A driving trajectory planning method and device, electronic equipment and storage medium

Abstract

This invention provides a method, apparatus, electronic device, and storage medium for planning driving trajectories. The method involves: collecting historical trajectories of traffic participants other than the vehicle to be planned; predicting the future trajectories of traffic participants using the historical trajectories; and using the future trajectories for distributed robust stochastic model predictive control to plan the optimal trajectory of the vehicle to be planned. In this scheme, the historical trajectories of traffic participants other than the vehicle to be planned are used to predict the future trajectories of traffic participants, and then distributed robust stochastic model predictive control is performed using the future trajectories of traffic participants to plan the optimal trajectory of the vehicle to be planned. The planning of driving trajectories fully considers the future trajectories of other traffic participants, thereby improving the planning effect of driving trajectories.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving and intelligent transportation technology, specifically to a method, apparatus, electronic device, and storage medium for planning driving trajectories. Background Technology

[0002] With the development of autonomous driving technology, it has gradually become popular in daily driving. The implementation of autonomous driving technology requires the planning of driving trajectories.

[0003] Currently, deterministic model predictive control (MPC) is commonly used for driving trajectory planning. However, when using MPC for driving trajectory planning, it is necessary to assume that the driving environment is known and deterministic, and it cannot handle uncertainties such as sensor noise and sudden changes in other vehicle driving intentions, resulting in poor driving trajectory planning performance. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method, apparatus, electronic device and storage medium for planning driving trajectories, in order to solve the problem of poor driving trajectory planning performance in MPC.

[0005] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0006] The first aspect of this invention discloses a method for planning a driving trajectory, the method comprising:

[0007] Collect historical trajectories of traffic participants other than the vehicles to be planned;

[0008] The future trajectories of the traffic participants are predicted using the historical trajectories.

[0009] The future trajectory is used for distributed robust stochastic model predictive control to plan the optimal trajectory of the vehicle to be planned.

[0010] Preferably, the collection of historical trajectories of traffic participants other than the vehicle to be planned includes:

[0011] Acquire environmental data collected by various sensors;

[0012] The environmental data is preprocessed;

[0013] Target tracking is performed using the preprocessed environmental data to obtain the historical trajectories of traffic participants other than the vehicle to be planned.

[0014] Preferably, the step of predicting the future trajectory of the traffic participant using the historical trajectory includes:

[0015] The historical trajectory is input into a preset spatiotemporal graph neural network for feature extraction to obtain the spatiotemporal feature representation of the traffic participants;

[0016] The spatiotemporal feature representation and the historical trajectory are input into a preset diffusion probability model to predict the trajectory, so as to obtain the future trajectory of the traffic participant.

[0017] Preferably, the step of using the future trajectory for distributed robust stochastic model predictive control to plan the optimal trajectory of the vehicle to be planned includes:

[0018] The future trajectories of different traffic participants are assigned to different computing units, and the computing units perform optimization solutions based on the assigned future trajectories, the optimization objective function of the distributed robust stochastic model predictive control, and the constraints. The constraints include chance constraints and control constraints. Each computing unit exchanges boundary information once every preset time interval during the optimization solution process.

[0019] The optimal trajectory of the vehicle to be planned is obtained by aggregating the optimization results of each computing unit.

[0020] Preferably, the step of tracking the target using the preprocessed environmental data to obtain the historical trajectories of traffic participants other than the vehicle to be planned includes:

[0021] Target tracking is performed using the preprocessed environmental data to obtain target detection results from each of the sensors;

[0022] The target detection results from each of the aforementioned sensors are correlated to form the historical trajectories of traffic participants other than the vehicle to be planned.

[0023] Preferably, after obtaining the future trajectory of the traffic participant, the method further includes:

[0024] Collect the real trajectories of the traffic participants;

[0025] The model parameters of the diffusion probability model are adjusted based on the error between the future trajectory and the actual trajectory.

[0026] Preferably, the step of associating the target detection results from each of the sensors to form the historical trajectories of traffic participants other than the vehicle to be planned includes:

[0027] The target detection results from each of the sensors are spatiotemporally aligned and feature matched to form the historical trajectories of traffic participants other than the vehicle to be planned.

[0028] A second aspect of this invention discloses a driving trajectory planning device, the device comprising:

[0029] The data collection module is used to collect the historical trajectories of traffic participants other than the vehicle to be planned.

[0030] The prediction module is used to predict the future trajectory of the traffic participant based on the historical trajectory.

[0031] The planning module is used to perform distributed robust stochastic model predictive control using the future trajectory to plan the optimal trajectory of the vehicle to be planned.

[0032] A third aspect of the present invention discloses an electronic device, comprising: a processor and a memory, the processor and the memory being connected via a bus; wherein, the processor is used to call and execute a program stored in the memory; the memory is used to store the program, the program being used to implement the driving trajectory planning method disclosed in the first aspect of the present invention.

[0033] A fourth aspect of the present invention discloses a storage medium storing computer-executable instructions for executing the driving trajectory planning method disclosed in the first aspect of the present invention.

[0034] Based on the above embodiments of the present invention, a method, apparatus, electronic device, and storage medium for planning driving trajectories are provided. The method involves: collecting historical trajectories of traffic participants other than the vehicle to be planned; predicting the future trajectories of traffic participants using the historical trajectories; and using the future trajectories to perform distributed robust stochastic model predictive control to plan the optimal trajectory of the vehicle to be planned. In this scheme, the historical trajectories of traffic participants other than the vehicle to be planned are used to predict the future trajectories of traffic participants, and then distributed robust stochastic model predictive control is performed using the future trajectories of traffic participants to plan the optimal trajectory of the vehicle to be planned. The planning of driving trajectories fully considers the future trajectories of other traffic participants, thereby improving the planning effect of driving trajectories. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0036] Figure 1 Example diagram of a vehicle model provided in an embodiment of the present invention;

[0037] Figure 2 A flowchart illustrating a driving trajectory planning method provided in an embodiment of the present invention;

[0038] Figure 3 A flowchart for collecting historical trajectories provided in an embodiment of the present invention;

[0039] Figure 4 A schematic diagram illustrating the overall principle architecture of a driving trajectory planning method provided in an embodiment of the present invention;

[0040] Figure 5 This is a structural block diagram of a driving trajectory planning device provided in an embodiment of the present invention. Detailed Implementation

[0041] 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.

[0042] In this application, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0043] With the development of autonomous driving technology, it has gradually become common in daily driving. The implementation of autonomous driving technology requires the planning of driving trajectories. Currently, common methods for driving trajectory planning include: Deterministic Model Predictive Control (MPC), Robust Model Predictive Control (RMPC), and deep learning methods.

[0044] Among them, the MPC method requires the assumption that the driving environment is known and certain, and cannot handle uncertainties such as sensor noise and sudden changes in other vehicle driving intentions, which may lead to overly conservative or risky decisions in complex traffic scenarios.

[0045] RMPC optimizes for the worst-case scenario, which can improve system robustness, but its overly conservative optimization strategy can lead to a significant decrease in average vehicle speed and traffic efficiency.

[0046] Deep learning methods mainly include two approaches: "end-to-end imitation learning" and "deep reinforcement learning." However, "end-to-end imitation learning" lacks clear safety constraints and makes it difficult to ensure driving safety. "Deep reinforcement learning" has low sample efficiency, an unstable training process, and lacks interpretability and safety boundary verification.

[0047] In summary, existing driving trajectory planning methods are not suitable for handling highly dynamic and uncertain traffic scenarios, and the planning effect of driving trajectories is poor.

[0048] To achieve real-time trajectory planning while ensuring safety, this invention proposes a method, device, electronic device, and storage medium for planning driving trajectories. It utilizes the historical trajectories of traffic participants other than the vehicle to be planned to predict their future trajectories. Then, it uses the future trajectories of the traffic participants for distributed robust stochastic model predictive control, thereby planning the optimal trajectory for the vehicle to be planned. By fully considering the future trajectories of other traffic participants during the planning of driving trajectories, the planning effect of driving trajectories is improved.

[0049] It is worth noting that, before performing vehicle trajectory planning, this scheme requires the pre-construction of a 14-DOF vehicle dynamics model to accurately describe the vehicle's dynamic behavior. The following section combines... Figure 1 The example diagram of the vehicle model shown will first provide a detailed explanation of the fourteen-degree-of-freedom vehicle dynamics model.

[0050] The 14-DOF vehicle dynamics model includes the following degrees of freedom: longitudinal motion of the vehicle body (v... x ), the lateral movement of the vehicle body (v y ), the yaw motion of the vehicle body (r), the rolling angular velocity of the four wheels ( ), vertical movement of the four suspension systems ( ), the vertical speed of the four suspension systems ( ).

[0051] It should be noted that, in the above content regarding degrees of freedom, " "These represent the rolling angular velocities of the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively." "These respectively represent the vertical movement of the left front suspension system, right front suspension system, left rear suspension system, and right rear suspension system." The symbols represent the vertical velocities of the left front suspension system, right front suspension system, left rear suspension system, and right rear suspension system, respectively. The dot (·) above the parameter indicates the time derivative. The subscripts... , , and These represent left front, right front, left back, and right back, respectively.

[0052] Therefore, the mathematical representation of the fourteen-degree-of-freedom vehicle dynamics model is shown in formulas (1) to (14).

[0053] (1);

[0054] (2);

[0055] (3);

[0056] (4);

[0057] (5);

[0058] (6);

[0059] (7);

[0060] (8);

[0061] (9);

[0062] (10);

[0063] (11);

[0064] (12);

[0065] (13);

[0066] (14);

[0067] The above fourteen-degree-of-freedom vehicle dynamics model is expressed in a continuous-time state-space form as shown in formula (15).

[0068] (15);

[0069] In formulas (1)-(15), x and y are the positions of the vehicle's center of mass in the global coordinate system, and θ is the heading angle / vehicle orientation. , , and These are driving force, braking force, front wheel steering angle, and rear wheel steering angle, respectively. To synthesize driving resistance (including rolling resistance, air resistance, etc.); and These are the lateral forces on the front axle and the lateral forces on the rear axle, respectively. , , and These represent the drive / braking torques of the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively. and These are the vertical support forces for the front and rear axles, respectively.

[0070] , and , respectively, are the equivalent masses of the whole vehicle, the front axle, and the rear axle; a and b are the distances from the center of mass to the front axle and the rear axle, respectively; Let the moment of inertia of the entire vehicle about its vertical axis be _____. and denoted as the equivalent moments of inertia of the front and rear wheels, respectively, and g is the acceleration due to gravity.

[0071] It is understandable that the state mode shown in formula (16) and the output mode shown in formula (17) represent the fourteen-degree-of-freedom vehicle dynamics model as a specific mathematical expression in the form of continuous-time state space.

[0072] (16);

[0073] (17);

[0074] In formulas (16) and (17), A c B is a continuous-time state matrix. c Let be the continuous-time input matrix, u be the unloaded input vector, C be the output matrix, D be the feedforward matrix, and w(t) be the random perturbation.

[0075] To transform the continuous-time vehicle dynamics model into a discrete-time model, the aforementioned state equations and output equations need to be discretized. Assuming the sampling time is ΔT, the discretized state-space model can be expressed as equations (18) and (19).

[0076] (18);

[0077] (19);

[0078] In formulas (18) and (19), A and B are discretization matrices, x k+1 Let x be the state vector of the vehicle at the current discrete time point k+1. k Let k be the state vector of the vehicle at the current discrete time point k. To apply the vehicle's control input vector at time point k, For random disturbances in the system, y k Let k be the system output vector at time point k.

[0079] Discretized matrices A and B can be calculated using formula (20).

[0080] (20);

[0081] in, It is the integral variable.

[0082] The discretized state-space model obtained from the above content is then applied to the optimization process of Distributed Robust Stochastic Model Predictive Control (DRSMPC) to plan the optimal trajectory. The following examples will explain how to plan the optimal trajectory.

[0083] See Figure 2 The flowchart illustrates a driving trajectory planning method provided by an embodiment of the present invention. The planning method includes:

[0084] Step S201: Collect the historical trajectories of traffic participants other than the vehicle to be planned.

[0085] In the specific implementation step S201, environmental data is collected by various sensors (such as vehicle-mounted lidar, cameras and millimeter-wave radar) on the vehicle to be planned. This environmental data includes, but is not limited to, point cloud data, image data and speed information.

[0086] By collecting environmental data, the historical trajectories of each traffic participant other than the vehicle to be planned can be determined.

[0087] It should be noted that each traffic participant is either a pedestrian or a vehicle; when the traffic participant is a vehicle, the historical trajectory is the historical driving trajectory; when the traffic participant is a pedestrian, the historical trajectory is the historical movement trajectory.

[0088] Step S202: Predict the future trajectories of traffic participants using historical trajectories.

[0089] It should be noted that a graph structure is pre-constructed to represent the spatiotemporal relationships of traffic participants, and a spatiotemporal graph neural network (ST-GNN) is used to extract the spatiotemporal features of nodes and edges in the graph structure. Specifically, each traffic participant is considered as a node in the graph structure, and the relative position and speed difference between nodes represent the features of the edges. The spatiotemporal feature representation of each node is learned through the information passing and aggregation mechanism of a multi-layer GNN.

[0090] In the specific implementation step S202, the historical trajectories of each traffic participant other than the vehicle to be planned are input into a preset spatiotemporal graph neural network to obtain the spatiotemporal feature representation of the traffic participants.

[0091] The spatiotemporal characteristics and historical trajectories of all traffic participants other than the vehicle to be planned are input into a pre-defined diffusion probability model (Diffusion Trajectory Decoder) to predict the future trajectories of the traffic participants.

[0092] Specifically, the spatiotemporal characteristics and historical trajectories of each traffic participant are input into the diffusion probability model. The diffusion probability model generates diverse future trajectories that conform to physical laws and traffic rules through a process of gradually adding noise and denoising.

[0093] Using the above method, the future trajectories of all traffic participants other than the vehicle to be planned can be predicted. These future trajectories are equivalent to multiple high-likelihood scenario trajectory samples. The uncertainty of the predicted future trajectories is quantified, and the result of the uncertainty quantification is then applied to the optimization solution process of DRSMPC, thereby planning the optimal trajectory of the vehicle to be planned.

[0094] The uncertainty quantification is achieved by calculating the mean of multiple future trajectories generated by the diffusion probability model. Covariance The overall distribution is obtained by weighting and combining the results using formula (21).

[0095] (twenty one);

[0096] In formula (21), This represents the overall mean of the future trajectories after the merger. For the first k The weighting coefficients of the predicted trajectories, For the first k The mean of the predicted trajectories, Let covariance matrix be the variance matrix. Let I be the weighting coefficient for the future trajectory, and let I be the identity matrix.

[0097] The overall distribution calculated by formula (21) provides environmental uncertainty input for the optimization solution process of DRSMPC.

[0098] It should be noted that the spatiotemporal graph neural network and diffusion probability model used above are trained using the following model training strategy: large-scale real traffic scene data is collected in advance, including various traffic conditions, weather conditions and road types; the real traffic scene data is labeled, including vehicle type, location, speed, acceleration and pedestrian trajectory, etc.

[0099] Based on the large-scale real traffic scene data collected, the spatiotemporal graph neural network and the diffusion probability model are pre-trained, enabling the spatiotemporal graph neural network and the diffusion probability model to learn the general behavioral patterns of traffic participants.

[0100] In addition, new environmental data is collected in real time during autonomous driving, and the spatiotemporal graph neural network and diffusion probability model are fine-tuned online using the new environmental data, so that the spatiotemporal graph neural network and diffusion probability model can continuously adapt to new traffic scenarios and driving habits.

[0101] Step S203: Utilize future trajectories for distributed robust stochastic model predictive control to plan the optimal trajectory of the vehicle to be planned.

[0102] In the specific implementation step S203, the future trajectories of each traffic participant are used to perform distributed robust stochastic model predictive control, thereby planning the optimal trajectory (i.e. the optimal driving trajectory) of the vehicle to be planned.

[0103] Specifically, the future trajectories of different traffic participants are assigned to different computing units (e.g., GPUs). The computing units then perform optimization solutions based on the assigned future trajectories, the objective function of the distributed robust stochastic model predictive control, and the constraints. The constraints include chance constraints and control constraints. Each computing unit exchanges boundary information every preset time interval (e.g., 5 milliseconds) during the optimization process.

[0104] For example, each computing unit is assigned a portion of the future trajectory; the computing unit performs optimization based on the assigned future trajectory, the objective function, and the constraints.

[0105] By aggregating the optimization results of each computing unit, the optimal trajectory of the vehicle to be planned is obtained.

[0106] It should be noted that the optimization objective function and constraints for distributed robust stochastic model predictive control will be explained below.

[0107] Optimization objective function: The optimization objective function is defined by comprehensively considering path tracking error, control input energy consumption and risk value (CVaR). The specific content of the optimization objective function is shown in formula (22).

[0108] (twenty two);

[0109] In formula (22), Q, R, Q f This is the weight matrix. For risk weighting coefficients, The confidence level for Value at Risk (CVaR) is given by N, where N is the prediction time domain length.

[0110] Opportunity constraint: Transform the safety constraint into an opportunity constraint, allowing the constraint to be violated with a very small probability, thereby balancing safety and passage efficiency. The specific content of the opportunity constraint is shown in formula (23).

[0111] (twenty three);

[0112] In formula (23), n k For the safety constraint normal vector, r c,k For the corresponding safety boundary / reference point, and For the linear lane constraint matrix and boundary, and P represents the upper limit of the probability of default, and P is the probability operator.

[0113] Control constraints: Limit the amplitude of control inputs to ensure the physical feasibility and safety of the vehicle dynamics system. The specific contents of control constraints are shown in formula (24).

[0114] (twenty four);

[0115] in, To control the upper limit of input.

[0116] It should be noted that in practical applications, the opportunity constraint can be transformed into an equivalent second-order cone constraint by using the sub-Bru bar optimization theory, thereby transforming the original optimization problem into a second-order cone programming (SOCP) problem that can be solved efficiently as shown in formula (25).

[0117] (25);

[0118] In formula (25), and for The mean and covariance, It is a 2-norm. The normal vector of the safety boundary. To constrain the default limit for opportunities.

[0119] When planning the optimal trajectory, all generated future trajectories are distributed to different computing units for parallel computation, with each computing unit assigned a portion of the future trajectories. The computing units perform optimization solutions using the Alternating Direction Multiplier Method (ADMM), exchanging boundary information every preset time interval (e.g., 5 milliseconds). The optimization results from each computing unit are aggregated by the main thread to obtain the optimal trajectory for the vehicle to be planned.

[0120] It should be noted that during the optimization process of distributed robust stochastic model predictive control, the optimization objective function guides the optimal performance, chance constraints ensure probabilistic safety, and control constraints ensure physical feasibility.

[0121] The boundary information exchanged between computational units consists of key variables that are exchanged to maintain consistency. For example, boundary information includes key variables used to coordinate global convergence, such as state continuity, ADMM multipliers, and residuals.

[0122] Understandably, in practical applications, in order to continuously improve the accuracy of trajectory planning, this solution proposes an online meta-learning and adaptive mechanism to update the model parameters of the diffusion probability model.

[0123] In some embodiments, the actual trajectories of traffic participants are collected; the model parameters of the diffusion probability model are adjusted based on the error between the future trajectory and the actual trajectory.

[0124] Specifically, the online meta-learning and adaptive mechanism includes the following parts: the goal and principle part, the triggering part, the data part, the adaptation part, and the deployment and scheduling part.

[0125] Objective: To enable the diffusion probability model to adapt quickly to new environments without lowering the control layer safety threshold, thereby improving the accuracy and calibration of future trajectory distribution predictions and reducing the conservative shrinkage of DRSMPC.

[0126] Principle: Only fine-tune a small number of model parameters or adapter modules at the end of the diffusion probability model, and maintain the control layer safety baseline (e.g., chance constraint default cap). CVaR confidence level The restrictions will not be relaxed.

[0127] Trigger: Real-time calculation of sliding window ADE and NLL; fine-tuning is triggered when either indicator exceeds the threshold for K consecutive periods, followed by a cooldown period T. c The entire process is shown in formulas (26) to (28).

[0128] (26);

[0129] (27);

[0130] (28);

[0131] In formulas (26) to (28), ADE represents the predicted future trajectory. and the actual trajectory The average Euclidean distance error over the entire prediction time domain T. Future trajectory The predicted position (two-dimensional coordinates) at time step t is equivalent to the actual trajectory. This is equivalent to the actual observed position at time step t. This means predicting the current position given historical information H (e.g., the trajectory over the past few seconds). The probability density.

[0132] For negative log-likelihood loss, NLL is negative log-likelihood.

[0133] The average ADE over a sliding time window, The average NLL within the same window, and This is the preset threshold.

[0134] Data: Includes a support set S and a query set Q, where the support set S is the most recent set T. s The data is in seconds, and the query set Q is the result of T. q Data in seconds (to avoid time overlap); in addition, the sample weight can be increased for high-risk segments.

[0135] Adaptation: Only the last L layers of the model are updated; the backbone structure is frozen. Additionally, the timeout budget is set to T. budget If the timeout occurs, the program will immediately terminate and exit.

[0136] Deployment and Scheduling: If the fine-tuning passes, it will take effect in the next trajectory planning cycle. If the fine-tuning conflicts with the critical synchronization phase, it will be postponed for one trajectory planning cycle before taking effect. Cooldown time T after fine-tuning. c When computing power is tight, skip this round of fine-tuning.

[0137] It should be noted that the spatiotemporal graph neural network and the diffusion probability model constitute the AI ​​structure. During the driving process of the vehicle to be planned, the trajectory planning of AI-DRSMPC is executed cyclically at a fixed high frequency period. The trajectory planning period can be set according to the actual situation. For example, the trajectory planning period can be set to 50 milliseconds.

[0138] It should be further noted that the execution cycle of trajectory planning can be set from aspects such as vehicle dynamics, environmental interaction requirements, system integration matching, and computational feasibility.

[0139] Vehicle dynamics: The trajectory planning cycle needs to accurately capture and control the vehicle's main lateral and longitudinal dynamic responses to ensure the smoothness and stability of trajectory tracking.

[0140] Environmental interaction requirements: When facing highly dynamic traffic scenarios, the trajectory planning cycle needs to provide the system with sufficient reaction speed and safety margin. For example, when a vehicle is traveling at 60 km / h, the displacement of other vehicles within 50 milliseconds is approximately 0.83 meters, which is within a controllable and plannable range.

[0141] System integration matching: The cycle of the execution trajectory planning is matched with the update frequency of the upstream sensing module (usually 10-20Hz) and the control bandwidth of the downstream wire actuator, which facilitates system integration.

[0142] Computational feasibility: The distributed parallel architecture proposed in this solution ensures that the continuous single-cycle computation time is lower than the "cycle of execution trajectory planning", meets the hard real-time constraint, and leaves a margin to cope with computational jitter.

[0143] To maximize the use of computing resources, at the beginning of each trajectory planning cycle, the latest environmental perception results are obtained and trajectory prediction and optimization are performed accordingly. The prediction step size used for optimization can be set to N=60 (for example only), which corresponds to a prediction time domain of 3 seconds.

[0144] After obtaining the optimal trajectory in each trajectory planning cycle, the optimal trajectory is transmitted to the vehicle chassis controller via the CAN bus. At the same time, a state observer is constructed using an extended Kalman filter (EKF) to perform multi-source information fusion and online error compensation on the outputs of vehicle positioning, inertial measurement, and sensing modules, effectively suppressing noise and deviation, and ultimately improving the overall accuracy and robustness of trajectory tracking.

[0145] Regarding the above embodiments of the present invention Figure 2 The content of "collecting historical trajectories of traffic participants other than the vehicle to be planned" involved in step S201, see [link to relevant documentation]. Figure 3 This illustrates a flowchart of the historical trajectory acquisition process provided in an embodiment of the present invention. Figure 3 Includes the following steps:

[0146] Step S301: Acquire environmental data collected by various sensors.

[0147] In the specific implementation step S301, environmental data is collected by various sensors (such as vehicle-mounted lidar, cameras and millimeter-wave radar) on the vehicle to be planned. This environmental data includes, but is not limited to, point cloud data, image data and speed information.

[0148] Step S302: Preprocess the environmental data.

[0149] In the specific implementation of step S302, the environmental data collected by each sensor is preprocessed.

[0150] Preprocessing includes filtering, denoising, and feature extraction. For example, during the preprocessing of this environmental data, lane lines and traffic signs are extracted using image segmentation algorithms, and traffic participants such as vehicles and pedestrians are identified using point cloud clustering algorithms.

[0151] Step S303: Perform target tracking using preprocessed environmental data to obtain the historical trajectories of traffic participants other than the vehicle to be planned.

[0152] In the specific implementation of step S303, target tracking is performed using preprocessed environmental data to obtain target detection results from various sensors.

[0153] Specifically, for each sensor, based on the preprocessed environmental data acquired by that sensor, a Kalman filter or extended Kalman filter algorithm is used to track traffic participants, thereby obtaining the target detection result from that sensor. The target detection result includes state information such as the position, velocity, and acceleration of the traffic participants. The target detection results from each sensor can be obtained through the aforementioned method.

[0154] The target detection results from various sensors are correlated to form the historical trajectories of traffic participants other than the vehicle to be planned.

[0155] Specifically, the method for associating target detection results from various sensors is as follows: the target detection results from various sensors are spatiotemporally aligned and feature matched to form the historical trajectories of traffic participants other than the vehicle to be planned.

[0156] above Figure 3 This is an explanation of how to collect the historical trajectories of traffic participants.

[0157] Based on the above embodiments, the overall principle architecture of the driving trajectory planning method proposed in this invention is shown in the following diagram. Figure 4 As shown, the architecture diagram includes the following parts: perception part, deep learning part, DRSMPC part, closed-loop execution part, and online meta-learning and adaptation part.

[0158] Perception section: Collects environmental data and transmits the environmental data to the deep learning section.

[0159] Deep learning component: Predicts the future trajectories of all traffic participants other than the vehicle to be planned, and transmits the predicted future trajectories to the DRSMPC component.

[0160] The DRSMPC section plans the optimal trajectory of the vehicle to be planned and transmits the optimal trajectory to the closed-loop execution section.

[0161] The closed-loop execution section consists of a control module, a vehicle chassis actuator, and feedback compensation. The control module generates control commands based on the optimal trajectory and transmits them to the vehicle chassis controller for execution. The vehicle chassis controller transmits information such as errors and actual status to the feedback compensation, which then feeds back the errors to the control module, thus forming a closed-loop execution.

[0162] Online meta-learning and adaptive part: Update the model parameters of the diffusion probability model based on the error feedback from the closed-loop execution part.

[0163] In summary, this solution integrates AI, DRSMPC, and online meta-learning for trajectory planning, and has the following beneficial effects:

[0164] Enhanced safety: By explicitly addressing multiple uncertainties in the traffic environment through the decomposed bar optimization framework, provable boundaries are provided for safety constraints, thereby systematically improving the safety and reliability of trajectory planning.

[0165] Improve traffic efficiency: By adopting stochastic optimization instead of traditional robust methods, the shortcomings of overly conservative planning are effectively overcome, thereby significantly improving average vehicle speed and overall traffic efficiency while ensuring safety.

[0166] High real-time performance: It adopts a distributed parallel computing architecture, which has extremely high single-cycle computing efficiency, meeting the high-frequency decision-making and real-time control requirements of autonomous driving.

[0167] Long-term robustness: Through online meta-learning and adaptation, it continuously adapts to environmental changes, maintains long-term stability and robustness, and ensures that it maintains high performance throughout long-term operation.

[0168] Corresponding to the driving trajectory planning method provided in the above embodiments of the present invention, see also... Figure 5 The present invention also provides a structural block diagram of a driving trajectory planning device, which includes: a data acquisition module 501, a prediction module 502, and a planning module 503.

[0169] The data acquisition module 501 is used to collect the historical trajectories of traffic participants other than the vehicle to be planned.

[0170] The prediction module 502 is used to predict the future trajectory of traffic participants based on historical trajectories.

[0171] In a specific implementation, the prediction module 502 is specifically used to: input the historical trajectory into a preset spatiotemporal graph neural network for feature extraction to obtain the spatiotemporal feature representation of the traffic participant; input the spatiotemporal feature representation and the historical trajectory into a preset diffusion probability model for trajectory prediction to obtain the future trajectory of the traffic participant.

[0172] Planning module 503 is used to perform distributed robust stochastic model predictive control using future trajectories in order to plan the optimal trajectory of the vehicle to be planned.

[0173] In its specific implementation, the planning module 503 is used to: assign the future trajectories of different traffic participants to different computing units, and enable the computing units to perform optimization solutions based on the assigned future trajectories, the optimization objective function of the distributed robust stochastic model predictive control, and the constraints. The constraints include chance constraints and control constraints. Each computing unit exchanges boundary information every preset time interval during the optimization solution process. The optimization solution results of each computing unit are aggregated to obtain the optimal trajectory of the vehicle to be planned.

[0174] Preferred, combined Figure 5 As shown, the acquisition module 501 includes an acquisition submodule, a processing submodule, and a tracking submodule. The execution principle of each submodule is as follows:

[0175] The acquisition submodule is used to acquire environmental data collected by various sensors.

[0176] The processing submodule is used to preprocess environmental data.

[0177] The tracking submodule is used to track targets using preprocessed environmental data to obtain the historical trajectories of traffic participants other than the vehicle to be planned.

[0178] In its specific implementation, the tracking submodule is used to: track targets using preprocessed environmental data and obtain target detection results from various sensors; and correlate the target detection results from various sensors to form the historical trajectories of traffic participants other than the vehicle to be planned.

[0179] This involves associating target detection results from various sensors to form the historical trajectories of traffic participants other than the vehicle to be planned. This includes: performing spatiotemporal alignment and feature matching on the target detection results from various sensors to form the historical trajectories of traffic participants other than the vehicle to be planned.

[0180] Preferred, combined Figure 5 The planning device, as shown, also includes:

[0181] The update module is used to collect the actual trajectories of traffic participants; the model parameters of the diffusion probability model are adjusted based on the error between the future trajectory and the actual trajectory.

[0182] Preferably, the present invention also provides an electronic device, including: a processor and a memory, the processor and the memory being connected via a bus; wherein, the processor is used to call and execute a program stored in the memory; the memory is used to store the program, the program being used to implement the driving trajectory planning method provided in the above method embodiments.

[0183] Preferably, the present invention also provides a storage medium storing computer-executable instructions for executing the driving trajectory planning method provided in the above method embodiments.

[0184] In summary, the embodiments of the present invention provide a method, apparatus, electronic device, and storage medium for planning driving trajectories. The method uses the historical trajectories of traffic participants other than the vehicle to be planned to predict the future trajectories of traffic participants, and then uses the future trajectories of traffic participants to perform distributed robust stochastic model predictive control, thereby planning the optimal trajectory of the vehicle to be planned. The method fully considers the future trajectories of other traffic participants when planning driving trajectories, thereby improving the planning effect of driving trajectories.

[0185] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0186] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0187] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for planning driving trajectories, characterized in that, The method includes: Collect historical trajectories of traffic participants other than the vehicles to be planned; The future trajectories of the traffic participants are predicted using the historical trajectories. The future trajectory is used for distributed robust stochastic model predictive control to plan the optimal trajectory of the vehicle to be planned.

2. The method according to claim 1, characterized in that, The collection of historical trajectories of traffic participants other than the vehicle to be planned includes: Acquire environmental data collected by various sensors; The environmental data is preprocessed; Target tracking is performed using the preprocessed environmental data to obtain the historical trajectories of traffic participants other than the vehicle to be planned.

3. The method according to claim 1, characterized in that, The method of predicting the future trajectory of the traffic participant through the historical trajectory includes: The historical trajectory is input into a preset spatiotemporal graph neural network for feature extraction to obtain the spatiotemporal feature representation of the traffic participants; The spatiotemporal feature representation and the historical trajectory are input into a preset diffusion probability model to predict the trajectory, so as to obtain the future trajectory of the traffic participant.

4. The method according to claim 1, characterized in that, The step of using the future trajectory to perform distributed robust stochastic model predictive control to plan the optimal trajectory of the vehicle to be planned includes: The future trajectories of different traffic participants are assigned to different computing units, and the computing units perform optimization solutions based on the assigned future trajectories, the optimization objective function of the distributed robust stochastic model predictive control, and the constraints. The constraints include chance constraints and control constraints. Each computing unit exchanges boundary information once every preset time interval during the optimization solution process. The optimal trajectory of the vehicle to be planned is obtained by aggregating the optimization results of each computing unit.

5. The method according to claim 2, characterized in that, The step of tracking the target using the preprocessed environmental data to obtain the historical trajectories of traffic participants other than the vehicle to be planned includes: Target tracking is performed using the preprocessed environmental data to obtain target detection results from each of the sensors; The target detection results from each of the aforementioned sensors are correlated to form the historical trajectories of traffic participants other than the vehicle to be planned.

6. The method according to claim 3, characterized in that, After obtaining the future trajectories of the traffic participants, the following is also included: Collect the real trajectories of the traffic participants; The model parameters of the diffusion probability model are adjusted based on the error between the future trajectory and the actual trajectory.

7. The method according to claim 5, characterized in that, The step of associating the target detection results from each of the sensors to form the historical trajectories of traffic participants other than the vehicle to be planned includes: The target detection results from each of the sensors are spatiotemporally aligned and feature matched to form the historical trajectories of traffic participants other than the vehicle to be planned.

8. A driving trajectory planning device, characterized in that, The device includes: The data collection module is used to collect the historical trajectories of traffic participants other than the vehicle to be planned. The prediction module is used to predict the future trajectory of the traffic participant based on the historical trajectory. The planning module is used to perform distributed robust stochastic model predictive control using the future trajectory to plan the optimal trajectory of the vehicle to be planned.

9. An electronic device, characterized in that, include: A processor and a memory are connected via a bus; wherein the processor is used to call and execute a program stored in the memory; The memory is used to store a program for implementing the driving trajectory planning method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores computer-executable instructions for executing the driving trajectory planning method as described in any one of claims 1-7.

Images