An automatic driving vehicle path planning method based on multi-agent interaction prediction

Abstract

The application discloses an automatic driving vehicle path planning method based on multi-agent interaction prediction and belongs to the technical field of automatic driving, which comprises the following steps: extracting the historical motion state features of a vehicle and surrounding traffic participants and the road environment features around the vehicle; inputting the extracted features into a multi-agent interaction trajectory prediction network to generate a multi-modal prediction result comprising multiple groups of predicted trajectories of all traffic participants and corresponding confidence; constructing a path planning cost function for each prediction mode in the multi-modal prediction result to generate a corresponding candidate path of the vehicle; constructing a comprehensive evaluation function to output an optimal vehicle path; and directly introducing the multi-agent interaction trajectory prediction result into the vehicle path planning process to generate a candidate path of the vehicle for different prediction traffic evolution modes in the planning stage, and the candidate paths are uniformly evaluated in combination with the prediction probability, so that the problem of decision deviation caused by planning only by relying on a single prediction result is effectively avoided.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to an autonomous vehicle path planning method based on multi-agent interaction prediction. Background Technology

[0002] When autonomous vehicles drive in real-world road environments, they need to generate safe and reasonable driving paths in real time, based on the motion states of surrounding traffic participants and road environment information, while adhering to traffic rules and vehicle motion constraints. Existing autonomous driving path planning methods are mostly based on rule constraints or optimization models. They generate driving paths by constructing objective functions and combining them with vehicle dynamics constraints, and can achieve stable results in scenarios with few traffic participants and simple traffic conditions.

[0003] However, existing technologies have significant drawbacks. For example, trajectory prediction and path planning are often designed as independent functional modules. Prediction results are typically provided only as a few candidate trajectories, which the planning module treats as external static constraints or obstacles, making it difficult to reflect the dynamic interactions between traffic participants over time. In complex traffic scenarios such as urban roads and intersections, the large number of surrounding traffic participants and their complex behavioral interactions exacerbate this deficiency, leading to discrepancies between prediction results and actual traffic conditions. This, in turn, affects the rationality and stability of autonomous vehicle path planning, failing to meet the demands of autonomous driving in complex scenarios.

[0004] Therefore, how to fully utilize multi-agent interactive prediction information in the route planning process to improve the reliability of planning results in complex traffic scenarios has become an urgent problem to be solved. Summary of the Invention

[0005] To address the problems of existing technologies, this invention provides a path planning method for autonomous vehicles based on multi-agent interaction prediction, the method comprising the following steps: S1. Extract the historical motion state characteristics of the vehicle and surrounding traffic participants, as well as the road environment characteristics around the vehicle; S2. Input the historical motion state features and road environment features into the multi-agent interactive trajectory prediction network to generate multiple sets of multimodal prediction results containing all traffic participants and their corresponding confidence levels. The multi-agent interaction trajectory prediction network includes a historical state encoding network, a multi-agent interaction modeling network, a road environment feature encoding network, a feature fusion network, and a multimodal trajectory generation network.

[0006] The historical state encoding network takes each traffic participant and assembles its motion state into a state sequence at consecutive historical moments. This sequence is then input into a feedforward neural network for feature mapping to obtain a high-dimensional representation. Subsequently, the mapped feature sequence is input into a temporal modeling network to model the motion evolution relationship of traffic participants in the time dimension and generate a historical feature vector representing its motion trend.

[0007] The multi-agent interaction modeling network establishes interaction relationships by using the historical characteristics of each traffic participant as node representations, and calculates the correlation weights between traffic participants based on the attention mechanism; combined with the relative motion states between traffic participants, the node features are weighted and aggregated to obtain multi-agent interaction features that integrate the influence of group behavior.

[0008] The road environment feature encoding network represents road structure information as a sequence of points or a set of vectors, and inputs it into a convolutional neural network for feature extraction to obtain environmental feature vectors that characterize the road topology and geometric constraints.

[0009] The feature fusion network integrates multi-agent interaction features with road environment features in the feature dimension, and performs nonlinear mapping through a feedforward neural network to generate a unified joint feature representation of the scene, enabling the model to simultaneously characterize the behavioral relationships between traffic participants and their constraints with the road structure.

[0010] The multimodal trajectory generation network sets up multiple parallel trajectory decoding branches based on the joint feature representation. Each branch is used to generate a future trajectory under a potential motion pattern. At the same time, the probability of each predicted trajectory is output through the mode scoring branch and normalized to obtain the corresponding confidence level, thereby generating multiple sets of candidate predicted trajectories and their probabilities to characterize the uncertainty and diversity of the future behavior of traffic participants.

[0011] S3. For each prediction mode in the multimodal prediction results, construct a path planning cost function that includes safety, comfort and traffic rule constraints, and generate the corresponding candidate path for the vehicle through a differentiable nonlinear optimization method. S4. Based on the confidence level of the multimodal prediction results, the calculation results of the path planning cost function, and the planning consistency among candidate paths, a comprehensive evaluation function is constructed, and the optimal vehicle path is selected according to the output of the comprehensive evaluation function.

[0012] Furthermore, the expression for the historical motion state features is: ; in, This represents the historical motion state characteristics of the i-th traffic participant; i represents the unique identifier of the traffic participant, signifying the i-th traffic participant; Indicates the first Traffic participants Two-dimensional coordinates at time; Indicates the first Traffic participants The heading angle at any given moment; Indicates the first Traffic participants in The speed of time; Indicates the first The attributes of a traffic participant, namely, the type identifier of the traffic participant; Indicates the total duration of state sampling; All historical motion state characteristics are uniformly transformed into the vehicle coordinate system to eliminate coordinate differences between different traffic participants, providing standardized motion state input data for multi-subject interactive trajectory prediction networks.

[0013] Furthermore, the road environment characteristics The expression is: ; in, ; in, This represents the road environment characteristics of the i-th lane; The dimension identifier of the lane sampling point represents the s-th sampling point; Indicates the first Lane 1 Two-dimensional coordinates of each sampling point; Indicates the first Lane 1 The heading angle of each sampling point; Indicates the first The attributes of each lane; S represents the total number of lane sampling points; The road environment features provide a basis for road structure constraints for the multi-agent interactive trajectory prediction network.

[0014] Furthermore, the multi-agent interactive trajectory prediction network characterizes the mutual influence relationships between different traffic participants through interaction features, the expression of which is: ; in, This represents the historical motion state characteristics of the i-th traffic participant after the fusion and interaction effects; A unique identifier representing a traffic participant, indicating the j-th traffic participant; Indicates traffic participants Traffic participants The interaction influence weight; This represents a function for modeling interaction relationships; Indicates traffic participants and The relative motion state between them; For the first Historical motion characteristics of each traffic participant; For the first Historical motion characteristics of each traffic participant; A unique identifier representing a traffic participant, indicating the k-th traffic participant; exp( ) represents the natural exponential function; This interactive feature is used to enhance the ability of multi-agent interactive trajectory prediction networks to model the dynamic correlation of traffic participants, avoiding treating the future movement of each traffic participant as an independent prediction problem.

[0015] Furthermore, the multi-agent interactive trajectory prediction network generates multiple sets of multimodal prediction results containing predicted trajectories of all traffic participants and their corresponding confidence levels, including: The historical motion state features are respectively subjected to feature mapping and temporal feature extraction to obtain the historical motion feature vector corresponding to each traffic participant; Feature aggregation is performed on road elements in the road environment characteristics to obtain the road feature vector corresponding to each road element; The historical motion vector and the road feature vector are fused to generate joint features; Based on the joint features, generate two-dimensional displacement sequences of traffic participants within a preset time period under different traffic interaction evolution modes and corresponding mode scores; accumulate the two-dimensional displacement sequences in the time dimension to obtain the future path point sequence of traffic participants under the corresponding traffic interaction evolution mode, i.e., the predicted trajectory of traffic participants; normalize the scores of the modes to obtain the confidence level of the corresponding traffic interaction evolution mode, i.e., the confidence level of the predicted trajectory.

[0016] Furthermore, the multi-agent interactive trajectory prediction network fuses the traffic participant features after interactive modeling with road environment features to generate a joint feature representation, the expression of which is: ; in, For the first The joint characteristics of each traffic participant; Indicates the feature fusion function; For the current scenario The characteristics of the road environment; This joint feature integrates motion state and environmental constraint information and is directly used to generate multimodal prediction results.

[0017] Furthermore, the expression for the multimodal prediction result is: ; in, Indicates the first The set of predicted trajectories for all traffic participants within the vehicle's perception range under a certain interactive evolutionary model; This represents the set of confidence levels for the corresponding predicted trajectory; This represents the predictive solution function; This represents the total number of all traffic participants within the vehicle's perception range; Indicates the total number of prediction patterns; Each prediction model corresponds to a potential traffic evolution scenario, providing differentiated constraints for constructing the path planning cost function.

[0018] Furthermore, the path planning cost function is expressed as follows: ; in, This represents the value of the path planning cost function constructed based on the k-th prediction model; It represents the safety-related costs and is used to measure the spatial relationship between the vehicle's path and the predicted trajectories of surrounding traffic participants; This represents the sequence of control variables for the vehicle. This represents the motion state of the vehicle at time t; Weighting coefficients representing security-related costs; Weighting coefficients representing comfort-related costs; Weighting coefficients representing the costs of traffic rules; It represents the cost related to comfort and is used to constrain changes in the vehicle's acceleration, jerk, etc., to avoid sudden acceleration, sudden deceleration, or sharp steering behavior; It represents the cost of traffic rules and is used to constrain the vehicle's path to meet lane constraints, traffic control, and road structure limitations, preventing behaviors that do not comply with traffic rules, such as crossing the line or driving against the flow of traffic. T represents the total duration of the planning time domain; This cost function provides a clear optimization objective for nonlinear optimization methods and directly determines the degree to which the constraints of candidate paths are satisfied.

[0019] Furthermore, the expression for the sequence of control variables is: ; in, Indicates that the car is in the first Control inputs for each planning time; Indicates the length of the planning time domain; The nonlinear optimization method solves... Obtain the optimal control sequence Based on this optimal control sequence and vehicle kinematics model, candidate paths for the vehicle are generated; This represents the optimal control sequence of the vehicle obtained based on the k-th prediction mode; The operator symbol represents the independent variable that minimizes the objective function; The expression for the vehicle kinematics model is: ; in, Indicates that the car is The state of motion at any given moment; This represents the state propagation function that satisfies the vehicle kinematic constraints. This represents the motion state of the vehicle at time t-1; The optimal control sequence is transformed into executable candidate paths for the vehicle through the vehicle kinematics model.

[0020] Furthermore, the expression for the comprehensive evaluation function is: ; in, Indicates the first A comprehensive score for each candidate path; These represent the weighting coefficients of the prediction confidence mapping function; The weighting coefficients represent the normalization function of the planning cost; Represents the weighting coefficients of the planning stability evaluation function; m represents the unique identifier of the prediction model, indicating the m-th prediction model. This represents the prediction confidence mapping function; high-confidence prediction patterns receive higher scores in path selection. This represents the normalization function of planning costs; candidate paths with lower total costs receive higher scores. This represents the planning stability evaluation function, used to calculate the vehicle path. The stability of the vehicle path planning in terms of spatial location relative to other candidate paths is considered to prevent the vehicle path planning from overreacting to minor changes in the prediction results and generating a strongly jittery planned path. This represents the candidate paths for the vehicle generated using the vehicle kinematics model. ; The optimal vehicle path passes through Sure; This represents the optimal interaction evolution mode; Plan the path for the final output vehicle; The optimal prediction model is determined by the maximum value of the comprehensive evaluation function, which makes the final path achieve the optimal balance in prediction matching degree, constraint satisfaction and driving stability.

[0021] The beneficial effects of this invention are: By directly incorporating multi-agent interactive trajectory prediction results into the autonomous vehicle path planning process, candidate paths for each predicted traffic evolution mode are generated during the planning phase. These paths are then uniformly evaluated based on the predicted probabilities, effectively avoiding the decision-making bias caused by relying solely on a single prediction result. This method can fully consider the uncertain movement trends of surrounding traffic participants in complex traffic scenarios, ensuring that the planned autonomous vehicle path aligns with the actual traffic evolution, thereby improving the stability and reliability of the path planning results.

[0022] By decoupling multi-agent trajectory prediction results into a finite number of prediction patterns and independently executing vehicle path planning under each prediction pattern, this method effectively avoids the problem of drastically increasing planning complexity caused by simultaneously handling multiple uncertain traffic evolution scenarios in a single planning process. This approach decomposes the originally high-dimensional, strongly coupled uncertain planning problem into multiple low-complexity deterministic planning sub-problems, ensuring that the path planning process remains solvable and computationally stable, thus achieving an effective balance between prediction uncertainty modeling and real-time path planning in complex traffic scenarios. Attached Figure Description

[0023] Figure 1 A flowchart illustrating the autonomous vehicle path planning method based on multi-agent interaction prediction provided by this invention; Figure 2 A flowchart illustrating the trajectory prediction method based on multi-subject interaction provided by this invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figure 1-2 This invention provides a path planning method for autonomous vehicles based on multi-agent interaction prediction, comprising: S1. First, the collected traffic dataset (such as the Waymo Open Motion Dataset) is filtered based on traffic rules and data integrity to remove abnormal data that does not conform to traffic rules, and a standard dataset for trajectory prediction and path planning is constructed. Then, with the vehicle as the center, surrounding traffic participants (such as the 15 closest ones, including motor vehicles, pedestrians and cyclists) are selected within its perception range. The historical motion state features of each traffic participant in the historical time domain (such as 20 frames, corresponding to a time length of 2 seconds) are extracted. The historical motion state features include two-dimensional position, heading angle, speed and traffic participant type identifier. At the same time, the road structure around the vehicle is vectorized and the road environment features such as lane center lines, lane boundaries and traffic control elements (such as pedestrian crossings) are extracted. The center lines of each lane are sampled at equal intervals (such as 20 sampling points) to represent their geometric shape and the lane attribute information is recorded. S2. The historical motion state features and road environment features are input into a multi-agent interaction trajectory prediction network. This network consists of a historical state encoding network, a multi-agent interaction modeling network, a road environment feature encoding network, a feature fusion network, and a multimodal trajectory generation network. The historical state encoding network employs a three-layer temporal feature extraction structure. The state at each moment is first mapped through a two-layer fully connected network (with a hidden dimension of 128 and ReLU activation function), and then fed into a single-layer temporal feature extraction network to obtain a historical motion feature vector. The multi-agent interaction modeling network takes the historical motion features of each traffic participant as input, and through the interaction feature mapping structure formed by the three-layer fully connected network, it introduces the relative motion states between traffic participants for weighted aggregation to obtain a feature table that integrates multi-agent interaction information. The road environment feature encoding network uses a three-layer one-dimensional convolutional neural network (each with a kernel size of 3 and channel numbers of 64, 128, and 128 respectively) to process the road point sequence. A max pooling operation is followed by the convolutional layer to achieve feature aggregation. The feature fusion network concatenates the traffic participant features and road environment features, and then performs fusion mapping through a two-layer fully connected network to generate a unified joint feature representation. The multimodal trajectory generation network adopts a multi-branch parallel output structure (such as 6 parallel decoding branches). Based on the shared joint feature representation, it outputs the two-dimensional displacement sequence of traffic participants within the next 30 frames and the corresponding mode score. The mode score is normalized by Softmax to form the confidence level corresponding to each prediction mode. Finally, multiple sets of multimodal prediction results containing all traffic participants and their corresponding confidence levels are generated. The historical motion state sequence and road environment features are input into a multi-agent interactive trajectory prediction network. This network consists of a historical state encoding network, a multi-agent interactive modeling network, a road environment feature encoding network, a feature fusion network, and a multimodal trajectory generation network.

[0026] The historical state encoding network employs a temporal encoding structure combining a multilayer perceptron and a long short-term memory network. For each traffic participant, a state vector is constructed at each historical moment, and feature mapping is performed through a two-layer fully connected network. The hidden dimension of each fully connected network is set to 128, and the activation function is ReLU. Subsequently, the mapped historical state sequence is input into a single-layer LSTM for temporal modeling. The hidden state is recursively updated to capture the motion evolution relationship of traffic participants between consecutive time steps, and a fixed-length historical motion feature vector is output.

[0027] The multi-agent interaction modeling network uses the historical motion features of each traffic participant as node representations to construct an interaction relationship graph between traffic participants and implements interaction modeling based on a self-attention mechanism. Specifically, the influence weights between traffic participants are calculated by the correlation between queries, keys, and values. Simultaneously, the relative motion states such as relative positions and relative speeds between traffic participants are combined to weighted aggregate of multi-agent features, resulting in a feature representation that integrates group interaction information. To enhance feature representation capabilities, the interaction feature mapping is constructed using a three-layer feedforward fully connected network, with each layer having a hidden dimension of 128 and using ReLU as the activation function.

[0028] The road environment feature encoding network uses a three-layer one-dimensional convolutional neural network to process road point sequences. The kernel size is set to 3 for all convolutional layers, and the number of output channels is 64, 128, and 128 respectively. A max pooling layer is set after the convolutional layers to achieve spatial feature compression and aggregation of key structural information, generating a road environment representation vector.

[0029] The feature fusion network concatenates the interaction features of traffic participants with the road environment features along the feature dimension, and performs nonlinear mapping through a two-layer fully connected network. The hidden dimensions are set to 256 and 128, and the activation function is ReLU. This generates a unified joint feature representation of the scene, enabling the model to simultaneously characterize the behavioral relationships between traffic participants and their constraints with the road structure.

[0030] The multimodal trajectory generation network consists of a shared feature encoding module, a trajectory decoding module, and a pattern scoring module. The shared feature encoding module takes the previously generated joint feature representation as input and performs feature transformation through a single fully connected network, with an output dimension of 128, to generate unified initial decoding features. The trajectory decoding module employs a multi-branch parallel decoding structure, containing six independent trajectory decoding branches with non-shared parameters. Each branch consists of two fully connected network layers with hidden dimensions of 256 and 128, respectively, and uses ReLU as the activation function. The decoding network takes the joint features as input and directly regresses the two-dimensional displacement sequence of the next 30 frames through linear mapping to represent the planar motion trajectory of traffic participants in the future time domain. The pattern scoring module inputs the joint features into a single fully connected network to generate a scoring vector of length 6, and performs normalization using the Softmax function to obtain the probability value corresponding to each predicted mode, representing the uncertainty of different motion modes.

[0031] Finally, the multimodal trajectory generation network outputs six candidate future trajectories and their corresponding probabilities, thus forming multimodal trajectory prediction results for multi-agent scenarios.

[0032] S3. For each prediction mode in the multimodal prediction results, construct a path planning cost function that includes safety, comfort, and traffic rule constraints. Initialize a 30-frame sequence of discrete path points for the vehicle (each frame includes two-dimensional position and heading angle). Use a differentiable nonlinear numerical optimization method relying on NumPy, SciPy, and Theseus differentiable optimization tools to iteratively check and adjust the path points in time sequence: calculate the minimum distance between the vehicle's path point and the predicted trajectory points of all traffic participants in this interaction mode; when the distance is less than 3m, adjust the position of the path point to increase the safety distance; check the smoothness of the path; when the acceleration change of adjacent path points exceeds 3m / s², ... 2 If the heading angle changes by more than 0.3 radians, the path segment is smoothed; road constraints are checked, and if the lateral distance from the path point to the nearest lane centerline exceeds 1.75m, a penalty is imposed on the path point and the path is guided back to the feasible area in subsequent iterations. The above checks and adjustments are repeatedly performed throughout the planning time domain (consistent with the path generation frequency). The corresponding candidate paths for the vehicle are generated by solving the optimal control sequence and combining it with the vehicle kinematics model. S4. Based on the confidence level of the multimodal prediction results, the calculation results of the path planning cost function, and the planning consistency among candidate paths in terms of spatial location, driving direction, and speed changes, a comprehensive evaluation function is constructed. The weight coefficients in the comprehensive evaluation function can be adjusted according to the actual scenario (e.g., α=0.5, β=0.3, γ=0.2). The prediction confidence mapping function enables high-confidence prediction modes to obtain higher scores, the planning cost normalization function enables candidate paths with lower total costs to obtain higher scores, and the planning stability evaluation function is used to avoid the autonomous vehicle path planning from overreacting to minor changes in the prediction results and generating strongly jittery planning paths. According to the output of the comprehensive evaluation function, the path with high confidence and the fewest constraint violations is selected first, and the candidate path with the best evaluation result is selected as the final execution path.

[0033] In some embodiments, the historical motion state feature expression is: ; in, This represents the historical motion state characteristics of the i-th traffic participant; i represents the unique identifier of the traffic participant, signifying the i-th traffic participant; Indicates the first Traffic participants in Two-dimensional coordinates of time, Indicates the first Traffic participants in The heading angle at any moment, Indicates the first Traffic participants in The speed of time, Indicates the first The attributes of a traffic participant Indicates the total duration of state sampling; All historical motion state feature information comes from a standard dataset that has been filtered for traffic rules and data integrity. Scene samples with obvious anomalies have been removed, and the data has been uniformly transformed to a local coordinate system with the vehicle as the origin to eliminate coordinate differences between different traffic participants. This provides standardized motion state input data for the multi-agent interactive trajectory prediction network, ensuring the consistency and effectiveness of the network input.

[0034] In some embodiments, the expression for the road environment characteristics is: ; in, This represents the road environment characteristics of the i-th lane; The dimension identifier of the lane sampling point represents the s-th sampling point; Indicates the first Lane 1 Two-dimensional coordinates of each sampling point Indicates the first Lane 1 The heading angle of each sampling point Indicates the first The attributes of each lane, such as straight lane, left-turn lane, right-turn lane, whether lane changing is allowed, etc., S represents the total number of lane sampling points, which can be set to 20 in practical applications. The geometry of the lane is characterized by equidistant sampling. The road environment features are obtained by vectorizing and extracting road elements such as lane centerlines, lane boundaries, and pedestrian crossings around the vehicle, providing road structure constraints for the multi-subject interactive trajectory prediction network, so that the predicted trajectory can conform to the actual road geometry constraints.

[0035] Furthermore, the multi-agent interactive trajectory prediction network generates multiple sets of multimodal prediction results containing predicted trajectories of all traffic participants and their corresponding confidence levels, including: The historical motion state features are respectively subjected to feature mapping and temporal feature extraction to obtain the historical motion feature vector corresponding to each traffic participant; Feature aggregation is performed on road elements in the road environment characteristics to obtain the road feature vector corresponding to each road element; The historical motion vector and the road feature vector are fused to generate joint features; Based on the joint features, generate two-dimensional displacement sequences of traffic participants within a preset time period under different traffic interaction evolution modes and corresponding mode scores; accumulate the two-dimensional displacement sequences in the time dimension to obtain the future path point sequence of traffic participants under the corresponding traffic interaction evolution mode, i.e., the predicted trajectory of traffic participants; normalize the scores of the modes to obtain the confidence level of the corresponding traffic interaction evolution mode, i.e., the confidence level of the predicted trajectory.

[0036] In some embodiments, the multi-agent interactive trajectory prediction network characterizes the mutual influence relationships between different traffic participants through interactive features, the expression of which is: ; in, This represents the historical motion state characteristics of the i-th traffic participant after the fusion and interaction effects; A unique identifier representing a traffic participant, indicating the j-th traffic participant; Indicates traffic participants Traffic participants The interaction influence weight, This represents a function for modeling interaction relationships. Indicates traffic participants and The relative motion state between them includes relative position, relative velocity, and relative heading. For the first Historical movement characteristics of traffic participants For the first Historical motion characteristics of each traffic participant; A unique identifier for a traffic participant, representing the k-th traffic participant; exp( ) represents the natural exponential function; This interactive feature is implemented through a multi-agent interactive modeling network, which adopts an interactive feature mapping structure composed of a three-layer fully connected network. During the feature mapping process, the relative motion state information between traffic participants is introduced. This feature is obtained by weighted aggregation of the features of surrounding traffic participants. This enhances the multi-agent interactive trajectory prediction network's ability to model the dynamic correlation of traffic participants, avoids treating the future motion of each traffic participant as an independent prediction problem, and thus generates a joint prediction result with scene consistency.

[0037] In some embodiments, the multi-agent interactive trajectory prediction network fuses the traffic participant features after interactive modeling with road environment features to generate a joint feature representation, the expression of which is: ; in, For the first The combined characteristics of traffic participants The feature fusion function is specifically implemented by concatenating the traffic participant features after interactive modeling with road environment features, and then performing fusion mapping through a two-layer fully connected network. For the current scenario The road environment features are integrated; this joint feature integrates the motion state information of traffic participants and the road environment constraint information, eliminating the limitations of a single feature dimension, and can be directly used to generate multimodal prediction results, providing comprehensive feature support for subsequent predictions.

[0038] In some embodiments, the expression for the multimodal prediction result is: ; in, Indicates the first The predicted trajectory set of all traffic participants within the vehicle's perception range under this interactive evolution mode is specifically the path point sequence obtained by accumulating the two-dimensional displacement sequences of traffic participants over the next 30 frames in the time dimension. This represents the set of confidence scores for the corresponding predicted trajectory, obtained by normalizing the pattern scores using Softmax. This represents the prediction and solution function. This represents the total number of all traffic participants within the vehicle's perception range, for example, 15. This indicates the total number of prediction modes, which can be adjusted according to the complexity of the scenario and the availability of computing resources, such as 3, 4, 6, etc. Each prediction model corresponds to a potential traffic evolution scenario. The predicted trajectories under different models reflect different interactive evolution trends among traffic participants, providing differentiated constraints for the construction of the path planning cost function, enabling path planning to cover a variety of possible traffic situations.

[0039] In some embodiments, the path planning cost function expression is: ; in, This represents the value of the path planning cost function constructed based on the k-th prediction model; This represents the safety-related cost, used to measure the spatial relationship between the vehicle's path and the predicted trajectory of surrounding traffic participants. When the minimum distance between the vehicle's path point and the predicted trajectory point of the traffic participants is less than 3m, it is determined that the safety cost exceeds the limit and triggers path adjustment. This represents the comfort-related cost, used to constrain changes in the vehicle's acceleration and jerk, preventing sudden acceleration, deceleration, or sharp steering. When the acceleration change between adjacent path points exceeds 3 m / s², this is considered a cost. 2 If the heading angle changes by more than 0.3 radians, it is determined that the comfort cost exceeds the limit and the route segment is smoothly adjusted. This represents the cost of traffic rules, used to constrain the vehicle's path to meet lane constraints, traffic control, and road structure limitations. When the lateral distance from a path point to the nearest lane centerline exceeds 1.75m, it is determined that the cost of traffic rules has exceeded the limit and a penalty is imposed. These are the weighting coefficients for the corresponding costs, which can be flexibly adjusted according to the priority requirements of the actual traffic scenario; This represents the sequence of control variables for the vehicle. This represents the motion state of the vehicle at time t; Weighting coefficients representing security-related costs; Weighting coefficients representing comfort-related costs; The weighting coefficient represents the cost of traffic rules; T represents the total duration of the planning time domain, which can be set to 30 frames in practical applications. This cost function provides a clear optimization objective for nonlinear optimization methods. By quantifying the degree to which the constraints of the three core dimensions of safety, comfort, and compliance are met, it directly determines the quality of candidate paths and ensures that the generated candidate paths meet the driving requirements of autonomous vehicles.

[0040] In some embodiments, the expression for the sequence of control variables is: ; in, Indicates that the car is in the first Control inputs at each planning time point This represents the length of the planning time domain and the total duration of the planning time domain. Keep it consistent, for example, 30 frames; The nonlinear optimization method solves... Obtain the optimal control sequence This nonlinear optimization method is a differentiable nonlinear numerical optimization method that relies on differentiable optimization tools such as NumPy, SciPy, and Theseus to achieve efficient solution; candidate paths for the vehicle are generated based on the optimal control sequence and the vehicle kinematic model. This represents the optimal control sequence of the vehicle obtained based on the k-th prediction mode; The operator symbol represents the independent variable that minimizes the objective function; The expression for the vehicle kinematics model is: ; in, Indicates that the car is The state of motion at any given moment This represents the state propagation function that satisfies the vehicle's kinematic constraints, used to transform control inputs into the actual motion state of the vehicle. This represents the motion state of the vehicle at time t-1; The optimal control sequence is transformed into executable candidate paths for the vehicle through the vehicle kinematics model, and the path generation process must be consistent with the path generation frequency to ensure the real-time performance and continuity of the path.

[0041] In some embodiments, the expression of the comprehensive evaluation function is: ; in, Indicates the first The overall score of the candidate paths, This represents the prediction confidence mapping function. High-confidence prediction models receive higher scores in path selection, directly reflecting the probability of the predicted trajectory matching the actual traffic situation. This represents the normalization function for planning costs. By standardizing the path planning costs, candidate paths with lower total costs receive higher scores, quantifying the degree to which the constraints are satisfied by the candidate paths. This represents the planning stability evaluation function, used to calculate the vehicle path. The stability of the autonomous vehicle path planning relative to other candidate paths in terms of spatial location, driving direction, and speed changes is assessed to prevent the path planning from overreacting to minor changes in the prediction results and generating a strongly jittery planned path. This represents the candidate paths for the vehicle generated using the vehicle kinematics model, i.e. ; , and This represents the weighting coefficients of each function. These represent the weighting coefficients of the prediction confidence mapping function; The weighting coefficients represent the normalization function of the planning cost; The weighting coefficients of the planning stability evaluation function are represented by m; m represents the unique identifier of the prediction mode, indicating the m-th prediction mode; in practical applications, it can be set to... =0.5、 =0.3、 =0.2, and the weight ratio of the three can also be adjusted according to the needs of the scenario to balance the three evaluation dimensions of prediction credibility, planning cost and driving stability.

[0042] In some embodiments, the optimal vehicle path is achieved through... Sure; in, This represents the optimal interactive evolutionary model, i.e., the prediction model with the highest overall score. Plan the path for the final output vehicle; The optimal prediction mode is determined by the maximum value of the comprehensive evaluation function. The autonomous vehicle path corresponding to this mode achieves the best balance in terms of prediction matching degree and high confidence prediction trajectory, constraint satisfaction and driving stability (small difference from other candidate paths). In practical applications, when the multimodal prediction results generate K candidate paths, such as 6, the optimal path with the best comprehensive performance is selected from the multiple candidate paths using this formula, which is then used as the final execution path of the autonomous vehicle to ensure the safety, rationality and stability of driving in complex traffic scenarios.

[0043] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A path planning method for autonomous vehicles based on multi-agent interaction prediction, characterized in that, The method includes the following steps: S1. Extract the historical motion state characteristics of the vehicle and surrounding traffic participants, as well as the road environment characteristics around the vehicle; S2. Input the historical motion state features and road environment features into the multi-agent interactive trajectory prediction network to generate multiple sets of multimodal prediction results containing all traffic participants and their corresponding confidence levels. S3. For each prediction mode in the multimodal prediction results, construct a path planning cost function that includes safety, comfort and traffic rule constraints, and generate the corresponding candidate path for the vehicle through a differentiable nonlinear optimization method. S4. Based on the confidence level of the multimodal prediction results, the calculation results of the path planning cost function, and the planning consistency among candidate paths, a comprehensive evaluation function is constructed, and the optimal vehicle path is selected according to the output of the comprehensive evaluation function.

2. The autonomous vehicle path planning method based on multi-agent interaction prediction according to claim 1, characterized in that, The expression for the historical motion state feature is: ； in, This represents the historical motion state characteristics of the i-th traffic participant; i represents the unique identifier of the traffic participant, signifying the i-th traffic participant; Indicates the first Traffic participants in Two-dimensional coordinates at time; Indicates the first Traffic participants in The heading angle at any given moment; Indicates the first Traffic participants in The speed of time; Indicates the first The attributes of a traffic participant, namely, the type identifier of the traffic participant; Indicates the total duration of state sampling; All historical motion state characteristics are uniformly transformed into the vehicle coordinate system to eliminate coordinate differences between different traffic participants, providing standardized motion state input data for multi-subject interactive trajectory prediction networks.

3. The autonomous vehicle path planning method based on multi-agent interaction prediction according to claim 1, characterized in that, The road environment characteristics The expression is: ； in, This represents the road environment characteristics of the i-th lane; The dimension identifier of the lane sampling point represents the s-th sampling point; Indicates the first Lane 1 Two-dimensional coordinates of each sampling point; Indicates the first Lane 1 The heading angle of each sampling point; Indicates the first The attributes of each lane; S represents the total number of lane sampling points; The road environment features provide a basis for road structure constraints for the multi-agent interactive trajectory prediction network.

4. The autonomous vehicle path planning method based on multi-agent interaction prediction according to claim 1, characterized in that, The multi-agent interactive trajectory prediction network generates multiple sets of multimodal prediction results containing predicted trajectories of all traffic participants and corresponding confidence levels, including: The historical motion state features are respectively subjected to feature mapping and temporal feature extraction to obtain the historical motion feature vector corresponding to each traffic participant; Feature aggregation is performed on road elements in the road environment characteristics to obtain the road feature vector corresponding to each road element; The historical motion vector and the road feature vector are fused to generate joint features; Based on the joint features, generate two-dimensional displacement sequences of traffic participants within a preset time period under different traffic interaction evolution modes and corresponding mode scores; accumulate the two-dimensional displacement sequences in the time dimension to obtain the future path point sequence of traffic participants under the corresponding traffic interaction evolution mode, i.e., the predicted trajectory of traffic participants; normalize the scores of the modes to obtain the confidence level of the corresponding traffic interaction evolution mode, i.e., the confidence level of the predicted trajectory.

5. The autonomous vehicle path planning method based on multi-agent interaction prediction according to claim 4, characterized in that, The multi-agent interactive trajectory prediction network characterizes the mutual influence relationships between different traffic participants through interactive features, and the expression of the interactive features is as follows: ； in, This represents the historical motion state characteristics of the i-th traffic participant after the fusion and interaction effects; A unique identifier representing a traffic participant, indicating the j-th traffic participant; Indicates traffic participants Traffic participants The interaction influence weight; This represents a function for modeling interaction relationships; Indicates traffic participants and The relative motion state between them; For the first Historical motion characteristics of each traffic participant; For the first Historical motion characteristics of each traffic participant; A unique identifier representing a traffic participant, indicating the k-th traffic participant; exp( ) represents the natural exponential function; This interactive feature is used to enhance the ability of multi-agent interactive trajectory prediction networks to model the dynamic correlation of traffic participants, avoiding treating the future movement of each traffic participant as an independent prediction problem.

6. The autonomous vehicle path planning method based on multi-agent interaction prediction according to claim 4, characterized in that, The multi-agent interactive trajectory prediction network fuses the traffic participant features after interactive modeling with road environment features to generate a joint feature representation. The expression for the joint feature is as follows: ； in, For the first The joint characteristics of each traffic participant; Indicates the feature fusion function; For the current scenario The characteristics of the road environment; This joint feature integrates motion state and environmental constraint information and is directly used to generate multimodal prediction results.

7. The autonomous vehicle path planning method based on multi-agent interaction prediction according to claim 6, characterized in that, The expression for the multimodal prediction result is: ； in, Indicates the first The set of predicted trajectories for all traffic participants within the vehicle's perception range under a certain interactive evolutionary model; This represents the set of confidence levels for the corresponding predicted trajectory; This represents the predictive solution function; This represents the total number of all traffic participants within the vehicle's perception range; Indicates the total number of prediction patterns; Each prediction model corresponds to a potential traffic evolution scenario, providing differentiated constraints for constructing the path planning cost function.

8. The autonomous vehicle path planning method based on multi-agent interaction prediction according to claim 1, characterized in that, The path planning cost function expression is as follows: ； in, This represents the value of the path planning cost function constructed based on the k-th prediction model; It represents the safety-related costs and is used to measure the spatial relationship between the vehicle's path and the predicted trajectories of surrounding traffic participants; This represents the sequence of control variables for the vehicle. This represents the motion state of the vehicle at time t; Weighting coefficients representing security-related costs; Weighting coefficients representing comfort-related costs; Weighting coefficients representing the costs of traffic rules; It represents the cost related to comfort and is used to constrain changes in the vehicle's acceleration, jerk, etc., to avoid sudden acceleration, sudden deceleration, or sharp steering behavior; It represents the cost of traffic rules and is used to constrain the vehicle's path to meet lane constraints, traffic control, and road structure limitations, preventing behaviors that do not comply with traffic rules, such as crossing the line or driving against the flow of traffic. T represents the total duration of the planning time domain; This cost function provides a clear optimization objective for nonlinear optimization methods and directly determines the degree to which the constraints of candidate paths are satisfied.

9. The autonomous vehicle path planning method based on multi-agent interaction prediction according to claim 8, characterized in that, The expression for the sequence of control variables is: ； in, Indicates that the car is in the first Control inputs for each planning time; Indicates the length of the planning time domain; The nonlinear optimization method solves... Obtain the optimal control sequence Based on this optimal control sequence and vehicle kinematics model, candidate paths for the vehicle are generated; This represents the optimal control sequence of the vehicle obtained based on the k-th prediction mode; The operator symbol represents the independent variable that minimizes the objective function; The expression for the vehicle kinematics model is: ; in, Indicates that the car is The state of motion at any given moment; This represents the state propagation function that satisfies the vehicle kinematic constraints. This represents the motion state of the vehicle at time t-1; The optimal control sequence is transformed into executable candidate paths for the vehicle through the vehicle kinematics model.

10. The autonomous vehicle path planning method based on multi-agent interaction prediction according to claim 1, characterized in that, The expression for the comprehensive evaluation function is: ； in, Indicates the first A comprehensive score for each candidate path; These represent the weighting coefficients of the prediction confidence mapping function; The weighting coefficients represent the normalization function of the planning cost; Represents the weighting coefficients of the planning stability evaluation function; m represents a unique identifier for the prediction pattern, indicating the m-th prediction pattern; This represents the prediction confidence mapping function; high-confidence prediction patterns receive higher scores in path selection. This represents the normalization function of planning costs; candidate paths with lower total costs receive higher scores. This represents the planning stability evaluation function, used to calculate the vehicle path. The stability of the vehicle path planning in terms of spatial location relative to other candidate paths is considered to prevent the vehicle path planning from overreacting to minor changes in the prediction results and generating a strongly jittery planned path. This represents the candidate paths for the vehicle generated using the vehicle kinematics model. ; The optimal vehicle path passes through Sure; This represents the optimal interaction evolution mode; Plan the path for the final output vehicle; The optimal prediction model is determined by the maximum value of the comprehensive evaluation function, which makes the final path achieve the optimal balance in prediction matching degree, constraint satisfaction and driving stability.

Images