A vehicle-road cooperative decision and vehicle motion planning method based on predictive risk assessment

Abstract

A kind of vehicle-road cooperative decision and vehicle motion planning method based on predictive risk assessment, first, obtain multi-source traffic information, complete target vehicle multi-modal trajectory prediction and driving intention recognition by fusing scene semantics;Then, based on the dynamic adjustment of scene semantic risk field parameters, the counterfactual risk field of behavior dependence is constructed by counterfactual reasoning, and the multi-vehicle interaction risk field is constructed by combining lane relationship weighting;Then, the local risk field of multiple vehicles is fused through vehicle-road cloud cooperative communication to generate a collaborative risk atlas;With the atlas as the core, risk assessment is carried out on the candidate driving behavior, and the integration of vehicle behavior decision and motion control is realized by adopting double-layer model predictive control;Finally, through the simulation platform combined with dynamic channel modeling, the risk confidence fusion weight is optimized.The present application realizes the prospective evaluation of vehicle future risk, solves the problems of lack of space-time integrated modeling in existing risk assessment method, disconnection of decision and control, and lack of global collaborative risk cognition.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and intelligent connected vehicle technology, and in particular to a vehicle-road cooperative decision-making and vehicle motion planning method based on predictive risk assessment. It is applicable to vehicle behavior decision-making and motion control in multi-vehicle interaction scenarios in intelligent connected environments, such as urban intersections, highway merging, multi-vehicle following, and sudden interference scenarios. Background Technology

[0002] The development of intelligent connected vehicles (ICVs) and vehicle-road-cloud collaborative systems is driving the evolution of road traffic from single-vehicle intelligence to group collaboration. In dense, dynamic multi-vehicle interaction scenarios, the safe and efficient operation of vehicles depends on accurate prediction of the future behavioral intentions of surrounding vehicles and precise, forward-looking quantitative assessment of collision risks. Current mainstream risk assessment and motion planning methods have many limitations, specifically:

[0003] Most methods assess instantaneous risk based on the geometric relationship at the current moment, lacking continuous and spatiotemporal integrated modeling of the uncertainty of the vehicle's future trajectory, making it difficult to characterize the dynamic evolution of the future risk field caused by the vehicle's driving intention;

[0004] The design is often hierarchical, with the risk assessment module output serving only as a heuristic input / soft constraint for the decision-making or planning layer. It does not form a unified risk measurement and driving mechanism that runs through the entire chain of perception, decision-making, and control, which can easily lead to decision oscillations or control lag.

[0005] Multimodal trajectory prediction models are decoupled from traffic scene semantics and counterfactual responses of other vehicles to the behavior of the vehicle itself, resulting in poor interpretability of prediction results and limiting the safety and rationality of critical decisions.

[0006] Optimizing trajectories primarily from the perspective of individual vehicles lacks risk information sharing and integration through vehicle-road-cloud collaboration, hindering the formation of collective consensus and making it difficult to achieve globally safe or optimal collaborative behavior.

[0007] Simulation verification of collaborative decision-making methods treats communication as an ideal link, ignoring the latency and packet loss issues of real channels. This can lead to risks such as information asynchrony and decision failure when the algorithm is actually deployed, and lacks a systematic verification method for the communication-control collaborative closed loop.

[0008] Therefore, there is an urgent need for a vehicle-road cooperative decision-making and motion planning method that can deeply integrate scene semantics and multi-agent interaction prediction, construct a forward-looking and computable collaborative risk map, and realize integrated decision-making and control collaboration, so as to solve the above-mentioned problems of existing technologies. Summary of the Invention

[0009] The purpose of this invention is to provide a vehicle-road cooperative decision-making and vehicle motion planning method based on predictive risk assessment, which enables forward-looking assessment of future vehicle risks in complex traffic scenarios, improves the driving safety, stability and cooperative decision-making capabilities of intelligent connected vehicles, and has good engineering feasibility and system scalability.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A vehicle-road cooperative decision-making and vehicle motion planning method based on predictive risk assessment, characterized by the following steps:

[0012] (1) Obtain historical driving status information, road topology information, traffic rule information and traffic scene semantic information of the target vehicle and its surrounding vehicles, construct scene semantic encoding vector, and use it together with historical status information and road topology information as input to subsequent modules;

[0013] (2) Based on the historical driving status information and scene semantic information, a multimodal vehicle trajectory prediction model that integrates scene semantics is constructed to predict the candidate driving trajectory of the target vehicle in multiple future time steps and to determine the driving intention corresponding to the candidate driving trajectory.

[0014] (3) Based on the scene semantic encoding vector, the key parameters of the risk field are dynamically adjusted through a learnable parameter adjustment mapping function to generate dynamic risk field adjustment parameters, so that the risk assessment has scene adaptive capability.

[0015] (4) For each candidate driving behavior, construct a counterfactual response model of surrounding vehicles based on counterfactual reasoning, predict the response trajectory of surrounding vehicles, and then construct a counterfactual risk field dependent on the behavior, and calculate the risk change of the behavior compared to the current behavior.

[0016] (5) Based on the candidate driving trajectory and the interaction relationship between vehicles, construct the multi-vehicle interaction risk field of the target vehicle in the prediction time domain, weight the risk contribution of different interactive vehicles according to the lane relationship of the vehicle, and integrate the counterfactual risk field and scene semantic parameters.

[0017] (6) Through vehicle-road-cloud collaborative communication, the local risk fields of multiple vehicles are extracted and compressed and encoded before transmission. The local risk fields of multiple vehicles are integrated at roadside units or cloud nodes to construct a collaborative risk map, and structured features are extracted from the map.

[0018] (7) Based on the collaborative risk map, risk assessment is performed on the candidate driving behaviors of the target vehicle, a comprehensive behavior cost function is constructed, and a two-layer model predictive control is used to generate vehicle control commands, thereby realizing the integrated collaboration of vehicle behavior decision-making and motion control. In the interactive scenario, group collaborative decision-making is achieved through the V2X negotiation mechanism.

[0019] (8) Based on the system-level simulation platform, the communication reliability of the transmission and fusion process of the collaborative risk map is simulated and verified. The real-time performance and reliability of risk information transmission are evaluated through dynamic channel modeling, and the risk confidence fusion weight is optimized.

[0020] Compared with the prior art, the technical solution of the present invention has the following advantages:

[0021] Scene-adaptive risk assessment: By introducing scene semantic perception and dynamic parameter adjustment, high-level traffic semantics are transformed into computable risk field parameters, enabling risk assessment to have scene-adaptive capabilities. This solves the problem of inaccurate risk quantification in fixed parameter models under different scenarios and improves the accuracy of risk modeling in complex traffic environments.

[0022] Proactive causal risk decision-making: By using a counterfactual risk reasoning mechanism, a hypothesis-response model is constructed for each candidate driving behavior, which proactively assesses the impact of different behaviors on surrounding vehicles and risk changes, providing a causal explanation for behavioral decisions and significantly enhancing the forward-looking nature and safety of decision-making;

[0023] Global collaborative risk perception: Breaking through the limitations of single-vehicle risk perception, a collaborative risk map construction mechanism based on V2X communication is proposed. This mechanism integrates local risk information from multiple vehicles to generate a globally consistent collaborative risk map, providing a shared risk perception foundation for multiple vehicles, enabling collaborative decision-making among groups, and improving the overall safety and traffic efficiency of the transportation system.

[0024] Integrated decision-making and control collaboration: The collaborative risk map is used as the core driver of the two-layer model predictive control. It serves as both the evaluation basis for upper-level behavioral decisions and the cost function and hard constraint for lower-level motion control, forming a unified risk measurement closed loop that runs through perception, decision-making, and control. This ensures the continuity from decision intention to control execution and improves the safety, smoothness, and stability of vehicle motion planning.

[0025] Excellent system scalability: The framework of this invention is compatible with different types of sensors, communication protocols and vehicle platforms, and can be directly deployed in large-scale intelligent connected transportation systems, demonstrating good engineering feasibility;

[0026] Closed-loop verification of communication and control: By integrating a V2X communication simulation platform, the channel transmission performance is quantified into risk confidence level, realizing closed-loop verification of communication quality and risk integration, improving the robustness of the system in weak signal and high interference environments, and providing a complete toolchain for performance verification before system deployment. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the specific process of the present invention.

[0028] Figure 2 This is a detailed framework diagram of the multimodal prediction of the present invention.

[0029] Figure 3 This is a schematic diagram of the risk field modeling process of the present invention.

[0030] Figure 4 This is a schematic diagram of the simulation platform system of the present invention. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0032] This embodiment takes a multi-vehicle interaction scenario at an urban intersection as an example to provide a detailed description of the vehicle-road cooperative decision-making and vehicle motion planning method based on predictive risk assessment of the present invention. The scenario includes target vehicles, straight-going social vehicles, and left-turning non-motorized vehicles, and has the semantic features of low visibility in rainy weather.

[0033] like Figure 1 As shown, a vehicle-road cooperative decision-making and vehicle motion planning method based on predictive risk assessment includes the following steps:

[0034] Step S1: Multi-source environmental perception and scene semantic information acquisition

[0035] The system acquires historical driving status information of the target vehicle and surrounding vehicles, including the position, speed, and relative positions of each vehicle at multiple historical moments. Simultaneously, it acquires road topology information, traffic rule information, and traffic scene semantic information to characterize lane structure, road boundaries, lane marking constraints, and semantic labels for traffic events, weather conditions, and special areas. Specifically, this includes:

[0036] 1. Obtain semantic tags such as traffic event information (e.g., accidents, construction, congestion), weather conditions (rain, snow, fog), special areas (schools, hospitals, intersections), and traffic participant types (pedestrians, bicycles, large vehicles) through vehicle sensors and communication;

[0037] 2. Construct scene semantic encoding vectors Each dimension corresponds to the strength or confidence level of a specific semantic feature;

[0038] 3. Use the semantic encoding vector, historical state information, and road topology information as inputs for subsequent modules.

[0039] Step S2: Multimodal trajectory prediction and driving intention recognition

[0040] like Figure 3 As shown, based on the historical driving state information and scene semantic information, a multimodal vehicle trajectory prediction model for environmental perception is constructed. The interaction relationship between vehicles is modeled using a graph structure, and the spatiotemporal features of vehicle motion are extracted through temporal modeling and attention mechanism to predict the driving trajectory of the target vehicle in multiple future time steps.

[0041] 1. Encoding Layer: The vehicle's trajectory is influenced by the interactions with surrounding vehicles and the traffic environment. The input to the encoding layer, based on the original trajectory and relative position information, adds local scene semantic encoding for each vehicle. Define the vehicle. At any moment The local scene semantic encoding vector is This vector is obtained from the global semantic information in step S1, based on the vehicle... The position is encoded to obtain the result. The input to the encoding layer is expanded as follows:

[0042]

[0043] in For the predicted vehicle at time The horizontal and vertical position coordinates, For the predicted vehicle at time The local scene semantic encoding vector.

[0044]

[0045] in For the first The lateral and longitudinal relative distances between the vehicle and the surrounding vehicles and the predicted vehicle. For the first The surrounding vehicles at all times Local scene semantic encoding vector.

[0046] After encoding through an LSTM network, the output is:

[0047]

[0048]

[0049] In the formula: and For a moment The hidden layer states contain historical trajectories and scene semantic features. This is the learnable weight parameter matrix for the LSTM encoding network.

[0050] 2. Graph Convolutional Layer: The graph convolutional layer is used to extract the interaction features between the predicted vehicle and surrounding vehicles, where the interaction occurs within a specific scene semantic context. At time [time value missing]... Below, the predicted vehicle and surrounding vehicles are treated as a graph structure. The vertices in the vector are encoded using LSTM to obtain the fused feature vector. and As vertex feature vectors, i.e. and A two-layer graph convolutional neural network is used to output the trajectory interaction features at time t, which include the trajectory's own features and contextual interaction features. :

[0051]

[0052] In the formula: For activation function, and These are the weights of the first and second layers of the graph convolutional network, respectively. It is an adjacency matrix.

[0053] 3. Attention Layer: An attention mechanism is added to the model to better understand and interpret vehicle interactions. Hidden layer vectors of surrounding vehicles encoded by LSTM Considered as key and value, the hidden layer vector of the vehicle being predicted Treating it as a query, the degree of interaction between surrounding vehicles and the predicted vehicle is explained by calculating the correlation between the key and the query. Because... and With the same dimensions, cosine similarity is selected as the function for calculating the attention score. The resulting attention scores are then subjected to softmax to output the interaction weights. The final attention vector is obtained by weighted summation of this vector and the value. The calculation expression is as follows:

[0054]

[0055]

[0056] 4. Intent Recognition Layer: The output of the prediction model includes the prediction of driving intent. Global scene semantic information influences the prior distribution of driving intent. Let the global scene semantic encoding be... The feature vector after convolution of the graph Attention vector and global scene semantic summarization The concatenated vectors, used as input for driving intention prediction, are then processed by softmax to obtain a scene-aware driving intention recognition vector. :

[0057]

[0058] in, These represent the scenario-dependent probabilities of changing lanes to the left, going straight, and changing lanes to the right, respectively. This is a summary function for global semantics. This is the weight matrix for the intent recognition layer. This is the bias vector for the intent recognition layer.

[0059] 5. Decoding layer: Finally, the graph convolutional features are... Attention vector Driving intention recognition vector and the final scene semantic encoding of the predicted vehicle The concatenated data is then fed into an LSTM decoding network to predict the probability distribution of future trajectories under different intentions.

[0060]

[0061] in, The distribution parameters of the predicted future trajectory sequence are output by the decoding layer. Scene semantic information participates in the decoding, which helps to generate reasonable trajectories that conform to scene constraints (such as curve curvature and intersection rules).

[0062] Step S3: Dynamic adjustment of risk field parameters for scene semantic awareness

[0063] Scene semantic encoding vector Dynamically adjust key parameters of the risk field to enable scenario-adaptive risk assessment. Define the parameter adjustment mapping function:

[0064]

[0065] in, , Learnable weight matrix , For bias terms, For risk field parameter dimensions.

[0066] Dynamic parameter set Includes: equivalent threshold for the same lane Cross-lane attenuation rate parameters Distance attenuation parameters , caliper potential field amplitude Direction correction parameters The above parameters will replace the fixed parameters in the original risk field formula, achieving scene adaptation, such as increasing the parameters in low visibility scenarios. Expand the scope of risk impact in the same lane; reduce risk in wet and slippery road surface scenarios. The risk decreases more slowly with distance; the school area is increasing. Strengthen lane marking constraints; adjust congestion scenarios This enhances the coupling of cross-lane risks. Step S4: Construction of a multi-behavioral risk field based on counterfactual reasoning.

[0067] For each candidate driving behavior Assuming the vehicle starts from the current moment... The behavior is executed, the response trajectories of surrounding vehicles are predicted, and a counterfactual risk field based on the behavior is constructed.

[0068] For candidate behavior Generate the vehicle's counterfactual trajectory For surrounding vehicles Construct a counterfactual response model:

[0069]

[0070] in, For the interactive feature extraction function, consider the vehicle's counterfactual trajectory for the vehicle. The direct impact.

[0071] Define surrounding vehicles Regarding the behavior of private vehicles Response sensitivity:

[0072]

[0073] in, This represents the dimension of the key vector in the attention mechanism, used to scale the dot product result to maintain numerical stability.

[0074] For each candidate action Construct a corresponding counterfactual risk field:

[0075]

[0076] in Based on vehicle Counterfactual prediction state calculation, using scenario dynamic parameters .

[0077] Define behavior Compared to current behavior Risk changes:

[0078]

[0079] Step S5: Construction of Multi-Vehicle Interaction Risk Field Based on Predicted Trajectory

[0080] Based on candidate driving trajectories and the relative positions of vehicles, a multi-vehicle interaction risk field for the target vehicle in the prediction time domain is constructed. This risk field comprehensively considers the motion state, relative distance, and relative orientation information of surrounding vehicles, and weights the risk contribution of different interacting vehicles according to their lane relationships, while also integrating counterfactual risk field and scene semantic parameters.

[0081] Single-vehicle risk field: To uniformly characterize the impact of surrounding vehicles on any position in space. The risk impact at the prediction time For each interactive vehicle Construct its risk potential field distribution The details are as follows:

[0082] 1. Risk items based on motion state: assuming interactive vehicles With assessment points The relative position vector is The relative velocity vector is , for and The angle between them. An exponential approach is used to characterize risk decay with higher forward risk and different horizontal and vertical distributions:

[0083]

[0084] in, This is the direction correction function. and These are the longitudinal and lateral risk attenuation scale parameters for scene adaptation, respectively, and their specific values ​​are calculated by the parameter dynamic adjustment mapping function in step S3.

[0085] 2. Distance-based risk component: Based on the physical characteristic that risk increases with closer distance, a risk contribution that decreases with increasing distance is defined:

[0086]

[0087] in, The preset safe distance threshold is used to characterize the radius of the absolutely high-risk area around the vehicle.

[0088] 3. Lane marking potential field: Different amplitude parameters are set for different types of markings (dashed lines, yellow lines, etc.) to characterize the binding force of road rules.

[0089]

[0090] 4. Road boundary potential field: Characterizes the hard constraint boundary risk of road edges on vehicle movement.

[0091]

[0092] 5. Lane Relationship Weighting Coefficient: For any interacting vehicles The lateral distance from its own lane centerline is defined as The lane width is Construct the weight function:

[0093]

[0094] In summary, the time was obtained The multi-vehicle interaction risk field that integrates counterfactual risk fields:

[0095]

[0096] in, To incorporate counterfactual risk weights, For behavior The predicted probability.

[0097] The multi-vehicle interaction risk field in the prediction time domain is obtained by superposition:

[0098]

[0099] in, This is the time discount factor.

[0100] Step S6: Collaborative Risk Map Construction and Dissemination

[0101] By leveraging vehicle-road-cloud collaborative communication, integrating local risk perception across multiple vehicles, constructing a collaborative risk map, and supporting collaborative decision-making among groups.

[0102] 1. Local risk contribution field generation: for each vehicle Based on its perceived scene semantics and counterfactual reasoning, calculate its contribution to local risk in the environment:

[0103]

[0104] 2. Risk Field Encoding and Transmission: To reduce the V2X communication load, feature extraction and compressed encoding are performed on the local risk field.

[0105]

[0106] broadcast via V2X and its confidence level .

[0107] 3. Collaborative Risk Map Fusion: At roadside units or cloud nodes, risk data uploaded by each vehicle is globally fused based on spatial attenuation weights.

[0108]

[0109] in For spatial decay weights, This is the decoding function.

[0110] 4. Feature Extraction: Structured features are extracted from the fused collaborative risk map to support subsequent group collaborative decision-making.

[0111]

[0112] Step S7: Risk-driven collaborative behavior decision-making and two-level model predictive control

[0113] Based on the collaborative risk map, the candidate driving behaviors of the target vehicle are assessed for risk, and a two-layer model is used to predict and control the vehicle to generate control commands.

[0114] 1. Construction of candidate behavior set: The candidate behavior set is provided by the multimodal prediction in step S2. .

[0115] 2. Risk cost calculation on the predicted trajectory: for any candidate trajectory (modality) Define its cumulative risk cost over the prediction time domain:

[0116]

[0117] 3. Behavioral cost function: For each candidate behavior Taking into account collaborative risks, changes in counterfactual risks, trajectory tracking errors, control smoothness, and prior intent, behavioral assessment indicators are constructed as follows:

[0118]

[0119] The items are: collaborative risk cost, counterfactual risk change cost, trajectory tracking error cost, control smoothness cost, and behavioral prior cost.

[0120] 4. Two-level model predictive control framework, such as Figure 2 As shown:

[0121] Upper-level MPC behavior decision-making: From the set of candidate behaviors, select the behavior that minimizes the overall cost function as the objective decision.

[0122]

[0123] Complete behavioral decision-making and behavior switching, and output reference trajectory.

[0124] Lower-level MPC: Performs longitudinal and lateral motion control of the vehicle based on behavioral decision results.

[0125] Predictive model motion control: A kinematic bicycle model is used as the predictive model, with state vectors... The continuous model is:

[0126]

[0127] Vertical MPC cost function:

[0128] Horizontal MPC cost function:

[0129] Risk-sensitive constraints: During the MPC optimization process, hard constraints based on scenario-adaptive risk thresholds are added to ensure driving safety.

[0130]

[0131] in, Risk thresholds are scenario-dependent.

[0132] 5. Collaborative Decision-Making Mechanism: In interactive scenarios such as lane changing, group collaboration is achieved through a V2X negotiation mechanism: vehicles broadcast their behavioral intentions and predicted trajectories, receive risk feedback and counterfactual assessments from surrounding vehicles, and choose the final behavior to maximize social welfare.

[0133]

[0134] in, For the benefit of the vehicle, For the benefit of his car, This is the synergy coefficient.

[0135] By incorporating the collaborative risk map and counterfactual risk assessment in the prediction time domain into the candidate behavior risk assessment and two-layer model predictive control optimization process, this invention can achieve integrated collaboration of behavior decision-making and motion control in complex traffic interaction scenarios, enabling vehicles to generate continuous, smooth, and lower-risk control commands while satisfying road and vehicle constraints, thereby improving driving safety, stability, and real-time adaptability.

[0136] Step S8: Collaborative risk assessment and communication simulation verification based on a system-level simulation platform

[0137] like Figure 4 As shown, to verify the effectiveness, stability, and real-time adaptability of the collaborative risk map constructed in this invention under complex traffic scenarios, a comprehensive simulation verification of the collaborative risk assessment and decision control process was conducted based on a system-level simulation platform. The simulation platform is used to uniformly evaluate the risk map generation, fusion, and its impact on vehicle decision control results; its specific process is shown in Table 1.

[0138] First, a typical traffic scenario matching the predictive risk assessment method is constructed through the simulation scenario definition module. This includes multi-vehicle interaction environments such as urban intersections, highway merging, multi-vehicle following, and sudden interference. Simultaneously, road topology, traffic rule constraints, the number of traffic participants, and their initial motion states are set. These scenario configuration parameters are passed to the simulation platform via interface 1 to initialize the runtime environment for the risk assessment and decision control algorithms.

[0139] Subsequently, during the simulation, the scene evolution parameters were dynamically adjusted via Interface 2 to simulate the impact of non-stationary factors such as changes in vehicle behavior, traffic density fluctuations, and changes in environmental constraints on the system, thereby testing the stability and adaptability of the collaborative risk map under complex dynamic conditions. The dynamically adjusted parameters include vehicle acceleration / deceleration disturbances, behavior switching frequency, prediction error magnitude, and changes in scene semantic weights.

[0140] During the collaborative risk map construction phase, each vehicle generates a local risk contribution field based on its own perception and prediction results. The simulation platform then uniformly calls Interface 3 to perform temporal alignment and spatial fusion of the local risk fields of multiple vehicles, resulting in a globally consistent collaborative risk map. Interface 3 outputs key evaluation indicators for the collaborative risk map, including risk field consistency indicators, spatial continuity indicators, and risk stability over time, which are used to measure the reliability and interpretability of the multi-vehicle risk fusion results.

[0141] During the decision-making and control verification phase, the simulation platform inputs the collaborative risk map into the risk-driven collaborative behavior decision-making and two-layer model predictive control module described in step S7 via interface 4, performing closed-loop simulation of the vehicle control command generation process. Interface 4 is used to statistically analyze the vehicle's decision-making results and control responses under different scenarios and parameter conditions, including indicators such as trajectory safety, risk exposure time, control smoothness, and behavior switching stability, thereby evaluating the actual constraint effect of the collaborative risk map on decision-making and motion control.

[0142] Through the above system-level simulation verification steps, the rationality of the construction of the collaborative risk map in multi-vehicle interaction scenarios, its dynamic stability, and its constraint effect on vehicle behavior decision-making and motion control can be comprehensively evaluated without relying on the communication model. This verifies the effectiveness and engineering feasibility of the method proposed in this invention in complex traffic environments.

[0143] Table 1. Examples of V2XDII library simulation processes

[0144] interface method Return parameters Interface 1 The initial configuration of the simulation scenario definition parameters (road topology, multi-vehicle interaction mode, traffic rules, scenario semantic labels) and risk assessment algorithm is synchronized to the system-level simulation platform. Scene initialization state, algorithm running parameters Interface 2 During the simulation, vehicle behavior disturbances, prediction errors, and scene semantic weights are dynamically adjusted to simulate complex environmental changes. Updated vehicle state sequence and dynamic scene parameters Interface 3 The local risk contribution fields of multiple vehicles are synchronized in time and fused spatially to generate a collaborative risk map; based on real-time vehicle-to-vehicle distances... Based on relative speed and dynamic channel state, calculate the local risk contribution field of inter-vehicle broadcasting. The transmission performance. The calculation process is as follows: formula (34)-(42). Risk information transmission success rate Total end-to-end latency of risk information Interface 4 Based on the collaborative risk map, we perform collaborative behavioral decision-making and two-level model predictive control simulation to calculate the collaborative risk map. Or the performance of issuing collaborative control commands. The calculation process is as follows: formula (43)-(50) Command / map transmission success rate Command / map issuance delay

[0145] Formulas (34)-(42) are used to calculate the local risk contribution field of inter-vehicle broadcasting. Communication quality at that time.

[0146] Path loss model:

[0147]

[0148] in, For the vehicle in step S5 and The relative distance between them.

[0149]

[0150] The probability of successful transmission of risk information packets. Given the length (in bits) of the data packet after risk field encoding, associate the risk field encoding and transmission in step S6:

[0151]

[0152] Channel contention access delay, simulating the delay caused by multiple vehicles simultaneously transmitting risk information in a dense scenario:

[0153]

[0154] Average timeslot duration, used to calculate access latency, reflects the efficiency of the MAC layer protocol:

[0155]

[0156] Data transmission latency, The channel capacity, based on Shannon's formula, determines the transmission speed of risk information:

[0157]

[0158] Total end-to-end latency of risk information The vehicle's local processing latency must be less than the patent decision control cycle; otherwise, the information becomes invalid.

[0159]

[0160] Risk transmission confidence level combines success rate and timeliness. The effective time threshold for risk information:

[0161]

[0162] Risk field corrected by the receiver, vehicles For vehicles Risk field information is weighted using link confidence to achieve communication-aware fusion:

[0163]

[0164] Formulas (43)-(50) are used to calculate the transmission collaboration risk map between the vehicle and the RSU / cloud. Or control the performance of commands.

[0165] V2R path loss model For antenna gain, Losses due to road obstruction:

[0166]

[0167] The formula for calculating the received signal-to-noise ratio (SNR) of a link, used to evaluate the downlink signal quality from the roadside unit (RSU) to the vehicle, is expressed as follows:

[0168]

[0169] in, This represents the channel gain from the RSU to the vehicle. This refers to the transmit power of the RSU. The path loss from RSU to vehicle. This represents the noise power spectral density.

[0170] Transmission success rate:

[0171]

[0172] in, The packet length (in bits) after encoding control instructions or global risk maps. This is the bit error rate function.

[0173] Access latency:

[0174]

[0175] in, The number of nodes (e.g., adjacent vehicles) competing for the channel with the RSU. The probability of sending RSU depends on the MAC layer protocol.

[0176] Average time slot duration:

[0177]

[0178] in, The probability of sending to at least one node. The conditional probability of successful RSU transmission. , , These correspond to the duration of idle, successful, and conflict time slots, respectively.

[0179] Transmission delay:

[0180]

[0181] in, The channel capacity of the V2R link (Shannon formula). To handle delays for RSU.

[0182] Total latency:

[0183]

[0184] Confidence level of decision instructions:

[0185]

[0186] in, This is the maximum allowable delay for control commands.

[0187] This invention proposes a scene semantic encoding vector and a local semantic pooling method to transform high-level traffic semantics (events, weather, region type) into computable feature vectors. These vectors are then deeply integrated into the entire encoding, graph convolution, and decoding process of a multimodal trajectory prediction model. This enables scene-aware driver intent recognition and trajectory generation, making the prediction results more consistent with the physical and rule constraints of traffic scenarios. By introducing scene semantic awareness and dynamic parameter adjustment, this invention enables risk assessment to have scene-adaptive capabilities, improving the accuracy of risk modeling in complex traffic environments.

[0188] This invention designs a parameter dynamic adjuster that uses a learnable mapping function to transform the global scene semantic encoding into core parameters of the risk field (such as the range of influence and the decay rate) in real time, enabling the risk potential field model to have scene adaptability and overcoming the problem of inaccurate risk quantification in fixed parameter models under different scenarios.

[0189] To address the uncertainty in interactive decision-making, this invention introduces a counterfactual reasoning mechanism. By constructing a hypothesis-response model for each candidate driving behavior, it predicts the possible reactions of other vehicles when the vehicle adopts that behavior, and calculates the corresponding counterfactual risk field. This provides a forward-looking causal risk comparison for behavior selection, significantly enhancing the forward-looking nature and safety of decision-making. Through the counterfactual risk reasoning mechanism, this invention can proactively assess the impact of different driving behaviors on surrounding vehicles and changes in risk, providing a causal explanation for behavioral decisions.

[0190] This invention overcomes the limitations of single-vehicle risk perception by proposing a collaborative risk graph construction and propagation mechanism based on V2X communication. Through encoding, broadcasting, and weighted fusion of local risk contributions from multiple vehicles, a globally consistent and consensus-based collaborative risk graph is generated, providing a shared risk perception foundation for multiple vehicles. By constructing the collaborative risk graph through vehicle-road-cloud collaboration, this invention achieves shared risk perception and collaborative decision-making among multiple vehicles, improving the overall safety and traffic efficiency of the transportation system.

[0191] This invention introduces collaborative risk mapping and counterfactual risk assessment into both behavioral decision-making and two-layer model predictive control processes, forming a unified risk-driven collaborative decision-making mechanism that improves the safety, smoothness, and stability of vehicle motion planning.

[0192] The framework proposed in this invention has good scalability and is compatible with different types of sensors, communication protocols and vehicle platforms, making it suitable for deployment and application in large-scale intelligent connected transportation systems.

[0193] This invention introduces V2X communication simulation and channel quality awareness mechanisms to enable the collaborative risk map to have adaptive adjustment capabilities for communication reliability, thereby improving the robustness and decision security of the system in weak signal and high interference environments, and providing a complete toolchain for performance verification before system deployment.

[0194] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments and examples. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the concept of the present invention.

Claims

1. A vehicle-road cooperative decision-making and vehicle motion planning method based on predictive risk assessment, characterized in that, Includes the following steps: (1) Obtain historical driving status information, road topology information, traffic rule information and traffic scene semantic information of the target vehicle and its surrounding vehicles, construct scene semantic encoding vector, and use it together with historical status information and road topology information as input to subsequent modules; (2) Based on the historical driving status information and scene semantic information, a multimodal vehicle trajectory prediction model that integrates scene semantics is constructed to predict the candidate driving trajectory of the target vehicle in multiple future time steps and to determine the driving intention corresponding to the candidate driving trajectory. (3) Based on the scene semantic encoding vector, the key parameters of the risk field are dynamically adjusted through a learnable parameter adjustment mapping function to generate dynamic risk field adjustment parameters, so that the risk assessment has scene adaptive capability. (4) For each candidate driving behavior, construct a counterfactual response model of surrounding vehicles based on counterfactual reasoning, predict the response trajectory of surrounding vehicles, and then construct a counterfactual risk field dependent on the behavior, and calculate the risk change of the behavior compared to the current behavior. (5) Based on the candidate driving trajectory and the interaction relationship between vehicles, construct the multi-vehicle interaction risk field of the target vehicle in the prediction time domain, weight the risk contribution of different interactive vehicles according to the lane relationship of the vehicle, and integrate the counterfactual risk field and scene semantic parameters. (6) Through vehicle-road-cloud collaborative communication, the local risk fields of multiple vehicles are extracted and compressed and encoded before transmission. The local risk fields of multiple vehicles are integrated at roadside units or cloud nodes to construct a collaborative risk map, and structured features are extracted from the map. (7) Based on the collaborative risk map, risk assessment is performed on the candidate driving behaviors of the target vehicle, a comprehensive behavior cost function is constructed, and a two-layer model predictive control is used to generate vehicle control commands, thereby realizing the integrated collaboration of vehicle behavior decision-making and motion control. In the interactive scenario, group collaborative decision-making is achieved through the V2X negotiation mechanism. (8) Based on the system-level simulation platform, the communication reliability of the transmission and fusion process of the collaborative risk map is simulated and verified. The real-time performance and reliability of risk information transmission are evaluated through dynamic channel modeling, and the risk confidence fusion weight is optimized.

2. The method according to claim 1, characterized in that, The multimodal vehicle trajectory prediction model in step (2) includes an encoding layer, a graph convolutional layer, an attention layer, an intent recognition layer, and a decoding layer; the encoding layer uses an LSTM network to fuse historical vehicle trajectory information with local scene semantic encoding; the graph convolutional layer extracts the interaction features between the predicted vehicle and surrounding vehicles in the context of scene semantics. The attention layer calculates the interaction weights between surrounding vehicles and the predicted vehicle, enhancing the model's interpretability; the intent recognition layer fuses graph convolutional features, attention vectors, and global scene semantic summaries to output a driving intent recognition vector; and the decoding layer fuses multi-dimensional features to predict the probability distribution of future trajectories under different intents.

3. The method according to claim 1, characterized in that, The key parameters of the risk field mentioned in step (3) include the equivalent threshold of the same lane, the attenuation rate parameter of the cross lane, the distance attenuation parameter, the amplitude of the lane marking potential field, and the direction correction parameter. In low visibility scenarios, the equivalent threshold of the same lane is increased; in wet and slippery road scenarios, the distance attenuation parameter is decreased; in school areas, the amplitude of the lane marking potential field is increased; and in congested scenarios, the cross lane attenuation rate parameter is adjusted.

4. The method according to claim 1, characterized in that, In step (5), when constructing the multi-vehicle interaction risk field, a single-vehicle risk field is first constructed for each interactive vehicle. The single-vehicle risk field includes risk terms based on motion state, risk terms based on distance, lane marking potential field, and road boundary potential field. Then, the lane relationship weighting coefficient, counterfactual risk fusion weight, and behavior prediction probability are combined to obtain the multi-vehicle interaction risk field. Finally, the risk fields at each time point in the prediction time domain are superimposed to obtain the multi-vehicle interaction risk field in the prediction time domain.

5. The method according to claim 1, characterized in that, In step (6), when constructing the collaborative risk map, each vehicle first calculates its local risk contribution field to the environment, then extracts and compresses the local risk field, broadcasts it via V2X, and finally performs global fusion of the risk data uploaded by each vehicle at the roadside unit or cloud node according to the spatial attenuation weight, and extracts the structured features of the map.

6. The method according to claim 1, characterized in that, The two-layer model predictive control described in step (7) includes upper-layer MPC behavior decision-making and lower-layer MPC motion control; the upper-layer MPC selects the behavior that minimizes the comprehensive behavior cost function from the candidate behavior set as the target decision and outputs the reference trajectory; The lower-level MPC uses a kinematic bicycle model as the prediction model and constructs longitudinal and lateral MPC cost functions respectively. Hard constraints based on scene adaptive risk thresholds are added during the optimization process. The comprehensive behavioral cost function integrates collaborative risk cost, counterfactual risk change cost, trajectory tracking error cost, control smoothness cost, and behavioral prior cost.

7. The method according to claim 1, characterized in that, The communication reliability simulation verification in step (8) includes vehicle-to-vehicle V2X communication simulation and vehicle-to-roadside unit / cloud V2R communication simulation. Path loss model, transmission success rate model, delay model and confidence model are constructed respectively. Physical layer indicators such as channel transmission success rate and delay are quantified into risk confidence, so as to realize the closed-loop verification of communication quality and risk integration.

8. The method according to any one of claims 1-7, characterized in that, The traffic scene semantic information includes traffic events, weather conditions, special areas, and types of traffic participants. It is obtained through vehicle-mounted sensors and communication and encoded to obtain a scene semantic encoding vector.

Images