Track generation method and device, vehicle and program product

Abstract

The embodiment of the invention provides a track generation method and device, a vehicle and a program product. The method comprises the following steps: acquiring a real-time environment sensing result and a current driving state of a vehicle; according to the real-time environment sensing result, a control quantity sequence of the vehicle is obtained through prediction; wherein the control quantity sequence comprises the accelerated speed and the steering angle of the vehicle at T moments in the future time period; t is an integer greater than or equal to 2; and performing trajectory recursion processing on the current driving state and the control quantity sequence of the vehicle to obtain a driving trajectory of the vehicle in a future time period. According to the method, the track meeting the environmental adaptability and the motion feasibility at the same time can be generated in a complex dynamic scene, and the safety and the performability of a track planning result are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of intelligent driving technology, and in particular to a trajectory generation method, device, vehicle, and program product. Background Technology

[0002] In scenarios involving dense traffic at urban road intersections, obstacle avoidance in narrow parking spaces, and obstacle avoidance on highways, vehicles must perceive environmental obstacles, assess dynamic risks, and generate executable driving trajectories within a limited time window to ensure driving safety and passenger comfort.

[0003] In existing technologies, driving trajectories are generated by explicitly embedding vehicle kinematic constraints through manual rules or optimal control theory based on search algorithms.

[0004] However, the above methods rely on preset grid maps or simplified obstacle models. In complex and dynamic scenes, due to weak environmental modeling capabilities and insufficient representation of complex scenes, trajectory planning fails. Summary of the Invention

[0005] This application provides a trajectory generation method, device, vehicle, and program product, which can generate trajectories that simultaneously meet environmental adaptability and motion feasibility in complex dynamic scenarios, significantly improving the safety and executability of trajectory planning results.

[0006] Firstly, this application provides a trajectory generation method, including:

[0007] Obtain the vehicle's real-time environmental perception results and current driving status;

[0008] Based on the real-time environmental perception results, a control quantity sequence for the vehicle is predicted; wherein, the control quantity sequence includes the vehicle's acceleration and steering angle at T time points within a future time period; T is an integer greater than or equal to 2;

[0009] The current driving state and control quantity sequence of the vehicle are processed by trajectory recursion to obtain the driving trajectory of the vehicle in the future time period.

[0010] In one possible implementation, predicting the vehicle's control quantity sequence based on the real-time environmental perception results includes:

[0011] Based on the real-time environmental perception results, the initial control quantity sequence of the vehicle is predicted;

[0012] Based on the physical constraints, the initial control quantity sequence is mapped to obtain the control quantity sequence of the vehicle.

[0013] In one possible implementation, predicting the initial control quantity sequence of the vehicle based on the real-time environmental perception results includes:

[0014] Feature extraction is performed on the real-time environment perception results to obtain the spatiotemporal features of the real-time environment perception results;

[0015] The spatiotemporal features are processed to obtain the initial control quantity sequence of the vehicle.

[0016] In one possible implementation, the process of recursively processing the current driving state and control quantity sequence of the vehicle to obtain the driving trajectory of the vehicle in a future time period includes:

[0017] Let t=1, and repeat the following process until t=T: Based on the vehicle's driving state at time t and the acceleration and steering angle at time t in the control quantity sequence, obtain the vehicle's driving state at time t+1; let t=t+1;

[0018] Based on the vehicle's driving status at each moment, the vehicle's driving trajectory for a future time period is generated.

[0019] In one possible implementation, the driving state includes driving speed, vehicle orientation, and coordinate information of the corresponding trajectory points in a preset coordinate system; obtaining the driving state of the vehicle at time t+1 based on the driving state of the vehicle at time t and the acceleration and steering angle at time t in the control quantity sequence includes:

[0020] The vehicle's speed at time t+1 is obtained based on the vehicle's speed at time t and the acceleration at time t in the control quantity sequence.

[0021] Based on the vehicle's speed and orientation at time t, and the steering angle at time t in the control quantity sequence, the vehicle's orientation at time t+1 is obtained.

[0022] Based on the vehicle's speed, orientation, and coordinates at time t, the coordinates of the vehicle at time t+1 are obtained.

[0023] In one possible implementation, the method further includes:

[0024] Acquire historical environmental perception results and historical driving status information of the vehicle;

[0025] Based on the initial model, the historical environmental perception results are processed to obtain a predictive control sequence for the vehicle; wherein, the control sequence includes the vehicle's acceleration and steering angle at N moments within a preset time period; N is an integer greater than or equal to 2;

[0026] The historical driving status information and predictive control quantity sequence of the vehicle are processed by trajectory recursion to obtain the predicted driving trajectory of the vehicle within a preset time period.

[0027] Based on the predicted driving trajectory of the vehicle within a preset time period, the initial model is trained to obtain a control quantity prediction network; wherein, the control quantity prediction network is used to implement the step of predicting the control quantity sequence of the vehicle based on the real-time environmental perception results.

[0028] In one possible implementation, training the initial model based on the predicted driving trajectory of the vehicle within a preset time period to obtain a control quantity prediction network includes:

[0029] Based on the vehicle's driving trajectory and actual trajectory within a preset time period, a multi-objective loss function is determined; wherein, the multi-objective loss function includes trajectory prediction loss, collision loss, lane centering loss, and comfort loss;

[0030] The initial model is optimized based on the multi-objective loss function to obtain the control quantity prediction network.

[0031] Secondly, this application provides a trajectory generation apparatus, comprising:

[0032] The acquisition module is used to acquire the vehicle's real-time environmental perception results and current driving status;

[0033] The prediction module is used to predict the control quantity sequence of the vehicle based on the real-time environmental perception results; wherein, the control quantity sequence includes the vehicle's acceleration and steering angle at T time points in the future time period; T is an integer greater than or equal to 2;

[0034] The recursive module is used to perform trajectory recursive processing on the current driving state and control quantity sequence of the vehicle to obtain the driving trajectory of the vehicle in a future time period.

[0035] In one possible implementation, the prediction module is specifically used to: predict the initial control quantity sequence of the vehicle based on the real-time environmental perception results; and map the initial control quantity sequence based on the physical constraint range to obtain the control quantity sequence of the vehicle.

[0036] In one possible implementation, the prediction module is specifically used to: extract features from the real-time environment perception results to obtain the spatiotemporal features of the real-time environment perception results; process the spatiotemporal features to obtain the initial control quantity sequence of the vehicle.

[0037] In one possible implementation, the recursive module is specifically used to: let t=1, and repeat the following process until t=T: obtain the driving state of the vehicle at time t+1 based on the driving state of the vehicle at time t and the acceleration and steering angle at time t in the control quantity sequence; let t=t+1; generate the driving trajectory of the vehicle in the future time period based on the driving state of the vehicle at each time.

[0038] In one possible implementation, the driving state includes driving speed, vehicle orientation, and coordinate information of the corresponding trajectory point in a preset coordinate system; the recursive module is specifically used to: obtain the driving speed of the vehicle at time t+1 based on the driving speed of the vehicle at time t and the acceleration at time t in the control quantity sequence; obtain the vehicle orientation at time t+1 based on the driving speed and vehicle orientation of the vehicle at time t and the steering angle at time t in the control quantity sequence; and obtain the coordinate information of the vehicle at time t+1 based on the driving speed, vehicle orientation, and coordinate information of the vehicle at time t.

[0039] In one possible implementation, the device is further configured to: acquire historical environmental perception results and historical driving state information of the vehicle; perform predictive processing on the historical environmental perception results based on an initial model to obtain a predictive control quantity sequence of the vehicle; wherein the control quantity sequence includes the vehicle's acceleration and steering angle at N moments within a preset time period; N is an integer greater than or equal to 2; perform trajectory recursive processing on the vehicle's historical driving state information and predictive control quantity sequence to obtain the vehicle's predicted driving trajectory within the preset time period; and train the initial model based on the vehicle's predicted driving trajectory within the preset time period to obtain a control quantity prediction network; wherein the control quantity prediction network is used to implement the step of predicting the vehicle's control quantity sequence based on the real-time environmental perception results.

[0040] In one possible implementation, the device is further configured to: determine a multi-objective loss function based on the vehicle's driving trajectory and actual trajectory within a preset time period; wherein the multi-objective loss function includes trajectory prediction loss, collision loss, lane centering loss, and comfort loss; and optimize the initial model based on the multi-objective loss function to obtain the control quantity prediction network.

[0041] Thirdly, this application provides an electronic device, including: a memory and a processor;

[0042] The memory stores computer-executed instructions;

[0043] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0044] Fourthly, embodiments of this application provide a vehicle, the vehicle including a vehicle body and electronic equipment as described in the third aspect.

[0045] Fifthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0046] In a sixth aspect, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0047] The trajectory generation method, apparatus, device, vehicle, storage medium, and program product provided in this application process the real-time environmental perception results of the vehicle to predict the vehicle's control quantity sequence, including the vehicle's acceleration and steering angle at multiple moments in the future time period. Then, the predicted control quantity sequence and the vehicle's current driving state are processed to obtain the vehicle's driving trajectory in the future time period. Furthermore, based on the architecture of predicted control quantity and recursive trajectory, a trajectory that simultaneously satisfies environmental adaptability and motion feasibility can be generated in complex dynamic scenarios, significantly improving the safety and executability of trajectory planning results. Attached Figure Description

[0048] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0049] Figure 1 This application provides an illustration of an application scenario.

[0050] Figure 2 A schematic flowchart illustrating a trajectory generation method provided in an embodiment of this application;

[0051] Figure 3 A flowchart illustrating another trajectory generation method provided in this application embodiment;

[0052] Figure 4 A flowchart illustrating a deep learning-assisted driving trajectory planning scheme provided in an embodiment of this application;

[0053] Figure 5 This is a schematic diagram of the structure of a trajectory generation device provided in an embodiment of this application;

[0054] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0055] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0056] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0057] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, they do not violate public order and good morals, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0058] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on user rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0059] Figure 1 This application provides an illustration of an application scenario, such as... Figure 1As shown, this application is applicable to trajectory planning scenarios for Advanced Driving Assistance Systems (ADAS) and autonomous vehicles, especially for complex dynamic environments (such as dense traffic flow at urban intersections, narrow spaces in parking lots, and sudden obstacles on highways).

[0060] Based on the above scenarios, it is evident that rule-based trajectory search optimization, such as A* search algorithm, Rapidly-exploring Random Tree (RRT), and optimization methods like Model Predictive Control (MPC) and Iterative Linear Quadratic Regulator (iLQR), suffers from low computational efficiency and struggles to achieve human-like decision-making in complex scenarios. While deep learning-based solutions can enhance environmental perception, they often overlook kinematic constraints such as vehicle turning radius and acceleration limits, leading to feasibility flaws in the generated trajectories (e.g., sharp turns exceeding maximum curvature).

[0061] The trajectory generation method provided in this application combines environmental perception capabilities with vehicle kinematics through an architecture based on predictive control variables and recursive trajectories, enabling the generation of trajectories that simultaneously satisfy environmental adaptability and motion feasibility in complex dynamic scenarios.

[0062] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0063] Figure 2 This is a flowchart illustrating a trajectory generation method provided in an embodiment of this application, as shown below. Figure 2 As shown, the method includes:

[0064] 201. Obtain the vehicle's real-time environmental perception results and current driving status.

[0065] For example, the execution subject of this embodiment may be an electronic device, hereinafter referred to as the device. The device may be a device deployed in a vehicle or in the cloud. The device obtains the real-time environmental perception results of the vehicle through sensors (such as LiDAR, cameras, and millimeter-wave radar) in the vehicle, including the perception results and historical information of surrounding vehicles and pedestrians, as well as environmental information such as lane lines and dynamic and static obstacles in the bird's-eye view (BEV) raster encoding. The device also obtains the current driving status of the vehicle through a real-time monitoring module in the vehicle, including, for example, position, speed, heading angle, pose, and also states such as load and tire adhesion coefficient.

[0066] 202. Based on the real-time environmental perception results, predict the vehicle's control quantity sequence; wherein, the control quantity sequence includes the vehicle's acceleration and steering angle at T time points in the future time period; T is an integer greater than or equal to 2.

[0067] For example, the device performs feature extraction and prediction processing on the acquired real-time environmental perception results based on preset processing rules, such as deep learning networks (e.g., Transformer sequence modeling network based on self-attention mechanism), to obtain a sequence of vehicle control quantities, including the vehicle's acceleration and steering angle at T time points in the future time period; T is an integer greater than or equal to 2.

[0068] In one example, acceleration can be longitudinal acceleration. Steering angle can be steering wheel angle, front wheel angle, or other angle information proportional to the front wheel angle. The front wheel angle is the actual rotation angle of the left and right front wheels around the kingpin, and can be the left front wheel angle, the right front wheel angle, or the average front wheel angle.

[0069] For example, in urban intersection scenarios, based on the control quantity prediction network, the real-time positions of surrounding vehicles and pedestrian trajectories are extracted and predicted to obtain the longitudinal acceleration and front wheel steering angle control quantity sequence within the next 5 seconds.

[0070] 203. Perform trajectory recursion processing on the vehicle's current driving state and control quantity sequence to obtain the vehicle's driving trajectory in the future time period.

[0071] For example, the device can invoke a preset physical model, which is a mathematical model built based on the principles of vehicle kinematics or dynamics, to simulate the evolution of the vehicle's state. Based on this physical model, using the current driving state and the acceleration and steering angle at the current moment in the control quantity sequence as initial values, the trajectory recursion process is executed cyclically to obtain the vehicle's driving trajectory in the future time period, including the vehicle's state at the next T moments.

[0072] For example, the device can invoke a differentiable trajectory generator, which can be built based on a bicycle model or a two-degree-of-freedom vehicle dynamics model. The physical model can be selected according to the scenario requirements (e.g., the bicycle model is used for low-speed scenarios, and the two-degree-of-freedom model is used for high-speed scenarios). Based on this differentiable trajectory generator, using the current driving state and the acceleration and steering angle at the current moment in the control quantity sequence as initial values, the bicycle model recursion is executed T times in a loop to describe the vehicle's steering geometry and trajectory generation process, obtaining the vehicle's state quantity trajectory for the next 5 seconds.

[0073] In one example, during the recursive process of the bicycle model, a self-attention mechanism dynamically focuses on key time steps (such as areas with dense obstacles) in the future trajectory. Combined with attention weights, it performs fine-tuning of constricted trajectories (such as sharp turns) at key time steps while maintaining the smoothness of the overall trajectory. Dynamically focusing on key trajectory segments through the self-attention module improves the accuracy and comfort of trajectory planning, reduces global trajectory jitter caused by local optimization, thereby reducing riding discomfort and lowering the risk of collisions.

[0074] This embodiment provides a trajectory generation method based on a predictive control variable and a physical model recursive trajectory architecture. The deep learning network learns the perception features of complex scenes through sensor data to improve the dynamic obstacle perception capability. The physical model recursive ensures that the generated trajectory strictly follows vehicle kinematic constraints (such as turning radius and acceleration limit). This achieves unified optimization of environmental modeling capability and physical constraints in complex dynamic scenes, ensuring that the generated trajectory simultaneously meets environmental adaptability and motion feasibility, and significantly improving the safety and executability of trajectory planning results.

[0075] Figure 3 A flowchart illustrating another trajectory generation method provided in this application embodiment is shown below. Figure 3 As shown, the method includes:

[0076] 301. Obtain the vehicle's real-time environmental perception results and current driving status.

[0077] For example, this step can be referred to as step 201, which will not be repeated here.

[0078] 302. Based on the real-time environmental perception results, the initial control quantity sequence of the vehicle is predicted.

[0079] For example, the device invokes a preset control quantity prediction network, such as a hybrid network of Convolutional Neural Network (CNN) and Transformer, to perform feature extraction and prediction processing on the acquired real-time environmental perception results (such as BEV raster map and LiDAR point cloud) to obtain the vehicle's initial control quantity sequence, including the vehicle's initial acceleration and initial steering angle at T time points in the future time period.

[0080] In one example, by using coordinate system transformation and spatiotemporal alignment algorithms, LiDAR point clouds, camera images, millimeter-wave radar data, and BEV rasterized maps are mapped to a unified BEV coordinate system. A cross-modal attention mechanism is designed to dynamically weight the feature contributions of different modalities, generating a comprehensive environmental feature vector and obtaining the acceleration and steering angle control quantity sequence within 5 seconds. Through multimodal data fusion, the comprehensiveness and robustness of environmental perception are significantly improved.

[0081] Specifically, point cloud clustering algorithms can be used to process LiDAR point clouds to extract obstacle boundaries and spatial distribution features; semantic segmentation networks can be used to process images captured by cameras to identify static environmental information such as lane lines and traffic signs; and millimeter-wave radar data can be processed to extract the relative velocity and acceleration information of dynamic obstacles.

[0082] In one possible implementation, step 302 includes the following steps:

[0083] The first step is to extract features from the real-time environmental perception results to obtain the spatiotemporal features of the real-time environmental perception results.

[0084] The second step is to process the spatiotemporal characteristics to obtain the initial control quantity sequence for the vehicle.

[0085] Specifically, the device extracts features from real-time environmental perception results based on the backbone network of the control quantity prediction network. For example, a convolutional neural network extracts local obstacle distribution features from the BEV gridded map, and a self-attention module performs spatiotemporal modeling of dynamic obstacle trajectories to obtain the spatiotemporal features of the real-time environmental perception results. The extracted spatiotemporal features are then processed using a fully connected layer through the output head of the control quantity prediction network to generate initial control quantity sequences (such as acceleration and steering angle) for future time periods. For example, this includes the acceleration and steering wheel angle control quantity sequences corresponding to the trajectory in the next 5 seconds.

[0086] By employing a hybrid network structure combining convolutional neural networks and attention mechanisms, multi-scale extraction of environmental features and temporal modeling of dynamic obstacle trajectories are achieved. Specifically, the convolutional neural network accurately captures static environmental features (such as lane lines and obstacle distribution), while the attention mechanism dynamically focuses on key dynamic obstacles (such as vehicles making sharp turns), thereby enhancing environmental perception capabilities in complex scenarios.

[0087] In one possible implementation, prior to step 302, the following is included:

[0088] The first step is to obtain the vehicle's historical environmental perception results and historical driving status information.

[0089] The second step involves processing the historical environmental perception results based on the initial model to obtain a predictive control sequence for the vehicle. The control sequence includes the vehicle's acceleration and steering angle at N moments within a preset time period, where N is an integer greater than or equal to 2.

[0090] The third step involves performing trajectory recursion processing on the vehicle's historical driving status information and predictive control quantity sequence to obtain the vehicle's predicted driving trajectory within a preset time period.

[0091] The fourth step is to train the initial model based on the predicted driving trajectory of the vehicle within a preset time period to obtain the control quantity prediction network; wherein, the control quantity prediction network is used to implement step 302.

[0092] Specifically, before invoking the control quantity prediction network, the device invokes an initial model and obtains historical environmental perception results (including LiDAR point clouds, camera images, millimeter-wave radar data, and BEV rasterized maps) and historical driving state information (including position, speed, heading angle, pose, etc.) from the training dataset within a historical time period. Based on this initial model, feature extraction and prediction processing are performed on the acquired historical environmental perception results to obtain a predicted control quantity sequence, including the vehicle's acceleration and steering angle at N moments within a preset time period; N is an integer greater than or equal to 2, and the preset time period is later than the historical time period. The device can invoke a preset physical model, which is a mathematical model built based on vehicle kinematics or dynamics principles to simulate the evolution of the vehicle's state. Based on this physical model, using the driving state at the first moment in the historical driving state information and the acceleration and steering angle at the first moment in the control quantity sequence as initial values, trajectory recursion processing is performed iteratively to obtain the predicted driving trajectory of the vehicle within the preset time period, including the vehicle's state at N moments. The predicted driving trajectory and the historical driving status information are compared with the state of each vehicle within the corresponding preset time period to optimize the model parameters of the initial model and obtain the trained model, which is the control quantity prediction network. Based on the control quantity prediction network, the real-time environmental perception results are processed to predict the initial control quantity sequence of the vehicle.

[0093] By using deep learning networks to learn complex scene features from multimodal sensor data (such as BEV gridded maps and LiDAR point clouds), the dynamic obstacle perception capability of the control variable prediction network is improved, and misjudgments are reduced. Through end-to-end optimization, the time-consuming steps of post-processing correction are reduced, planning efficiency is improved, and millisecond-level response requirements are met.

[0094] In one possible implementation, the fourth step includes:

[0095] Step 1: Determine the multi-objective loss function based on the vehicle's driving trajectory and actual trajectory within a preset time period; the multi-objective loss function includes trajectory prediction loss, collision loss, lane centering loss, and comfort loss.

[0096] Step 2: Optimize the initial model based on the multi-objective loss function to obtain the control quantity prediction network.

[0097] Specifically, based on the multi-objective loss function calculation formula, the vehicle's driving trajectory within a preset time period in the predicted driving trajectory and the vehicle's actual trajectory within a preset time period in the historical driving status information are calculated to obtain the multi-objective loss function. Based on this multi-objective loss function, the model parameters of the initial model are optimized and trained to obtain the control quantity prediction network.

[0098] In one example, the multi-objective loss function includes trajectory prediction loss, collision loss, lane centering loss, and comfort loss. The trajectory prediction loss function can be calculated using the formula to calculate the vehicle's trajectory within a preset time period in the predicted trajectory, as well as the vehicle's actual trajectory within the preset time period from historical driving status information. This results in a trajectory prediction loss value, ensuring the predicted trajectory approximates the expert driver's reference trajectory. Similarly, the collision loss function can be calculated using the formula to penalize collisions between the predicted trajectory and obstacles. Finally, the lane centering loss function can be calculated using the formula to penalize collisions between the predicted trajectory and obstacles. The lane centering loss function value is calculated by analyzing the vehicle's actual trajectory within a preset time period and historical driving status information to keep the vehicle centered in the lane. A comfort loss function value is also calculated using the formula to penalize drastic changes in control parameters and ensure passenger comfort. Finally, a weighted combination of the trajectory prediction loss function value, collision loss function value, lane centering loss function value, and comfort loss function value yields the result corresponding to the multi-objective loss function.

[0099] In one example, the weights of each component in the loss function can be dynamically adjusted based on the scene complexity metric. For instance, a collision loss weight can be increased in complex scenes, while a comfort loss weight can be increased in simple scenes. This scene complexity metric includes obstacle density (the number of obstacles per unit area) and road curvature (the degree of road curvature). By dynamically adjusting the loss function weights, scene adaptive optimization is achieved. In complex scenes, obstacle avoidance capabilities are prioritized, while in simple scenes, ride comfort is optimized. This balances planning needs under different environments and improves the adaptability and stability of subsequent trajectory planning based on the control variable prediction network.

[0100] For example, Figure 4 A flowchart illustrating a deep learning-assisted driving trajectory planning scheme provided in this application embodiment is shown below. Figure 4As shown, a large-scale dataset collected from expert drivers is acquired, including multimodal environmental perception results and historical driving state information. This dataset is input into a feature extraction network, which outputs environmental features. A control prediction network then processes these environmental features to produce an initial predicted control sequence. A physical constraint layer maps this initial predicted control sequence to obtain the final predicted control sequence, including acceleration and steering wheel angle control sequences corresponding to the trajectory in the next 5 seconds. A differentiable kinematics model performs trajectory recursion processing on the predicted control sequence and historical driving state information to obtain the predicted driving trajectory of the vehicle within a preset time period, and then outputs the planned trajectory. Based on the predicted driving trajectory output from this planned trajectory, and combined with the driver's trajectory as the true prediction value, a data-driven approach is used to supervise the learning of the control prediction network and the physical constraint layer. Specifically, the loss gradient is backpropagated to the control prediction network and the physical constraint layer through an optimizer. Deep learning and gradient descent methods are used to reduce the multi-objective loss function, thereby optimizing the model's prediction performance. The differentiable kinematics model acts as a "differentiable intermediate layer," enabling joint optimization of data-driven and kinematic constraints.

[0101] By using a differentiable intermediate layer, the predicted trajectory generated by the physical model is compared with the actual trajectory, a multi-objective loss function is calculated, and the loss gradient is backpropagated to the deep learning network to adjust its prediction results to meet physical constraints, thereby achieving joint optimization of data-driven and physical constraints.

[0102] 303. Based on the physical constraint range, the initial control quantity sequence is mapped to obtain the vehicle control quantity sequence.

[0103] For example, the device acquires a pre-defined range of physical constraints, such as the range of vehicle kinematic limits (e.g., maximum acceleration, maximum steering angle). Using an activation function, the output of the control quantities (initial acceleration and initial steering angle) at each moment in the initial control quantity sequence is limited to the range [-1, 1]. These are then mapped to the vehicle's actual physical limits (e.g., maximum acceleration, maximum steering angle) using a scaling factor, generating a trajectory sequence that conforms to the kinematic constraints, i.e., predicting the vehicle's control quantity sequence.

[0104] In one example, the output range is hard-constrained by the hyperbolic tangent function tanh activation function, including: longitudinal acceleration a = a max ·tanh(oa), front wheel steering angle δ=δ max ·tanh(oδ), where oa and oδ are the initial longitudinal acceleration and initial front wheel angle at each moment in the initial control sequence output by the network prediction layer, respectively, and the maximum acceleration a max and maximum front wheel steering angle δ maxThese are pre-set vehicle kinematic limit values ​​to ensure that the predicted control quantities do not exceed the physical constraints.

[0105] By using the hyperbolic tangent function, the control quantity is limited to the vehicle's kinematic limits, avoiding the generation of control quantities that exceed physical feasibility (such as rapid acceleration or excessive steering angles). This ensures that the trajectory generated by the physical model strictly follows the vehicle's dynamic constraints, thereby improving the safety and feasibility of trajectory planning.

[0106] 304. Let t=1, and repeat the following process until t=T: Based on the vehicle's driving state at time t and the acceleration and steering angle at time t in the control quantity sequence, obtain the vehicle's driving state at time t+1; let t=t+1.

[0107] For example, based on the physical model, let t=1, iteratively calculate the vehicle's driving state at time t and the acceleration and steering angle at time t in the control quantity sequence to obtain the vehicle's driving state at time t+1. Let t=t+1, iteratively calculate the vehicle's driving state at time t+1 and the acceleration and steering angle at time t+1 in the control quantity sequence to obtain the vehicle's driving state at time t+2. Repeat the above process until the driving state at time T is generated.

[0108] For example, the control sequence includes acceleration and steering angle for each 1 second within the next 3 seconds. The current driving state of the vehicle is determined as the driving state of the vehicle in the 1st second. Based on the differentiable trajectory generator, the driving state of the vehicle in the 1st second and the acceleration and steering angle in the control sequence for the 1st second are calculated to obtain the driving state of the vehicle in the 2nd second. The driving state of the vehicle in the 2nd second and the acceleration and steering angle in the control sequence for the 2nd second are calculated to obtain the driving state of the vehicle in the 3rd second. Thus, the driving state of the vehicle for each 1 second within the next 3 seconds is obtained.

[0109] In one possible implementation, the driving state includes driving speed, vehicle orientation, and coordinate information of the corresponding trajectory point in a preset coordinate system; step 304 includes:

[0110] Step 1: Based on the vehicle's speed at time t and the acceleration at time t in the control sequence, obtain the vehicle's speed at time t+1.

[0111] Step 2: Based on the vehicle's speed and orientation at time t, and the steering angle at time t in the control sequence, obtain the vehicle's orientation at time t+1.

[0112] Step 3: Based on the vehicle's speed, orientation, and coordinates at time t, obtain the vehicle's coordinates at time t+1.

[0113] Specifically, the vehicle's driving state at each moment includes driving speed, vehicle orientation, and coordinate information of the trajectory point in the preset coordinate system at the corresponding moment; wherein, the vehicle orientation may include the vehicle orientation (i.e., absolute orientation, such as heading angle) of the trajectory point in the preset coordinate system at the corresponding moment, and the coordinate information includes the first coordinate (such as the horizontal coordinate) and the second coordinate (such as the vertical coordinate) of the trajectory point in the preset coordinate system at the corresponding moment. For the calculation process of the driving state at each moment, the driving speed of the vehicle at moment t and the acceleration predicted in the control sequence at moment t are calculated using a preset calculation formula to obtain the driving speed of the vehicle at moment t+1. The driving speed, vehicle orientation, and steering angle at moment t in the control sequence at moment t are calculated using a preset calculation formula to obtain the vehicle orientation at moment t+1. The driving speed, vehicle orientation, and the first coordinate (e.g., the horizontal coordinate) in the coordinate information of the vehicle at moment t are calculated using a preset calculation formula to obtain the first coordinate (e.g., the horizontal coordinate) in the coordinate information of the vehicle at moment t+1. The driving speed, vehicle orientation, and the second coordinate (e.g., the vertical coordinate) in the coordinate information of the vehicle at moment t are calculated to obtain the second coordinate (e.g., the vertical coordinate) in the coordinate information of the vehicle at moment t+1.

[0114] In one example, a differentiable trajectory generator is constructed based on a bicycle model using the following formula:

[0115] ,

[0116] Regarding the driving state [x] at the previous time t t y t θ t v t Longitudinal acceleration and front wheel cornering Recursively calculate the driving state [x] at the next time step t+1. t+1 y t+1 θ t+1 v t+1 ], where L is the vehicle's wheelbase, Δt is the prediction time step, and v t Let θ be the speed at the previous moment t. t+1 Let x be the vehicle orientation of the trajectory point corresponding to the previous time t in the preset coordinate system. t Let y be the x-coordinate of the trajectory point corresponding to the previous time t in the preset coordinate system.t v is the ordinate of the trajectory point corresponding to the previous time t in the preset coordinate system. t+1 Let θ be the speed at the next time step t+1. t+1 Let y be the vehicle orientation of the trajectory point at the next time t+1 in the preset coordinate system. t+1 Let x be the x-coordinate of the trajectory point at the next time t+1 in the preset coordinate system. t+1 This is the ordinate of the trajectory point at the next time t+1 in the preset coordinate system.

[0117] By using the current vehicle state as the initial value, future trajectories are generated recursively based on a physical model, ensuring that the generated trajectories strictly follow vehicle kinematic constraints (such as turning radius and acceleration limits), thereby achieving safety, comfort, and real-time performance in generating trajectories in complex scenarios.

[0118] 305. Generate the vehicle's trajectory for future time periods based on the vehicle's driving status at each time point.

[0119] For example, the trajectories of the vehicle's driving state at various times can be merged to obtain the vehicle's driving trajectory over a future time period. For instance, the trajectories of the vehicle's driving state at each 1-second interval within the next 3 seconds can be merged to obtain the vehicle's driving trajectory over the next 3 seconds.

[0120] In one example, the vehicle's driving status at various times is merged to obtain the initial driving trajectory of the vehicle in the future time period. Then, the vehicle's load, tire adhesion coefficient, tire lateral stiffness and other state parameters are obtained in real time. The initial driving trajectory is dynamically corrected based on these state parameters to obtain the final driving trajectory of the vehicle in the future time period.

[0121] By dividing the trajectory planning task into two stages, "control quantity prediction" and "physical recursion", deep learning networks are used to predict control quantities (such as acceleration and steering angle), and the trajectory is generated recursively through a physical model, which can reduce the physical infeasibility of trajectory points in the driving trajectory.

[0122] In this embodiment, based on the above embodiments, on the one hand, the vehicle kinematics model is integrated with a deep learning architecture to generate trajectories that strictly follow steering geometry and dynamic constraints, thereby reducing the risk of physically infeasible trajectories; on the other hand, through an end-to-end intelligent decision-making mechanism, the environmental understanding and obstacle avoidance capabilities in complex dynamic scenarios are significantly enhanced, achieving the safety and comfort of humanoid driving; at the same time, the real-time planning performance is greatly improved, reducing deficiencies in computational efficiency and scenario generalization capabilities, and improving the reliability, intelligence, and efficiency of the assisted driving system.

[0123] Figure 5 This is a schematic diagram of the structure of a trajectory generation device provided in an embodiment of this application, as shown below. Figure 5 As shown, the device includes:

[0124] The acquisition module 401 is used to acquire the vehicle's real-time environmental perception results and current driving status;

[0125] The prediction module 402 is used to predict the vehicle's control quantity sequence based on the real-time environmental perception results; wherein, the control quantity sequence includes the vehicle's acceleration and steering angle at T time points in the future time period; T is an integer greater than or equal to 2;

[0126] The recursive module 403 is used to perform trajectory recursive processing on the vehicle's current driving state and control quantity sequence to obtain the vehicle's driving trajectory in the future time period.

[0127] In one possible implementation, the prediction module 402 is specifically used to: predict the initial control quantity sequence of the vehicle based on the real-time environmental perception results; and map the initial control quantity sequence based on the physical constraint range to obtain the vehicle's control quantity sequence.

[0128] In one possible implementation, the prediction module 402 is specifically used for: extracting features from the real-time environment perception results to obtain the spatiotemporal features of the real-time environment perception results; and processing the spatiotemporal features to obtain the initial control quantity sequence of the vehicle.

[0129] In one possible implementation, the recursive module 403 is specifically used to: let t=1, and repeat the following process until t=T: obtain the vehicle's driving state at time t+1 based on the vehicle's driving state at time t and the acceleration and steering angle at time t in the control quantity sequence; let t=t+1; generate the vehicle's driving trajectory in the future time period based on the vehicle's driving state at each time.

[0130] In one possible implementation, the driving state includes driving speed, vehicle orientation, and coordinate information of the corresponding trajectory point in a preset coordinate system; the recursive module 403 is specifically used to: obtain the driving speed of the vehicle at time t+1 based on the driving speed of the vehicle at time t and the acceleration at time t in the control quantity sequence; obtain the vehicle orientation at time t+1 based on the driving speed and vehicle orientation of the vehicle at time t and the steering angle at time t in the control quantity sequence; and obtain the coordinate information of the vehicle at time t+1 based on the driving speed, vehicle orientation, and coordinate information of the vehicle at time t.

[0131] In one possible implementation, the device is further configured to: acquire historical environmental perception results and historical driving state information of the vehicle; perform predictive processing on the historical environmental perception results based on an initial model to obtain a predictive control quantity sequence for the vehicle; wherein the control quantity sequence includes the vehicle's acceleration and steering angle at N moments within a preset time period; N is an integer greater than or equal to 2; perform trajectory recursive processing on the vehicle's historical driving state information and the predictive control quantity sequence to obtain the vehicle's predicted driving trajectory within the preset time period; and train the initial model based on the vehicle's predicted driving trajectory within the preset time period to obtain a control quantity prediction network; wherein the control quantity prediction network is used to implement the step of predicting the vehicle's control quantity sequence based on real-time environmental perception results.

[0132] In one possible implementation, the device is further specifically used to: determine a multi-objective loss function based on the vehicle's driving trajectory and actual trajectory within a preset time period; wherein the multi-objective loss function includes trajectory prediction loss, collision loss, lane centering loss, and comfort loss; and optimize the initial model based on the multi-objective loss function to obtain a control quantity prediction network.

[0133] The apparatus in this embodiment can execute the technical solutions in the above method. Its specific implementation process and technical principles are the same, and will not be repeated here.

[0134] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 6 As shown, the electronic device includes: a memory 501 and a processor 502; the memory 501 is a memory used to store instructions executable by the processor 502.

[0135] The processor 502 is configured to perform the method provided in the above embodiments.

[0136] The electronic device also includes a receiver 503 and a transmitter 504. The receiver 503 is used to receive instructions and data sent by other devices, and the transmitter 504 is used to send instructions and data to external devices.

[0137] The specific implementation process of the processor can be found in the above method embodiments, and its implementation principle and technical effect are similar, so it will not be repeated here.

[0138] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0139] This application also provides a vehicle, which includes a vehicle body and the electronic equipment as described above, to implement the technical solutions of the above embodiments.

[0140] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed on a computer, cause the computer to perform the technical solutions described above.

[0141] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory, electrically erasable programmable read-only memory, erasable programmable read-only memory, programmable read-only memory, read-only memory, magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0142] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. The readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in a device.

[0143] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium, and when the at least one processor executes the computer program, it can implement the technical solutions in the above embodiments.

[0144] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as magnetic disks or optical disks.

[0145] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A trajectory generation method, characterized in that, include: Obtain the vehicle's real-time environmental perception results and current driving status; Based on the real-time environmental perception results, a control quantity sequence for the vehicle is predicted; wherein, the control quantity sequence includes the vehicle's acceleration and steering angle at T time points within a future time period; T is an integer greater than or equal to 2; The current driving state and control quantity sequence of the vehicle are processed by trajectory recursion to obtain the driving trajectory of the vehicle in the future time period.

2. The method according to claim 1, characterized in that, The step of predicting the control quantity sequence of the vehicle based on the real-time environmental perception results includes: Based on the real-time environmental perception results, the initial control quantity sequence of the vehicle is predicted; Based on the physical constraints, the initial control quantity sequence is mapped to obtain the control quantity sequence of the vehicle.

3. The method according to claim 2, characterized in that, The step of predicting the initial control quantity sequence of the vehicle based on the real-time environmental perception results includes: Feature extraction is performed on the real-time environment perception results to obtain the spatiotemporal features of the real-time environment perception results; The spatiotemporal features are processed to obtain the initial control quantity sequence of the vehicle.

4. The method according to claim 1, characterized in that, The process of recursively processing the current driving state and control quantity sequence of the vehicle to obtain the driving trajectory of the vehicle in a future time period includes: Let t=1, and repeat the following process until t=T: Based on the vehicle's driving state at time t and the acceleration and steering angle at time t in the control quantity sequence, obtain the vehicle's driving state at time t+1; let t=t+1; Based on the vehicle's driving status at each moment, the vehicle's driving trajectory for a future time period is generated.

5. The method according to claim 4, characterized in that, The driving state includes driving speed, vehicle orientation, and coordinate information of the corresponding trajectory points in a preset coordinate system; obtaining the driving state of the vehicle at time t+1 based on the driving state of the vehicle at time t and the acceleration and steering angle at time t in the control quantity sequence includes: The vehicle's speed at time t+1 is obtained based on the vehicle's speed at time t and the acceleration at time t in the control quantity sequence. Based on the vehicle's speed and orientation at time t, and the steering angle at time t in the control quantity sequence, the vehicle's orientation at time t+1 is obtained. Based on the vehicle's speed, orientation, and coordinates at time t, the coordinates of the vehicle at time t+1 are obtained.

6. The method according to any one of claims 1-5, characterized in that, The method further includes: Acquire historical environmental perception results and historical driving status information of the vehicle; Based on the initial model, the historical environmental perception results are processed to obtain a predictive control sequence for the vehicle; wherein, the control sequence includes the vehicle's acceleration and steering angle at N moments within a preset time period; N is an integer greater than or equal to 2; The historical driving status information and predictive control quantity sequence of the vehicle are processed by trajectory recursion to obtain the predicted driving trajectory of the vehicle within a preset time period. Based on the predicted driving trajectory of the vehicle within a preset time period, the initial model is trained to obtain a control quantity prediction network; wherein, the control quantity prediction network is used to implement the step of predicting the control quantity sequence of the vehicle based on the real-time environmental perception results.

7. The method according to claim 6, characterized in that, The step of training the initial model based on the predicted driving trajectory of the vehicle within a preset time period to obtain a control quantity prediction network includes: Based on the vehicle's driving trajectory and actual trajectory within a preset time period, a multi-objective loss function is determined; wherein, the multi-objective loss function includes trajectory prediction loss, collision loss, lane centering loss, and comfort loss; The initial model is optimized based on the multi-objective loss function to obtain the control quantity prediction network.

8. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

9. A vehicle, characterized in that, The vehicle includes a vehicle body and the electronic equipment as described in claim 8.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-7.

Images