chain. Specifically, this decision is made using an acceptance probability, which is the product of the ratio of the joint probability density of the candidate sequence to the joint probability density of the current state sequence and an adjustment factor associated with the proposal distribution. A uniform random number between 0 and 1 is generated. If this random number is less than the calculated acceptance probability, the candidate sequence is accepted, and the Markov chain state transitions to the new sequence; otherwise, the candidate sequence is rejected, and the Markov chain state remains unchanged.

[0089] The core stage of sampling is the repeated execution of state transition steps until the Markov chain reaches a stationary distribution. This involves repeatedly executing the complete steps from defining the proposal distribution to deciding whether to accept a candidate state. Each repetition constitutes an iteration of the Markov chain's state transition. The preset number of samples needs to be large enough to ensure that after a sufficient number of iterations, the Markov chain's state distribution converges to the target distribution, i.e., the probability density distribution of future positions. This process is called the Markov chain reaching a stationary distribution. The state transition process stops when the preset number of iterations is reached.

[0090] In practice, states are extracted as final samples from the Markov chain after iteration has reached a stationary distribution. States are then extracted from the Markov chain after iteration has stopped at preset intervals. These preset intervals are designed to reduce autocorrelation between samples and ensure their independence or weak correlation. For example, the chain's states are extracted every fixed number of iterations. Each extracted state, i.e., a complete sequence of future trajectory points, is treated as an independent future trajectory sampling point. Extraction continues until a predetermined number of future trajectory sampling points are obtained; these future trajectory sampling points together constitute a sample set. This sample set is extracted from a complex high-dimensional position probability density distribution and can be used to approximate the position probability density distribution at future times, providing discrete, computable trajectory samples for subsequent risk field calculations.

[0091] Optionally, the risk field construction module calculates the spatial risk field intensity value generated by each future trajectory sampling point. The calculation process relies on the vehicle's equivalent mass, the human occupancy model, and the future time discounting weight function. The vehicle's equivalent mass is obtained by looking up its physical parameters in a table. The vehicle's equivalent mass integrates the vehicle's actual mass and dynamic characteristics and is a scalar parameter. Based on the human occupancy model, the vehicle's planar position represented by the future trajectory sampling point is expanded into a polygonal spatial region occupied by the vehicle in that planar position. The human occupancy model defines the safety boundary for the vehicle's outline expansion, and the polygonal spatial region is a convex polygon or rectangular region that considers the vehicle's geometric dimensions and safety margin. A future time discounting weight is defined for each predicted future moment. This future time discounting weight is a function of the predicted time step, and its function value monotonically decreases as the predicted moment is delayed, making the influence of more distant future moments on the current risk assessment smaller. Drivers on icy and snowy roads tend to be more conservative or hesitant in their actions. The manual occupancy model fully incorporates the characteristics of driver behavior on icy and snowy roads when quantifying individual driver behavior. The setting of future time discount weights also takes into account the temporal evolution characteristics of risk events on icy and snowy roads, making the calculation of risk field intensity more consistent with the lane-changing risk characteristics on icy and snowy roads.

[0092] The specific calculation process for the risk field intensity value involves traversing all sampling points in the future trajectory sampling point sample set and all spatial grid points within the polygonal spatial region corresponding to each sampling point. For any future trajectory sampling point in the sample set, this sampling point corresponds to a specific predicted time sequence. The future time discount weight associated with this sampling point is denoted as... ,in This represents the prediction time step index of the sampling point. The equivalent vehicle mass corresponding to this sampling point is denoted as... The probability of this sampling point occurring itself is denoted as... This probability value can be directly evaluated from the probability density distribution that generated the sampling point. For any spatial grid point within the polygonal spatial region, the basic value of the risk field intensity contributed by the future trajectory sampling point to that spatial grid point is... The calculation follows the following relationship:

[0093]

[0094] in: This represents the basic value of the risk field intensity generated by a single future trajectory sampling point at a single spatial grid point. Indicates the equivalent mass of the vehicle. This represents the future time discount weight corresponding to the prediction time of this sampling point. This represents the probability of the future trajectory sampling point itself appearing. All spatial grid points within the polygonal spatial region are traversed, and a basic risk field intensity value is calculated for each grid point. After calculating for one future trajectory sampling point, the above calculation process is repeated for the next future trajectory sampling point in the sample set. Finally, for the same spatial grid point in space, the basic risk field intensity values ​​contributed by all future trajectory sampling points to that point are summed to obtain the final vehicle risk field intensity value for that spatial grid point. The same summation calculation is performed on all spatial grid points, and the vehicle risk field intensity values ​​of all spatial grid points together constitute a continuous spatial distribution of risk intensity; this distribution is the vehicle risk field. The vehicle risk field is stored and represented in the form of a spatial grid, where the value of each grid cell represents the comprehensive intensity of the risk caused by the target vehicle appearing at that location in the future.

[0095] In one embodiment of the invention, the binary conflict calculation module uses a product summation method to calculate the conflict intensity caused by the overlap of the risk fields of the self-vehicle and surrounding vehicles. The product summation method presupposes mapping the self-vehicle's risk field and the surrounding vehicles' risk fields onto the same high-resolution spatial discrete grid. The spatial discrete grid covers the entire area required for evaluation and has a unified spatial coordinate origin, coordinate axis direction, and grid cell size. Both the self-vehicle's and the surrounding vehicles' risk fields are sampled or interpolated onto this unified spatial discrete grid, so that each grid cell stores the self-vehicle's risk field intensity value and the surrounding vehicles' risk field intensity values. After mapping, the conflict intensity is calculated by traversing each grid cell of the spatial discrete grid. For any given grid cell, the intensity value stored in the self-vehicle's risk field and the intensity value stored in the surrounding vehicles' risk fields are read. These two intensity values ​​represent the probabilistic intensity of the self-vehicle and surrounding vehicles occupying the grid cell and generating risk in the future. The lateral slip characteristics of vehicles on icy and snowy roads lead to greater uncertainty in the boundaries of the actual space occupied by the vehicle. High-resolution spatial discrete meshes can accurately capture such spatial occupancy deviations caused by icy and snowy roads, making the risk field spatial matching between the vehicle and surrounding vehicles more accurate and adapting to the needs of conflict intensity calculation on icy and snowy roads.

[0096] It can be understood that the calculation of the grid cell conflict intensity contribution value is based on the product of two intensity values. The read risk field intensity value of the vehicle is multiplied by the risk field intensity value of the surrounding vehicles, and the result is taken as the contribution value of the grid cell to the conflict intensity between the vehicle and the surrounding vehicles. This product operation implies the probabilistic meaning of the simultaneous occurrence of risks. It is then determined whether the calculated grid cell conflict intensity contribution value is greater than zero. If the grid cell conflict intensity contribution value is greater than zero, it means that the vehicle risk fields of the vehicle and the surrounding vehicles in that grid cell have non-zero intensities at the same time, that is, the future risk distributions of the two vehicles overlap at that spatial location. The grid cell is marked as a risk field overlap region. For all grid cells marked as risk field overlap regions, the calculated grid cell conflict intensity contribution values ​​are summed. The scalar result obtained by summing is the binary conflict field scalar value between the vehicle and the surrounding vehicle. This binary conflict field scalar value quantitatively describes the overall conflict risk level between the vehicle and a single surrounding vehicle in terms of spatial occupancy; the larger the value, the more severe the potential spatial conflict. Table 1 shows the risk field intensity values ​​of the vehicle itself, the risk field intensity values ​​of the surrounding vehicles, and the calculated contribution values ​​of the grid cell conflict intensity for three grid cells marked as overlapping risk fields on a simplified spatial discrete grid.

[0097] Table 1: Examples of Calculating Binary Collision Fields Using the Product Summation Method

[0098]

[0099] Note: The scalar value of the binary conflict field = 0.48 + 0.63 + 0.20 = 1.31.

[0100] In some embodiments, the binary conflict calculation module uses the Monte Carlo integration method to calculate the conflict intensity caused by the overlap of the risk fields of the self-vehicle and surrounding vehicles. The Monte Carlo integration method is used to approximate the integral of two risk fields over a continuously overlapping region. The Monte Carlo integration method first needs to determine the spatially overlapping region between the self-vehicle's risk field and the surrounding vehicles' risk fields; this overlapping spatial region is a continuous spatial region. Within the overlapping spatial region, a specified number of sampling points are randomly generated according to a uniform distribution. The specified number of values ​​is determined based on accuracy requirements and computational resources, and the spatial coordinates of the sampling points follow a uniform distribution within the overlapping spatial region. For each randomly generated sampling point, the intensity value of that point in both risk fields needs to be evaluated. Since vehicle risk fields are typically stored in discrete grid form, while the coordinates of the sampling points are continuous values, an interpolation method is needed to obtain the intensity value. A bilinear interpolation method is used to calculate the intensity value of the sampling point in the self-vehicle's risk field and the intensity value of the sampling point in the surrounding vehicles' risk fields, respectively. Bilinear interpolation uses a weighted average of the stored values ​​of the four nearest grid cells surrounding the sampling point.

[0101] Optionally, the contribution of each sampling point to the conflict intensity is calculated using the same principle as the product summation method on the grid cell. The calculated intensity value of the sampling point in its own vehicle's risk field is multiplied by the intensity value of the sampling point in the surrounding vehicles' risk fields; the result of this multiplication is taken as the contribution of that sampling point to the conflict intensity. After calculating the contribution for all specified number of sampling points, the average contribution of all random sampling points is calculated. The Monte Carlo integration method is used for the scalar values ​​of the binary conflict field. An approximate estimate is calculated using the following relationship:

[0102]

[0103] in: Let represent the binary conflict field scalar value to be determined between the vehicle and surrounding vehicles. This represents the area of ​​a continuous spatial region where the vehicle's risk field overlaps with the risk fields of surrounding vehicles. This represents the total number of randomly generated sampling points within the spatial region where the risk fields overlap. Indicates the first The spatial coordinate vector of a random sampling point Represents the first value obtained through bilinear interpolation. The intensity value of each sampling point in the vehicle's risk field. Represents the first value obtained through bilinear interpolation. The intensity value of each sampling point in the vehicle risk field of surrounding vehicles. Indicates all The contribution values ​​of each sampling point are summed. The resulting approximate estimate is the scalar value of the binary conflict field. The Monte Carlo integration method is particularly suitable for situations where the spatial regions of overlapping risk fields have irregular shapes or the risk fields themselves have complex analytical forms. It provides a flexible and easy-to-implement integration approximation scheme through random sampling.

[0104] See Figure 4 This is a heatmap showing the spatial distribution of the risk field for autonomous vehicles. The distribution pattern, with high risk at the center and a decreasing gradient outwards, visually verifies whether the risk field model aligns with physical intuition and the actual risk patterns on icy and snowy roads. The color bars on the right represent "risk / conflict intensity," gradually changing from blue to red, with the highest risk intensity in the central area, consistent with the core risk concentration area during lane changes. Based on the probability distribution of the predicted future trajectory of the autonomous vehicle, the risk field constructed using the vehicle's equivalent mass, human occupancy model, and time-discounted weights visually demonstrates the potential risk intensity of the vehicle at various locations in the future space. During system development, comparing the risk field heatmaps under different operating conditions allows for rapid identification of model defects, guiding parameter optimization and algorithm iteration.

[0105] In one embodiment of the present invention, the comprehensive risk assessment module generates a real-time total conflict field for the interaction scenario between the vehicle and multiple surrounding vehicles. First, it needs to calculate an independent binary conflict field scalar value for each surrounding vehicle within the vehicle's current perception range. The calculation process uses the vehicle's own vehicle risk field as the reference field, and for each surrounding vehicle within the perception range, its own vehicle risk field is used as the comparison field. For each surrounding vehicle, the binary conflict calculation module is invoked, and the vehicle risk field of the vehicle and that surrounding vehicle are input. The binary conflict calculation module can internally use a product summation method or a Monte Carlo integration method for calculation. The result is a single scalar value, which is the binary conflict field scalar value between the vehicle and that specific surrounding vehicle. For example, in a typical lane-changing scenario involving three surrounding vehicles, assume that the three surrounding vehicles are located to the left rear, directly behind, and right rear of the vehicle, respectively. The system will sequentially calculate the binary conflict field scalar values ​​between the vehicle and the vehicle to its left rear, between the vehicle and the vehicle directly behind, and between the vehicle and the vehicle to its right rear. Each calculation is independent and complete.

[0106] It is understandable that all calculated binary conflict field scalar values ​​need to be systematically stored for subsequent comprehensive processing. The system maintains a vehicle identifier for each participating surrounding vehicle, which can be a unique tracking identifier based on sensor data. After each binary conflict field scalar value is calculated, the system stores this scalar value and its corresponding vehicle identifier in a dynamically updated data structure, such as a mapping table with the vehicle identifier as the key and the binary conflict field scalar value as the value. This storage method allows the system to track the independent contribution of each surrounding vehicle to the comprehensive risk and facilitates updates or removals at subsequent time steps.

[0107] In some embodiments, the core operation for generating the current real-time total conflict field scalar value is to perform arithmetic summation on all stored binary conflict field scalar values. The system reads all recorded binary conflict field scalar values ​​from the storage structure at the current moment. These scalar values ​​represent the independent conflict intensity between the vehicle and each valid surrounding vehicle in the scene. All these binary conflict field scalar values ​​are summed using an addition operation, and the result is the current real-time total conflict field scalar value. Assuming there are three binary conflict field scalar values ​​in the storage structure, namely 2.15, 3.80, and 0.65, then the real-time total conflict field scalar value is 2.15 + 3.80 + 0.65 = 6.60. This real scalar value represents the overall global conflict risk level faced by the vehicle at a specific moment.

[0108] As vehicle movement and sensor data are updated, the system repeatedly executes the aforementioned calculation and accumulation process periodically or in an event-triggered manner. In a new calculation cycle, the system regenerates the vehicle risk field for the vehicle and all surrounding vehicles based on the latest environmental perception information, recalculates the binary conflict field scalar value between the vehicle and each updated surrounding vehicle, updates the storage structure, and performs the arithmetic accumulation operation again to generate a new real-time total conflict field scalar value. This continuous execution of the process generates a time-varying numerical sequence, which constitutes a quantitative expression of the real-time total conflict field in the time dimension, reflecting the dynamic evolution of the vehicle's overall risk during lane changing.

[0109] Optionally, the dynamic risk update module is responsible for coordinating the refresh cycle of the entire risk state at preset fixed time intervals. The system sets a fixed update cycle, which determines the time resolution of the risk assessment, for example, 100 milliseconds. At the beginning of each update cycle, the dynamic risk update module first synchronously acquires the latest sensor data of the vehicle and all surrounding vehicles. Based on the latest sensor data, the system updates the vehicle motion features and vehicle interaction features. Using the updated feature data, the system refreshes the time-series feature input matrix. The environmental attention network uses the refreshed time-series feature input matrix to re-execute trajectory prediction, generating a new position probability density distribution for the future moment that reflects the latest situation. Based on the newly generated position probability density distribution, the system re-executes all subsequent steps completely: Markov chain Monte Carlo sampling to obtain a new set of future trajectory sampling points; constructing a new vehicle risk field for the vehicle and each surrounding vehicle; calculating the binary conflict field scalar value of the vehicle and each surrounding vehicle; and superimposing all the binary conflict field scalar values ​​to generate a new real-time total conflict field scalar value. The output of the dynamic risk update module is the real-time total conflict field scalar value calculated in the current update cycle, as well as the binary conflict field scalar values ​​between the vehicle and each key surrounding vehicle. The entire system achieves dynamic updating of the real-time total conflict field through this periodic, end-to-end recalculation. Its update mechanism can be described as follows:

[0110]

[0111] in: Indicates the first The real-time total collision field scalar value output in each update cycle. Indicates the first The total number of surrounding vehicles within the vehicle's perception range is updated every cycle. Indicates the first The scalar values ​​of the binary conflict field of all surrounding vehicles within each period are summed. This represents a function for calculating binary conflict, which internally implements either product summation or Monte Carlo integration. Indicates the first Each update cycle calculates the vehicle risk field based on the latest data. Indicates the first The calculation of the first update cycle yields the [number]th [number]. The vehicle risk field surrounding the vehicle.

[0112] See Figure 5This is a line graph showing the evolution of the binary conflict intensity between the vehicle and multiple vehicles behind it over time. It visually demonstrates that during lane changes, the vehicle directly behind is the primary source of risk, followed by the vehicle to the left, while the vehicle to the right carries the lowest risk. This provides a clear priority basis for system decision-making and driver warnings. The conflict intensity of the vehicle to the left peaks at approximately 2.5 seconds (approximately 2.1), then rapidly decreases. The conflict intensity of the vehicle directly behind peaks at approximately 2.5 seconds (approximately 3.8), representing the highest risk among the three, and its decrease is relatively slow. The conflict intensity of the vehicle to the right peaks at approximately 2 seconds (approximately 1.2), exhibiting the lowest overall risk level and the fastest decrease. The conflict intensity of all vehicles follows a pattern of "rapid increase, reaching a peak, and then gradually decreasing to near zero," consistent with the risk evolution logic of "approach-convergence-distance" during lane changes.

[0113] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A dynamic assessment system for vehicle lane-changing risk on icy and snowy roads based on trajectory prediction, characterized in that, The system includes: The prediction sampling module obtains the probability density distribution of the target vehicle's position at future times, and extracts a series of future trajectory sampling points from the probability density distribution of the position at future times using the Markov chain Monte Carlo method. The risk field construction module calculates the risk field intensity value generated in space for each of the future trajectory sampling points based on the vehicle's equivalent mass, human occupation model, and future time discounting weight function, and aggregates them to form a vehicle risk field that characterizes the intensity of the risk that the vehicle will encounter at various points in space in the future. The binary conflict calculation module, when assessing the risk interaction between the vehicle and surrounding vehicles, spatially matches the vehicle risk field of the vehicle with the vehicle risk fields of the surrounding vehicles, calculates the sum of the products of the vehicle risk field intensity value of the vehicle and the vehicle risk field intensity value of the surrounding vehicles in the overlapping spatial region, and obtains the binary conflict field characterizing the conflict intensity between the vehicle and a single surrounding vehicle. The comprehensive risk assessment module calculates the binary conflict field between the vehicle and each of the surrounding vehicles in the scenario of interaction between the vehicle and multiple surrounding vehicles, and superimposes all the binary conflict fields to generate a real-time total conflict field that reflects the comprehensive risks faced by the vehicle during the entire lane-changing process. The sum of the products of the vehicle risk field intensity value of the self-vehicle and the vehicle risk field intensity values ​​of surrounding vehicles within the overlapping spatial region yields a binary conflict field characterizing the conflict intensity between the self-vehicle and a single surrounding vehicle, including: The vehicle risk field of the vehicle itself and the vehicle risk fields of the surrounding vehicles are mapped onto the same high-resolution spatial discrete grid, so that the two vehicle risk fields have the same spatial coordinate definition and grid cell division. Traverse each grid cell of the spatial discrete grid, read the intensity value of the grid cell in the vehicle risk field of the vehicle itself, and read the intensity value of the grid cell in the vehicle risk field of the surrounding vehicles. The risk field intensity value of the self-vehicle is multiplied by the risk field intensity values ​​of the surrounding vehicles to obtain the conflict intensity contribution value on the grid cell. Determine whether the conflict intensity contribution value on the grid cell is greater than zero. If it is greater than zero, mark the grid cell as a risk field overlap region. The conflict intensity contribution values ​​of all grid cells marked as overlapping areas of the risk field are summed to obtain the scalar value of the binary conflict field, which quantitatively describes the overall conflict risk between the vehicle and surrounding vehicles in terms of space occupation.

2. The dynamic assessment system for vehicle lane-changing risk on icy and snowy roads based on trajectory prediction according to claim 1, characterized in that, The acquisition of the probability density distribution of the target vehicle's future location includes: The vehicle motion features and vehicle interaction features are obtained from continuous historical moments of the vehicle and surrounding vehicles. The vehicle motion features include instantaneous velocity values, instantaneous acceleration values, planar position coordinates, and vehicle sideslip angle. The vehicle interaction features include the relative velocity and relative position of the vehicle and each surrounding vehicle. The vehicle motion features and vehicle interaction features at the continuous historical moments are time-aligned and dimension-stitched to form a time series feature input matrix containing multi-dimensional features of multiple vehicles at continuous time steps. The time-series feature input matrix is ​​processed by an environmental attention network to extract the temporal dependencies, spatial dependencies, and interactive attention relationships between vehicles. Based on these temporal dependencies, spatial dependencies, and interactive attention relationships, the probability density distribution of the target vehicle's future position is predicted.

3. The dynamic assessment system for vehicle lane-changing risk on icy and snowy roads based on trajectory prediction according to claim 2, characterized in that, The step of processing the time-series feature input matrix through an environmental attention network to extract temporal dependencies, spatial dependencies, and interactive attention relationships between vehicles, and predicting the probability density distribution of the target vehicle's future position based on these dependencies, includes: The time series feature input matrix is ​​processed using a long short-term memory encoder. The long short-term memory encoder encodes the historical trajectory sequence of each vehicle, extracts and outputs the encoded feature vector of each vehicle at each historical moment, and the encoded feature vector contains the temporal dependency relationship of the vehicle. The encoded feature vectors of all vehicles at the same time are used as nodes to construct a graph structure based on the spatial distance relationship between vehicles. The graph structure is then input into a graph attention network. The graph attention network adaptively calculates the attention weights between nodes to generate an interaction attention weight matrix that reflects the relative importance between vehicles and outputs an interaction feature vector that contains the interaction attention relationship between vehicles. Based on the interaction feature vector, a three-dimensional social tensor centered on the target vehicle is constructed. The two planar dimensions of the three-dimensional social tensor represent a spatial grid, and the channel dimension is filled with the interaction feature vectors of other vehicles in the corresponding spatial grid. The three-dimensional social tensor is input into a convolutional social pooling network with a squeeze excitation structure. The convolutional social pooling network with a squeeze excitation structure extracts the local spatial patterns around the target vehicle through convolution operations and uses a channel attention mechanism to weight different spatial feature channels, outputting spatial context features that contain the spatial dependency relationship between the target vehicle and the surrounding environment. The final hidden state corresponding to the temporal dependency, the interaction feature vector corresponding to the interaction attention relationship, and the spatial context feature corresponding to the spatial dependency are fused and input into the combined structure of the long short-term memory decoder and the hybrid density network. The long short-term memory decoder is responsible for decoding the future time sequence, and the hybrid density network processes the output of the decoder at each future time to estimate the two-dimensional Gaussian mixture distribution parameters of the target vehicle's planar position at the future time. The two-dimensional Gaussian mixture distributions of all future times together constitute the position probability density distribution of the future time.

4. The dynamic assessment system for vehicle lane-changing risk on icy and snowy roads based on trajectory prediction according to claim 3, characterized in that, From the probability density distribution of the future positions, a series of future trajectory sampling points are extracted using the Markov chain Monte Carlo method, including: Initialize a Markov chain whose initial state is a sequence of future trajectory points randomly drawn from the position probability density distribution of the future time; Based on the parameters of the two-dimensional Gaussian mixture distribution, a proposal distribution is defined to generate a new sequence of candidate future trajectory points from the current Markov chain state; Calculate the joint probability density of the candidate future trajectory point sequence under the position probability density distribution at the current future time, and compare it with the joint probability density of the current state. Based on a preset acceptance criterion, decide whether to accept the candidate future trajectory point sequence as a new state of the Markov chain. Repeat the steps from defining the proposed distribution to deciding whether to accept the candidate state until the state transition of the Markov chain reaches a preset number of samples, which ensures that the chain reaches a stable distribution; From the Markov chain that has reached a stationary distribution, states are extracted at preset intervals. Each extracted state is used as a complete sampling point of the future trajectory, and finally a set of samples is obtained to approximate the position probability density distribution of the future time.

5. The dynamic assessment system for vehicle lane-changing risk on icy and snowy roads based on trajectory prediction according to claim 4, characterized in that, The method of calculating the spatial risk field intensity value for each future trajectory sampling point based on the vehicle's equivalent mass, human occupancy model, and future time discounting weight function includes: The equivalent mass of the target vehicle is obtained by looking up the corresponding table based on its physical parameters. The equivalent mass of the vehicle combines the actual mass and dynamic characteristics of the vehicle. Based on the artificial occupancy model, the vehicle planar position represented by the future trajectory sampling point is expanded into a polygonal spatial region occupied by the vehicle in the planar position. The artificial occupancy model defines the safety boundary of the vehicle outline expansion. A future time discount weight is defined for each predicted future moment. The future time discount weight decreases monotonically as the predicted moment is delayed, and the more distant the future moment is, the smaller the weight of its impact on the current risk. For each spatial grid point within the polygonal spatial region, the equivalent mass of the vehicle, the future time discount weight, and the probability value of the future trajectory sampling point itself are multiplied to obtain the basic value of the risk field intensity contributed by the future trajectory sampling point to the spatial grid point. Traverse all sampling points in the future trajectory sampling point sample set. For the same spatial grid point, accumulate the basic value of the risk field intensity contributed by all future trajectory sampling points to it to obtain the final risk field intensity value of the spatial grid point. The risk field intensity values ​​of all spatial grid points together constitute a continuous risk intensity spatial distribution, that is, the vehicle risk field.

6. The dynamic assessment system for vehicle lane-changing risk on icy and snowy roads based on trajectory prediction according to claim 5, characterized in that, The step of spatially matching the vehicle risk field of the self-vehicle with the vehicle risk fields of surrounding vehicles, and calculating the sum of the products of the vehicle risk field intensity values ​​of the self-vehicle and the vehicle risk field intensity values ​​of surrounding vehicles in the overlapping spatial region to obtain a binary conflict field characterizing the conflict intensity between the self-vehicle and a single surrounding vehicle, further includes the step of approximating continuous spatial integration through Monte Carlo integration: Within the spatial region where the risk fields of the vehicle and surrounding vehicles overlap, a specified number of sampling points are randomly generated, and the spatial coordinates of the sampling points follow a uniform distribution within the spatial region. For each randomly generated sampling point, the intensity value of the sampling point in the vehicle risk field of its own vehicle and the intensity value of the sampling point in the vehicle risk field of the surrounding vehicles are calculated by bilinear interpolation. Multiply the calculated risk field intensity value of the vehicle by the risk field intensity values ​​of the surrounding vehicles to obtain the contribution of the sampling point to the conflict intensity. Calculate the average contribution of all random sampling points, and multiply the average by the area of ​​the overlapping spatial region of the risk field to obtain an approximate estimate of the continuous spatial integral. The approximate estimate is the scalar value of the binary conflict field.

7. The dynamic assessment system for vehicle lane-changing risk on icy and snowy roads based on trajectory prediction according to claim 6, characterized in that, The process of calculating the binary conflict fields between the vehicle and each surrounding vehicle, and then superimposing all the binary conflict fields to generate a real-time total conflict field reflecting the comprehensive risks faced by the vehicle throughout the lane-changing process, includes: For each surrounding vehicle within the current perception range of the vehicle, a binary conflict field scalar value is calculated using the product summation method or the Monte Carlo integration method based on the vehicle risk field of the vehicle and the vehicle risk fields of the surrounding vehicles. The calculated binary conflict field scalar values ​​of the vehicle and each surrounding vehicle are stored according to the vehicle identifier. Perform an arithmetic summation operation on all stored binary collision field scalar values ​​at the current moment, and the sum is the real-time total collision field scalar value at the current moment; As time progresses and the scenario is updated, the calculation and accumulation operations are repeatedly performed to generate a real-time total conflict field scalar value sequence that changes over time. This real-time total conflict field scalar value sequence constitutes the real-time total conflict field that reflects the dynamic evolution of risk.

8. The dynamic assessment system for vehicle lane-changing risk on icy and snowy roads based on trajectory prediction according to claim 7, characterized in that, Also includes: The dynamic risk update module repeatedly generates a real-time total conflict field at preset fixed time intervals, dynamically updating and outputting the risk status of continuously changing lane-changing scenarios. Specifically, this includes: Set a system update cycle, and at the beginning of each update cycle, synchronously acquire the latest sensor data of the vehicle and all surrounding vehicles; The vehicle motion features and vehicle interaction features are updated based on the latest sensor data, and the time series feature input matrix is ​​refreshed accordingly. Using the refreshed time-series feature input matrix, the environmental attention network is driven to re-execute trajectory prediction, generating an updated position probability density distribution for future moments. Based on the updated location probability density distribution for future moments, the entire process of Markov chain Monte Carlo sampling, vehicle risk field construction, binary conflict field calculation, and real-time total conflict field generation is re-executed. Output the real-time total conflict field scalar value calculated in the current update cycle, as well as the binary conflict field scalar value between the vehicle and each key surrounding vehicle.

9. The dynamic assessment system for vehicle lane-changing risk on icy and snowy roads based on trajectory prediction according to claim 8, characterized in that, The instantaneous acceleration values ​​in the vehicle motion characteristics are longitudinal and lateral accelerations corrected for the adhesion coefficient of icy and snowy road surfaces. The correction process includes: The estimated ice and snow adhesion coefficient of the current road surface is obtained through vehicle-mounted sensors or road condition recognition modules; Read the raw longitudinal acceleration request and raw lateral acceleration request output by the vehicle dynamics controller; Multiply the original longitudinal acceleration request by the estimated ice and snow adhesion coefficient to obtain the upper limit of safe longitudinal acceleration under icy and snowy road surface. If the original request exceeds the upper limit of safe longitudinal acceleration, the upper limit of safe longitudinal acceleration is used as the instantaneous value of longitudinal acceleration actually used. Multiply the original lateral acceleration request by the estimated ice and snow adhesion coefficient to obtain the upper limit of safe lateral acceleration under icy and snowy road surface. If the original request exceeds the upper limit of safe lateral acceleration, the upper limit of safe lateral acceleration is used as the instantaneous value of lateral acceleration actually used. The safe longitudinal acceleration upper limit value and the safe lateral acceleration upper limit value, after being processed by the upper limit constraint, are used as the final instantaneous acceleration values ​​and input into the time series feature input matrix.

Images