Vehicle trajectory generation method, computer device and vehicle

Abstract

The application discloses a vehicle driving track generation method, a computer device and a vehicle, and belongs to the technical field of vehicle control. The method comprises the following steps: generating a tree structure comprising multiple levels based on continuous time steps, one time step corresponding to one level, each level comprising multiple nodes, and each node corresponding to a driving action of the vehicle at the current level; determining a second driving action of a second level based on a first driving action of a first level in the tree structure, the second level being located after the first level; determining multiple driving action sequences according to the second driving action and the first driving action, the driving action sequence comprising driving actions corresponding to effective nodes of each level; and determining a vehicle driving track of the continuous time steps based on the multiple driving action sequences. The application can improve the real-time performance of vehicle driving track generation and the safety of vehicle driving.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and in particular to a method for generating vehicle driving trajectory, a computer device, and a vehicle. Background Technology

[0002] With the rapid development of autonomous driving technology, autonomous vehicles have begun to be applied in various road traffic scenarios. Generating vehicle driving trajectories, as a core component of autonomous driving systems, directly impacts the driving safety and passenger comfort of autonomous vehicles.

[0003] In related technologies, autonomous vehicles collect information about their surrounding environment through onboard sensors and combine this information with the vehicle's own operating status and road network information to plan their driving trajectory, thereby achieving autonomous driving control. When planning their driving trajectory, autonomous vehicles typically exhaustively extrapolate feasible driving paths over a future time period and calculate the driving cost of each feasible path to select the optimal trajectory, thus obtaining the final vehicle driving trajectory.

[0004] However, the process of generating vehicle driving trajectories involves a large amount of computation and takes a long time, resulting in poor real-time performance and affecting vehicle driving safety. Summary of the Invention

[0005] This application provides a method for generating vehicle driving trajectories, a computer device, and a vehicle, which can improve the real-time performance of vehicle driving trajectory generation and the safety of vehicle operation. The technical solution is as follows: On one hand, embodiments of this application provide a method for generating a vehicle driving trajectory, the method comprising: A tree structure with multiple levels is generated based on continuous time steps. One time step corresponds to one level, and each level contains multiple nodes. Each node corresponds to the vehicle's driving action at the current level. The second driving action of the second level is determined based on the first driving action of the first level in the tree structure, and the second level is located after the first level. Multiple driving action sequences are determined based on the second driving action and the first driving action, and the driving action sequences include driving actions corresponding to valid nodes at each level; The vehicle trajectory with the continuous time step is determined based on the multiple driving action sequences.

[0006] On the other hand, embodiments of this application provide a vehicle trajectory generation apparatus, the apparatus comprising: The generation module is used to generate a tree structure with multiple levels based on continuous time steps. One time step corresponds to one level, and each level contains multiple nodes. Each node corresponds to the vehicle's driving action at the current level. The determining module is used to determine a second driving action at a second level based on a first driving action at a first level in the tree structure, wherein the second level is located after the first level. The determining module is further configured to determine multiple driving action sequences based on the second driving action and the first driving action, wherein the driving action sequences include driving actions corresponding to valid nodes at each level; The determining module is further configured to determine the vehicle trajectory with the continuous time steps based on the multiple driving action sequences.

[0007] In one possible implementation, the first level includes a first sub-level and a second sub-level that are temporally adjacent, and the first sub-level is located before the second sub-level. In the first level, the first driving action of any sub-level includes a first lateral driving action, and the second driving action of the second level includes a second lateral driving action. The second lateral driving action is consistent with the first lateral driving action of the second sub-level.

[0008] In one possible implementation, the first driving action includes a first longitudinal driving action, and the second driving action includes a second longitudinal driving action, which is the same as the first longitudinal driving action.

[0009] In one possible implementation, the driving action corresponding to each node includes the lateral driving action and the longitudinal driving action of the vehicle, and the effective node of the second level is the node that matches the second lateral driving action and the second longitudinal driving action in the driving action corresponding to each node of the second level.

[0010] In one possible implementation, the determining module is configured to determine, based on the valid nodes corresponding to the first driving action and the valid nodes corresponding to the second driving action, an abnormal action sequence including abnormal driving actions from the plurality of driving action sequences. The abnormal driving action is a driving action corresponding to an abnormal node. The abnormal node includes at least one of risk nodes or invalid action nodes at any level. The risk node is a node where the driving action has a driving safety hazard, and the invalid action node is a node where the driving action cannot be executed normally. The abnormal action sequences are removed from the plurality of driving action sequences to obtain valid action sequences; the vehicle driving trajectory is determined based on the valid action sequences.

[0011] In one possible implementation, the device further includes a prediction module, which is used to predict the obstacle position corresponding to the reference time of the second level based on the vehicle's environmental data, wherein the reference time is determined based on the current time and the time step. The determining module is further configured to determine the risk coefficient of each effective node in the second level using the obstacle position, the expected collision time, and the vehicle position, wherein the vehicle position is the position indicated by the effective node corresponding to the second driving action; and to determine the effective node whose risk coefficient is greater than or equal to the risk threshold as the risk node.

[0012] In one possible implementation, the determining module is further configured to determine a reference cost using the actual cost and predicted cost corresponding to the plurality of driving action sequences, wherein the actual cost is the driving cost generated by extrapolating the first driving action along the driving action sequence to the target driving action, and the predicted cost is the expected driving cost extrapolated from the driving action after the target driving action in the driving action sequence; and to determine valid nodes whose reference cost is greater than or equal to a cost threshold as invalid action nodes.

[0013] In one possible implementation, the determining module is configured to determine multiple basic driving trajectories based on the multiple driving action sequences; perform verification on each basic driving trajectory to obtain a verification result; and determine the vehicle driving trajectory from the multiple basic driving trajectories based on the verification result.

[0014] In one possible implementation, the verification result includes at least one of the following: lateral safety distance verification result, longitudinal safety distance verification result, or braking distance verification result; The lateral safety distance verification result is determined based on the vehicle driving posture corresponding to the valid nodes in the basic driving trajectory and the lateral relative distance between the vehicle and surrounding obstacles. The longitudinal safety distance verification result is determined based on the vehicle speed corresponding to the valid nodes in the basic driving trajectory and the longitudinal relative distance between the vehicle and the surrounding obstacles. The braking distance verification result is determined based on the vehicle speed corresponding to the valid nodes in the basic driving trajectory.

[0015] In one possible implementation, the verification result includes a driving efficiency verification result, which is determined based on at least one of the following: speed cost, acceleration cost, lateral offset cost, safety distance cost, heading angle change cost, and rule cost corresponding to the effective nodes in the basic driving trajectory.

[0016] In one possible implementation, the determining module is configured to obtain the vehicle state vector corresponding to each driving action in the plurality of driving action sequences; generate trajectory segments using the vehicle state vectors corresponding to the driving actions; and sequentially splice the trajectory segments according to the driving action sequence to obtain the plurality of basic driving trajectories.

[0017] On the other hand, this application provides a vehicle, which includes a vehicle body and a vehicle control system; the vehicle control system is used to generate a vehicle driving trajectory based on the above method, and control the vehicle body to drive based on the vehicle driving trajectory.

[0018] On the other hand, embodiments of this application provide a computer device, which includes a processor and a memory. The memory stores at least one piece of program code, which is loaded and executed by the processor to enable the computer device to implement any of the above-described methods for generating vehicle driving trajectories.

[0019] On the other hand, a computer-readable storage medium is also provided, wherein at least one piece of program code is stored in the computer-readable storage medium, the at least one piece of program code being loaded and executed by a processor to enable a computer to implement any of the above-described methods for generating vehicle driving trajectories.

[0020] On the other hand, a computer program or computer program product is also provided, wherein the computer program or computer program product stores at least one computer instruction, which is loaded and executed by a processor to enable the computer to implement any of the above-mentioned methods for generating vehicle driving trajectories.

[0021] The technical solution provided in this application has at least the following beneficial effects: This application generates a tree structure with multiple levels based on continuous time steps. The nodes in each level of the tree structure correspond one-to-one with the vehicle's driving actions. Then, the first driving action of the preceding first level is used to determine the second driving action of the following second level. Combining the first and second driving actions, a valid sequence of driving actions corresponding to valid nodes across multiple levels is determined. This avoids redundant calculations caused by driving actions with invalid nodes in the valid sequence. The vehicle's driving trajectory is determined based on the driving action sequence with continuous time steps, improving the accuracy of vehicle driving trajectory generation, reducing invalid calculations in the trajectory generation process, and reducing the generation time of the vehicle driving trajectory, thereby improving the real-time performance of the vehicle driving trajectory and the safety of vehicle driving. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of an implementation environment provided in an embodiment of this application; Figure 2 This is a flowchart of a method for generating a vehicle driving trajectory provided in an embodiment of this application; Figure 3 This is a schematic diagram of a vehicle trajectory generation device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a server provided in an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0025] It should be noted that the terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. 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.

[0026] Figure 1 This is a schematic diagram of an implementation environment provided in an embodiment of this application. For example... Figure 1 As shown, the implementation environment may include a vehicle 101 and a vehicle control system 102. The vehicle control system 102 is used to generate a vehicle driving trajectory and control the vehicle 101 to drive based on the vehicle driving trajectory. The vehicle 101 is a vehicle with autonomous driving environment perception capability.

[0027] The vehicle control system 102 can be located inside the vehicle 101, for example, the vehicle control system 102 is an in-vehicle terminal; the vehicle control system 102 can also be located outside the vehicle 101, for example, the vehicle control system 102 is a cloud control system.

[0028] The vehicle control system 102 can be a single server, or it can be a server cluster consisting of multiple servers that perform different functions, or it can be a cloud computing center.

[0029] Vehicle 101 may also have communication capabilities. Vehicle 101 may be equipped with a communication module that supports wireless or wired communication technologies. Vehicle 101 interacts with the vehicle control system 102 via this communication module. For example, vehicle 101 can transmit its real-time driving status data to the vehicle control system 102 via the communication module and receive the vehicle's driving trajectory from the vehicle control system 102, thus enabling data interaction between the two and ensuring the smooth execution of trajectory generation and vehicle driving control.

[0030] Based on the above Figure 1 The implementation environment shown in this application embodiment provides a method for generating a vehicle driving trajectory. The method for generating a vehicle driving trajectory includes, but is not limited to, the following steps: Step S1: Divide the planned time period into multiple time steps. Each time step corresponds to multiple candidate driving actions of the vehicle. The multiple time steps include a first time step and a second time step that follows the first time step.

[0031] Step S2: Configure the driving action of the second time step to be consistent with the driving action of the first time step, where the driving action of the first time step is any one of the candidate driving actions.

[0032] Step S3: Determine multiple driving action sequences based on the driving actions at the first time step, the driving actions at the second time step, and the temporal relationship of the time steps.

[0033] Step S4: Generate the vehicle driving trajectory corresponding to the planned time period based on multiple driving action sequences.

[0034] In the exemplary embodiment of the application, steps S1 to S4 can be used to plan the vehicle's driving trajectory using a tree structure corresponding to the planning time period. The planning time period is divided into multiple time steps, and any time step can be used as a level of the tree structure. Multiple candidate driving actions of the vehicle corresponding to any time step can be used as nodes in the level. The tree structure is obtained by sorting each level according to the order of the time steps.

[0035] The first time step corresponds to the first level of the tree structure, and the second time step corresponds to the second level. The temporal sequence of the first time step is earlier than that of the second time step. A first driving action is selected from the candidate driving actions of the first time step, and the driving action corresponding to the second time step is determined to be consistent with the first driving action. The nodes corresponding to the first driving action at the first level and the second driving action at the second level are considered valid nodes. The driving actions corresponding to the valid nodes at each level are arranged according to the temporal relationship of the time steps, resulting in multiple driving action sequences. Finally, based on the obtained multiple driving action sequences, the vehicle driving trajectory corresponding to the planned time period is generated.

[0036] It should be noted that, besides using a tree structure, the vehicle trajectory generation method of this application can also be implemented using topology graphs or tables. The principles of topology graphs and tables are similar to those of tree structures, and will not be elaborated further here. This application does not limit the implementation method of the vehicle trajectory generation method. This application uses a tree structure as an example for illustration.

[0037] like Figure 2 As shown, taking the method for generating vehicle driving trajectory applied to vehicle control system 102 as an example, the method may include steps 201 to 204.

[0038] In step 201, a tree structure with multiple levels is generated based on continuous time steps. One time step corresponds to one level, and each level contains multiple nodes. Each node corresponds to the vehicle's driving action at the current level.

[0039] Autonomous vehicles must execute their actions sequentially according to time. If an autonomous vehicle makes decisions based solely on its current driving status, it cannot anticipate subsequent driving conditions, leading to issues such as stuttering in action transitions and delayed command responses, thus affecting driving safety. Therefore, autonomous vehicles need to predict and plan their driving trajectories in advance for future time periods (planned time periods), pre-planning the driving actions that can be executed at different times to ensure smooth transitions between actions and timely responses.

[0040] In an exemplary embodiment of this application, after obtaining the current driving time, the planned future driving duration can be determined based on the current driving time. The planned future driving duration can be determined based on a preset planned duration, vehicle operating status, and driving environment complexity. For example, the planned future driving duration can be 6 seconds, 7 seconds, or 8 seconds.

[0041] A time step refers to the smallest unit of time obtained by dividing the planned future travel time into equal or unequal segments. It is used to break down the overall future travel time into multiple consecutive time segments. In this embodiment, the planned future travel time is evenly divided at fixed time intervals to obtain several consecutive time steps, thereby realizing the segmented planning of travel actions along the time dimension. For example, the time step can be 100ms (milliseconds). If the planned future travel time is set to 7 seconds, then 70 consecutive time steps can be obtained.

[0042] Each time step corresponds to a level in the tree structure. These levels are arranged sequentially according to time, with time as the vertical axis, generating a tree structure. Each level can include multiple nodes, and each node corresponds to a driving action that the vehicle can choose to execute at that level's time step. For example, using a horizontal and vertical decoupling design, a single node corresponds to a set of combined action pairs consisting of lateral and longitudinal driving actions, representing the complete driving action that the vehicle can choose to execute at that level's time step.

[0043] For example, the vehicle's driving motion includes lateral driving motion and longitudinal driving motion, with the lateral driving motion decoupled from the longitudinal driving motion. For instance, the overall driving motion can be expressed by the formula A=A lat ×A lon Characterization, A lat Represents lateral movement, A lon This represents longitudinal driving actions. Lateral driving actions include, but are not limited to, lane keeping, changing lanes to the left (lane 1), driving over the lane line (lane 2), and returning to the lane; longitudinal driving actions include, but are not limited to, constant speed driving, smooth acceleration, smooth deceleration, normal braking, emergency braking, and following other vehicles. Since changing lanes to the right is against traffic regulations, this action can be disabled.

[0044] Taking lateral movement as an example with 4 selectable actions and longitudinal movement as an example with 6 selectable actions, by combining them pairwise according to the Cartesian product rule, a total of 24 joint action pairs can be generated, meaning that this level can include 24 nodes. For example, if the lateral movement actions are labeled as H1, H2, H3, H4, and the longitudinal movement actions are labeled as Z1, Z2, Z3, Z4, Z5, Z6, then the movement actions corresponding to each node are (H1, Z1), (H1, Z2)...(H1, Z6), (H2, Z1)...(H4, Z6).

[0045] Nodes at different levels form multiple hierarchical branches through the succession of driving actions. Each hierarchical branch corresponds to a sequence of driving actions. The succession of driving actions means that the formal actions corresponding to the nodes conform to the rationality of vehicle driving, traffic rules, and action connection logic. By using the succession of driving actions, the driving action node corresponding to the previous time step is sequentially linked with the optional driving action node corresponding to the next time step, thereby generating multiple independent hierarchical branches. Each hierarchical branch represents a complete set of temporal driving action combinations, corresponding to a candidate driving path to be selected.

[0046] For example, 70 consecutive time steps correspond to 70 levels of the tree structure. All levels are arranged vertically in chronological order, and multiple joint action nodes composed of lateral movement and vertical movement are configured for each level to obtain a 70-level tree structure.

[0047] When determining hierarchical branches, a node is selected in each level, and the action sequence relationship between adjacent levels is set according to driving logic and traffic rules: for example, a node containing lane keeping and constant speed driving can be followed by a node that combines lane keeping with constant speed and smooth acceleration. By connecting nodes in this way level by level, multiple different hierarchical branches can be formed.

[0048] In step 202, the second driving action of the second level is determined based on the first driving action of the first level in the tree structure, and the second level is located after the first level.

[0049] In the embodiments of this application, the first level and the second level may include one level in the tree structure, or multiple levels in the tree structure. Taking a tree structure with N levels as an example, for instance, the first level may be a level with k=1, and the second level may be a level from k=2 to N. Another example is that the first level may be two consecutive levels with k=1 and k=2, and the second level may correspond to levels from k=3 to N. It should be noted that the selection range and number of the first and second levels in this application are illustrative and are not intended to limit the scope of the application.

[0050] The first level is used for intent detection, which comprehensively traverses various driving intents through the driving actions corresponding to each node within this level; the second level is used for intent locking, which fixes the direction and type of driving actions in all subsequent levels based on the driving intents obtained from the first level detection.

[0051] The first driving action refers to the driving action to be performed by the vehicle corresponding to the first-level valid node; the second driving action refers to the driving action to be performed by the vehicle corresponding to the second-level valid node. The second driving action can be determined based on the first driving action. Since the first driving action can include a first lateral driving action and a second longitudinal driving action, and the second driving action can include a second lateral driving action and a second longitudinal driving action, when determining the second driving action, it is necessary to combine the lateral and longitudinal driving actions in the first driving action to determine the lateral control action and longitudinal control action in the second driving action, and then determine the combined lateral and longitudinal driving action to be performed by the vehicle at the second level, i.e., the second driving action.

[0052] In an exemplary embodiment of this application, the first level includes a first sub-level and a second sub-level that are sequentially adjacent, and the first sub-level is located before the second sub-level. In the first level, the first driving action of any sub-level includes a first lateral driving action, and the second driving action of the second level includes a second lateral driving action. The second lateral driving action is consistent with the first lateral driving action of the second sub-level.

[0053] Taking the first level, which includes two levels (k=1 and k=2), and the second level, which corresponds to levels k=3 to N, as an example, the first level consists of a first sub-level (k=1) and a second sub-level (k=2) arranged sequentially. These two sub-levels correspond to different execution times, with the execution time of the k=1 sub-level preceding that of the k=2 sub-level. The first driving actions corresponding to the k=1 and k=2 sub-levels both include lane keeping and left lane changing, among other first lateral driving actions. When determining the second driving actions for subsequent levels k=3 to N in the second level, the second lateral driving actions included in these sub-levels directly adopt the first lateral driving actions of the k=2 sub-level, ensuring consistency in the lateral driving actions between levels.

[0054] It should be noted that the first level is the intent exploration level. The lateral driving actions within the first level can be random. All available lateral driving actions can be traversed to explore and filter different driving intentions.

[0055] For example, let the lateral movement action selected by the first sub-level k=1 be... The lateral movement action selected for the second sub-level k=2 is: After completing two levels of intent probing, the process enters the intent locking phase and triggers a guided branching mechanism, imposing strong constraints on subsequent lateral movement actions from k=3 to N levels: when and When the values ​​are the same, that is and The identical lateral movement indicated means that the driving intentions of two consecutive levels are consistent, forcing all subsequent levels to meet the requirement. = ;when and When the values ​​are different, that is and The indicated lateral movement actions differ, but the subsequent lateral movement actions from k=3 to N are also forced to be unified as follows: = .

[0056] For example, the lateral movement selected at level k=1. To change lanes to the left (borrow lane 1), the lateral driving action selected at level k=2. If the action is to drive over the line (borrowing lane 2), then all subsequent lateral driving actions from k=3 to N will be forcibly locked as driving over the line (borrowing lane 2).

[0057] The embodiment of this application consists of a first sub-level and a second sub-level arranged sequentially, and a setting where the second lateral driving action is consistent with the first lateral driving action of the second sub-level. After completing the initial lateral driving action screening, the action space of the subsequent levels can be reduced from a full combination of horizontal and vertical dimensions to an extension along the vertical dimension. This reduces the scope and computational load of the spatiotemporal search of the subsequent levels, ensures the uniformity of lateral driving actions at each time stage, improves the overall efficiency of action decision-making, thereby reducing the generation time of vehicle driving trajectory and improving the real-time performance and driving safety of vehicle driving trajectory.

[0058] In an exemplary embodiment of this application, the first driving action includes a first longitudinal driving action, and the second driving action includes a second longitudinal driving action, which is the same as the first longitudinal driving action.

[0059] For example, the explanation will take a first level that includes levels with k=1, and a second level that corresponds to levels with k=2 to N. k=1, and the selected longitudinal driving action is... Force constraints are applied to the longitudinal driving actions of subsequent levels k=2 to N, forcing all subsequent levels to satisfy the constraints. = .

[0060] For example, the longitudinal driving action selected at level k=1. To ensure uniform speed, the longitudinal movement of all subsequent levels from k=2 to N will be forcibly locked to uniform speed.

[0061] The embodiments of this application have a setting where the first driving action includes a first longitudinal driving action, the second driving action includes a second longitudinal driving action, and the two types of longitudinal driving actions are consistent. After determining the first longitudinal driving action, the longitudinal driving content of each subsequent level can be fixed. This reduces the traversal range of action combinations in conjunction with the driving action filtering process, reduces the amount of computation in the hierarchical search process, and maintains the consistency of longitudinal driving actions at different time-series levels. This improves the overall operational efficiency of vehicle driving action decision-making, thereby reducing the generation time of vehicle driving trajectory and improving the real-time performance and driving safety of vehicle driving trajectory.

[0062] In step 203, multiple driving action sequences are determined based on the second driving action and the first driving action. The driving action sequences include the driving actions corresponding to the valid nodes at each level.

[0063] In the exemplary embodiments of this application, the effective nodes at each level refer to the nodes selected from all nodes at the corresponding level, used to match driving actions at different levels to form a driving action sequence. The effective nodes can be determined by arbitrarily selecting one node from all nodes at the level; or, the effective nodes can be determined by the driving actions corresponding to the effective nodes of the preceding level.

[0064] For example, the valid nodes in the first level are the nodes corresponding to the first driving action. The first level includes the first temporal level of the current planning window, without considering the action constraints of the preceding levels. Therefore, the valid nodes in the first level can also be any node among all the nodes in the first level. That is to say, different valid nodes can be selected in the first level, thus corresponding to different first driving actions, thereby deriving differentiated second driving actions. Here, the planning window refers to the continuous time interval covered by a single trajectory planning. The duration of the planning window is shorter than the duration of the future time period. For example, if the duration of the future time period is 7 seconds, the duration of the planning window is 1 second.

[0065] After determining the second lateral and second longitudinal driving actions, the valid nodes of the second level can be identified from among multiple nodes in the second level. The valid nodes of the second level are the nodes corresponding to the second driving actions.

[0066] In the tree structure hierarchy, each node corresponds to a driving action that includes both lateral and longitudinal driving actions. Valid nodes in the second level are those nodes whose driving actions match the second lateral and second longitudinal driving actions within the second driving action hierarchy. Matching means that the lateral and longitudinal driving actions corresponding to a node are the same as the second lateral and second longitudinal driving actions, respectively.

[0067] For example, the driving actions corresponding to each node in the tree structure hierarchy include the vehicle's lateral driving actions and longitudinal driving actions. When filtering valid nodes in the second level, the lateral driving actions and longitudinal driving actions corresponding to each node in the second level are compared one by one with the second lateral driving actions and second longitudinal driving actions included in the second driving actions. When the lateral driving action corresponding to a node is consistent with the second lateral driving action, and the longitudinal driving action corresponding to that node is consistent with the second longitudinal driving action, it is determined that the node has completed the action matching, and that node is a valid node in the second level.

[0068] This application embodiment is based on the nodes in the hierarchy carrying lateral and longitudinal driving actions. By comparing and matching the driving actions in the two directions, valid nodes are selected from the nodes in the second level. This can select nodes that meet the driving action requirements, narrow the selection range of subsequent hierarchical branches, reduce the computational overhead caused by invalid nodes, and ensure the consistency between node selection and driving action requirements.

[0069] After determining the valid nodes at the second level, valid hierarchical branches are determined based on these valid nodes. The process of determining hierarchical branches includes, but is not limited to, steps A1 to A3.

[0070] In step A1, an abnormal action sequence including abnormal driving actions is determined from multiple driving action sequences. The abnormal driving action is the driving action corresponding to the abnormal node.

[0071] For example, the first level is the driving intention detection level, which allows random selection of lateral and longitudinal driving actions. Therefore, the valid node corresponding to the first driving action is all nodes within the first level that can be executed normally and cover all selectable lateral and longitudinal driving actions. For instance, the first level includes all nodes obtained by combining all selectable lateral and longitudinal driving actions, and the valid node corresponding to the first driving action is all nodes within this level.

[0072] The first level includes all nodes corresponding to various optional lateral and longitudinal driving actions, all of which are considered valid nodes corresponding to the first driving action. The second level selects nodes whose driving actions are consistent with the predetermined second driving action as valid nodes. When determining multiple basic level branches among multiple level branches, all level branches formed by the tree structure are traversed. Level branches that simultaneously connect the valid nodes corresponding to the first driving action and the matching valid nodes corresponding to the second driving action, and that conform to the driving action succession relationship, are determined as multiple basic level branches. Each basic level branch corresponds to a driving action sequence.

[0073] After identifying multiple basic-level branches, abnormal branches, including those with abnormal nodes, are then identified within these basic-level branches. Abnormal nodes correspond to abnormal driving actions, and abnormal branches correspond to abnormal action sequences. Abnormal nodes include at least one of two types: risk nodes or invalid action nodes at any level. Risk nodes are those where the driving action poses a driving safety hazard, while invalid action nodes are those where the driving action cannot be executed normally.

[0074] For example, the tree structure includes multiple nodes in its hierarchy, and these nodes can include valid nodes and abnormal nodes. Abnormal nodes include, but are not limited to, risk nodes and invalid action nodes. Abnormal nodes at each level can be identified before identifying abnormal branches.

[0075] In the exemplary embodiment of this application, a risk node is a node whose driving action poses a potential safety hazard. The process of determining a risk node includes, but is not limited to, steps A11 to A13.

[0076] In step A11, the obstacle position corresponding to the second-level reference time is predicted based on the vehicle's environmental data. The reference time is determined based on the current time and the time step.

[0077] For example, vehicle environmental data refers to perceived data of the vehicle's surrounding driving environment, used to predict the location of obstacles at subsequent reference times. The vehicle can be equipped with onboard cameras and LiDAR. For instance, the vehicle can be equipped with six surround-view high-definition cameras and one mechanical LiDAR.

[0078] The first environmental data, which is image data, is acquired by multiple sets of vehicle-mounted cameras installed at preset locations around the vehicle. These vehicle-mounted cameras include, but are not limited to, front-view cameras, side-view cameras, rear-view cameras, and surround-view cameras. Each camera acquires images synchronously at a preset frame rate. During image acquisition, the parameters of the vehicle-mounted cameras are automatically adjusted to adapt to different lighting conditions (e.g., strong light, weak light, backlight) and weather conditions (e.g., rainy days, foggy days).

[0079] A second environmental data point cloud can be collected by a LiDAR system installed on a vehicle. This LiDAR includes multi-line LiDAR systems installed at key locations such as the vehicle's roof, front, rear, and sides to perform comprehensive scanning of the vehicle's surroundings. The LiDAR emits laser beams and receives the echo signals reflected back from target objects, acquiring the three-dimensional spatial coordinates, reflection intensity, and distance information of each sampling point in the surrounding environment, forming a point cloud containing a large number of discrete spatial points. This point cloud data can characterize the three-dimensional contours, sizes, positions, and spatial distribution of targets such as pedestrians, other vehicles, non-motorized vehicles, obstacles, traffic signs, and road edges around the vehicle.

[0080] The vehicle is also equipped with a Global Positioning System (GPS) and an Inertial Measurement Unit (IMU) to collect the vehicle's real-time position, attitude, and motion information.

[0081] By connecting GPS with vehicle-mounted cameras and LiDAR through hardware synchronization, and using GPS timestamps as a reference for hardware triggering, the consistency of environmental data and real-time vehicle information across time is ensured. Synchronizing GPS and environmental data ensures consistency in the timing of data collection, providing a reliable data foundation for subsequent data processing and thus improving the accuracy of obstacle location.

[0082] Optionally, after acquiring environmental data, the environmental data can be preprocessed to obtain preprocessed environmental data. Preprocessing includes, but is not limited to, distortion correction, point cloud denoising, ground segmentation, and mosaic data augmentation.

[0083] Distortion correction processing primarily targets the primary environmental data captured by vehicle-mounted cameras. Due to the optical characteristics of vehicle-mounted camera lenses and installation deviations, the original captured images are prone to optical distortion problems such as radial and tangential distortion, causing the position and contour of target objects in the image to shift and become distorted. Distortion correction processing can correct the distortion deviations of the original image, restore the realistic scene imaging effect, ensure the accuracy of the position and size information of target objects such as obstacles, road edges, and traffic signs in the image, and improve the acquisition accuracy of image environmental data.

[0084] Point cloud denoising primarily targets point cloud-based secondary environmental data acquired by LiDAR. During the scanning and acquisition of the surrounding environment, LiDAR is susceptible to interference from lighting, weather, long-distance scanning errors, and stray environmental reflections, resulting in a large number of discrete, disordered, and meaningless noise points and isolated points in the original point cloud data. Point cloud denoising effectively filters out various invalid noise points, retaining only valid point clouds that accurately represent the vehicle's surrounding environment, and avoiding interference from noise points in obstacle location identification and environmental contour perception.

[0085] Ground segmentation processing is used for LiDAR point cloud data or image data. Let's take point cloud data as an example. Raw point cloud data contains a large number of road surface points and target point clouds of various obstacles and traffic participants. Ground point clouds consume significant computing power and interfere with obstacle recognition. Ground segmentation processing can accurately distinguish and separate ground point clouds from non-ground target point clouds in the point cloud data, eliminating redundant ground environment information and retaining only the effective non-ground point cloud data such as vehicles, pedestrians, and road obstacles, simplifying the computational dimensions for subsequent environmental perception and obstacle prediction.

[0086] Mosaic data augmentation is used to enhance and optimize the collected image and point cloud environmental data. This processing method can stitch together and randomly transform multiple sets of original environmental data under different scenes, lighting conditions, and weather conditions to generate rich composite environmental data samples. This expands the scene coverage of environmental data, compensates for the limitations of single-scene data collection, and improves the environmental adaptability and robustness of subsequent obstacle location prediction and driving status judgment.

[0087] After obtaining the preprocessed environmental data, the obstacle position corresponding to the reference time at the second level can be predicted based on the preprocessed environmental data. For example, the process of determining the obstacle position includes: extracting features from the preprocessed environmental data to obtain image semantic features and point cloud features; spatially aligning and fusing the image semantic features and point cloud geometric features to obtain environmental features; and using the environmental features to predict the obstacle position corresponding to the reference time.

[0088] For example, a deep convolutional network (such as ResNet-50) is used to extract the semantic features of the preprocessed image data, and a three-dimensional (3D) sparse convolutional network is used to extract features from the preprocessed point cloud environment data to obtain the point cloud geometric features. Then, the image semantic features and point cloud features are uniformly mapped to the BEV space through depth prediction and BEV (Bird's Eye View) pooling.

[0089] Among them, depth prediction is used to infer the depth distribution of the scene based on the image pixel information, obtain the real spatial distance information corresponding to each pixel in the image, and establish the correspondence between two-dimensional image pixels and three-dimensional spatial positions. Then, combined with BEV pooling operation, the image semantic features carrying depth information are transformed from image view projection to top view features, and the feature grid arrangement is completed through spatial pooling aggregation. Finally, all image semantic features are uniformly mapped to the bird's-eye view spatial coordinate system consistent with the point cloud features, realizing the spatial dimension unification of image features and LiDAR features.

[0090] Optionally, during the process of unifying spatial dimensions, sparse areas of point cloud features can be supplemented using an adaptive filling method. Point cloud features generated by LiDAR acquisition suffer from sparsity issues, such as uneven spatial distribution and missing data in local areas. For the geometric features of the point cloud projected onto the bird's-eye view space, appropriate filling rules and filling data can be dynamically selected based on the distribution patterns, numerical features, and spatial correlations of surrounding effective features. This process fills in sparse feature areas and areas with missing data, ensuring the integrity and continuity of the overall feature map in the spatial dimension.

[0091] After unifying the spatial dimensions of image semantic features and point cloud geometric features and completing sparse region filling, deformable convolution is used to perform precise spatial alignment and channel-level fusion processing on the two types of features. Deformable convolution can adaptively adjust the sampling position and receptive field range according to the actual spatial shape of targets such as obstacles and road boundaries, adapting to irregular target contour features and avoiding feature shift and detail loss caused by fixed convolution sampling. Based on precise spatial alignment, channel-level feature concatenation and weight fusion are performed on image semantic features and point cloud geometric features, adaptively assigning weights to the two types of features, preserving rich semantic category information of the image and three-dimensional geometric position information of the point cloud, and fusing to obtain global environmental features that combine semantic integrity and spatial accuracy.

[0092] After acquiring global environmental features, temporal motion features of obstacles are extracted based on multi-frame global environmental features in a continuous time sequence. The historical displacement trend, velocity, and direction of movement of the obstacles are learned to construct trajectory constraints. Combining the obstacle state features at the current moment with a preset time step, the subsequent motion state of the obstacles is iteratively extrapolated and trajectory fitted to predict, obtaining the continuous motion trajectory of the obstacles in the future time period. Based on the fitted obstacle motion trajectory, the position of the obstacles at the reference time corresponding to the second level is obtained.

[0093] In step A12, the risk coefficient of each effective node in the second level is determined using the obstacle position, the expected collision time, and the vehicle position, where the vehicle position is the position indicated by the effective node corresponding to the second driving action.

[0094] In the exemplary embodiments of this application, the estimated collision time refers to the estimated duration of the collision between the vehicle and the obstacle, calculated by combining the vehicle's driving state and the predicted obstacle position. The vehicle position refers to the vehicle's planned driving position by the second-level effective node.

[0095] The process of determining the risk coefficient of each effective node in the second level using obstacle location, expected collision time, and vehicle location includes: calculating the time risk item and spatial risk item corresponding to a single node respectively; and combining the time risk item and spatial risk item with preset weight coefficients to obtain the risk coefficient corresponding to the node.

[0096] For example, the time dimension risk is calculated based on the estimated collision time corresponding to each effective node at the current moment, and the spatial dimension risk is calculated based on the lateral and longitudinal distance between the vehicle position and the obstacle position. Then, the time risk weight coefficient and the spatial risk weight coefficient are configured respectively, and the calculation results of the two types of risks are weighted and calculated to obtain the overall risk coefficient of each effective node in the second level, thereby characterizing the degree of danger of the driving action corresponding to each node.

[0097] For example, the risk coefficient corresponding to the i-th effective node in the second level at time t. It can be calculated using formula (1): Formula (1) In the above formula (1), This is the time risk weighting coefficient, used to characterize the degree of impact of the estimated collision time on the overall risk; This is the spatial risk weighting coefficient, used to characterize the degree of influence of the relative position of the vehicle obstacle on the overall risk; Let be the estimated collision time corresponding to the i-th valid node. This is a time safety margin threshold used to distinguish between safe and dangerous driving conditions; Let be the lateral relative distance between the vehicle and the obstacle at the i-th valid node; Let be the longitudinal relative distance between the vehicle and the obstacle at the i-th valid node; and These are the standard deviations of the Gaussian distribution corresponding to the horizontal and vertical distances, respectively, used to constrain the decay law of risk values ​​due to spatial distance.

[0098] In step A13, valid nodes with risk coefficients greater than or equal to the risk threshold are identified as risk nodes.

[0099] In the exemplary embodiment of this application, the risk coefficients corresponding to all valid nodes in the second level are traversed, and valid nodes with risk coefficients greater than or equal to the risk threshold are selected and identified as risk nodes. The risk threshold can be set based on actual driving conditions, and this application does not impose any restrictions on it.

[0100] This application's embodiments predict obstacle locations based on vehicle environmental data. By combining obstacle location, estimated collision time, and vehicle position, the risk coefficient of each effective node is calculated. Risk nodes are then selected based on risk thresholds. This allows for the quantification of the safety level of each driving node layer by layer according to a time-series process, identifying nodes with potential safety hazards and proactively detecting potential driving risks. Furthermore, subsequent abnormal branch judgment and processing are performed on the selected risk nodes, eliminating the need for complex calculations on all nodes, reducing unnecessary data processing, and reasonably conserving equipment computing resources. This improves overall data processing efficiency while ensuring vehicle driving safety.

[0101] In the exemplary embodiments of this application, an invalid action node is a node in which the driving action cannot be executed normally. The process of determining an invalid action node includes, but is not limited to, steps A14 to A15.

[0102] In step A14, a reference cost is determined using the actual cost and predicted cost corresponding to multiple driving action sequences. The actual cost is the driving cost generated by extrapolating from the driving action sequence to the target driving action, and the predicted cost is the expected driving cost extrapolated from the driving action after the target driving action in the driving action sequence.

[0103] During the process of a vehicle performing corresponding driving actions along a driving action sequence, driving costs may arise due to factors such as changes in driving state, path deviation, and changes in driving time. Driving cost refers to the comprehensive quantitative indicator consumed by the vehicle in the process of extrapolating from a specified driving action sequence to the target driving action. For example, driving cost can characterize the comprehensive quantitative result of multiple indicators such as vehicle driving energy consumption, path deviation degree, driving smoothness loss, and driving time, and is used to evaluate the execution quality of a single node path.

[0104] After determining the sequence of driving actions, the actual cost corresponding to the sequence can be determined by deducing the complete execution process from the initial driving action to the target driving action. The actual cost refers to the total driving cost actually incurred during the process of deducing the first driving action along the sequence to the target driving action.

[0105] For example, traverse all driving actions in the current driving action sequence from the starting driving action to the target driving action, collect energy consumption, path deviation, driving time and other sub-quantitative data for each driving action, and sum them up to obtain the historical total driving cost of the simulated road segment, that is, the actual cost.

[0106] After determining the actual cost, the predicted cost can be determined by extrapolating the trend of the subsequent driving sequence from the target driving action. The predicted cost refers to the total estimated driving cost that will occur as the target driving action continues into the subsequent driving actions in the driving sequence.

[0107] For example, the state corresponding to the target driving action is input into a lightweight multilayer perceptron (MLP) value network trained offline on a simulation platform. This value network combines inference rules and sample patterns to complete the calculation and outputs the expected minimum driving cost from the target driving action to the end of the driving action, thus obtaining the predicted cost.

[0108] The reference cost for a single driving action sequence is obtained by summing the actual cost and the predicted cost. A driving action sequence can be mapped to a node path in a tree structure, and each driving action can be mapped to a valid node that makes up that path.

[0109] For example, the reference cost can be calculated by combining the tree structure and the following formula (2): Formula (2) In the above formula (2), Indicates including nodes The reference cost corresponding to the node path; Indicates the state from the starting node of the path Step-by-step deduction to the target effective node The actual cost accumulated during the process is the sum of the actual travel costs obtained from the actual statistics of the completed road segments of the route; For lightweight multilayer perceptron (MLP) value networks targeting effective nodes The predicted cost of the calculated output is the cost from the target effective node. Continue extrapolating to the expected minimum travel cost required to reach the final node of the entire path; These are the network parameters of a lightweight multilayer perceptron value network, which are determined after training on offline data from a simulation platform.

[0110] In step A15, valid nodes with reference costs greater than or equal to the cost threshold are identified as invalid action nodes.

[0111] In the exemplary embodiment of this application, the cost threshold is a preset critical quantification standard used to evaluate the value of node path deduction, and is used to distinguish between high-quality feasible node paths and redundant and inefficient node paths. After calculating the reference cost corresponding to each level of effective node, the reference cost of each intermediate effective node is compared with the preset cost threshold one by one. If the reference cost corresponding to the effective node is greater than or equal to the cost threshold, it is determined that the comprehensive driving cost of the branch path to which the effective node belongs is too high and the path deduction benefit is extremely low, and a feasible vehicle driving trajectory cannot be generated. Therefore, this type of effective node is marked as an invalid action node.

[0112] Optionally, the cost threshold can be the cost of the global complete trajectory ( The sum of the global complete trajectory cost (δ) and the relaxation tolerance (δ). The global complete trajectory cost is the minimum driving cost corresponding to the current complete driving trajectory obtained through real-time system search; the relaxation tolerance is a pre-calibrated small tolerance value used to avoid excessive pruning leading to the loss of potential optimal paths, ensuring the robustness and completeness of trajectory search. The relaxation tolerance can be set based on the actual driving environment scenario or dynamically and adaptively based on the vehicle's driving parameters; this application does not impose any restrictions on this.

[0113] For example, for low-speed driving scenarios in the park, the preset relaxation tolerance can be set to a fixed small value; for complex urban road conditions, the value of the relaxation tolerance can be dynamically adjusted based on the number of obstacles and vehicle speed perceived in real time.

[0114] This application embodiment calculates the reference cost by jointly using the actual cost generated during the real-world deduction process of the driving action sequence and the predicted cost estimated by the subsequent deduction of the target driving action. This approach takes into account both the actual losses of the already driven road segments and the expected losses of the undeduced road segments, quantifying the comprehensive deduction value of each driving action sequence and avoiding the bias of judgment caused by a single cost assessment. By filtering the reference costs of each effective node through a cost threshold, invalid action nodes with excessively high reference costs and low deduction value are obtained, eliminating inefficient and redundant driving branches in advance. This eliminates the need for indiscriminate traversal deduction of all level nodes, reducing the amount of invalid computation in the trajectory search process. While ensuring the accuracy and feasibility of vehicle driving path planning, this approach saves onboard computing resources and improves the overall operational efficiency of vehicle driving decision-making and trajectory planning.

[0115] For example, after identifying anomalous nodes such as risk nodes and invalid action nodes, all basic-level branches are traversed, and the node types contained in each branch are checked one by one. Basic-level branches containing anomalous nodes are designated as anomalous branches, where an anomalous branch is a branch containing at least one type of anomalous node, namely risk nodes or invalid action nodes. The driving action sequence corresponding to the anomalous branch is then determined as an anomalous action sequence.

[0116] In step A2, abnormal action sequences are removed from multiple driving action sequences to obtain valid action sequences.

[0117] In the exemplary embodiments of this application, after determining the abnormal branch corresponding to the abnormal action sequence, the abnormal branch is removed from the basic level branch corresponding to the driving action sequence, and the basic level branch consisting entirely of valid nodes, without security risks and with compliant deduction costs is retained to obtain the valid level branch. The valid level is classified into valid action sequences.

[0118] In step A3, the vehicle's trajectory is determined based on the effective action sequence.

[0119] The process of determining the vehicle's trajectory based on the effective action sequence is explained in detail in step 204 and will not be repeated here.

[0120] The driving action sequence in this application embodiment can completely cover the deduction path corresponding to different combinations of driving actions, ensuring the comprehensiveness of driving scheme exploration. Furthermore, abnormal action sequences can be identified from the driving action sequence, accurately distinguishing action sequences that pose driving safety hazards or cannot be executed normally. By removing abnormal action sequences from the driving action sequence and obtaining valid action sequences, and using the valid action sequences to generate the vehicle's driving trajectory, the scope of subsequent data processing and action deduction can be reduced, unsafe and infeasible driving paths can be avoided, and resource consumption caused by redundant calculations can be reduced. Under the premise of ensuring driving safety and the feasibility of driving actions, the overall computing efficiency can be improved.

[0121] In step 204, the vehicle trajectory with consecutive time steps is determined based on multiple driving action sequences.

[0122] In the exemplary embodiment of the application, the process of determining the vehicle's driving trajectory includes, but is not limited to, steps B1 to B3.

[0123] In step B1, multiple basic driving trajectories are determined based on multiple driving action sequences.

[0124] For example, a basic driving trajectory refers to a preliminary driving path formed by sequentially connecting driving actions and corresponding vehicle states in a valid action sequence according to time steps. The process of determining multiple basic driving trajectories includes, but is not limited to, steps B11 to B13.

[0125] In step B11, the vehicle state vector corresponding to each driving action is obtained from multiple driving action sequences.

[0126] For example, a vehicle state vector is a multidimensional data set composed of multiple vehicle operating parameters. The vehicle state vector is used to quantitatively describe the overall driving state of the vehicle at a corresponding node (driving action) and time. For example, by traversing each driving action sequence, combining the current real-time vehicle state and time step extrapolation rules, the vehicle operating parameters corresponding to each driving action at future time are predicted, thus obtaining the vehicle state vector.

[0127] It should be noted that the vehicle's state vector is a state vector in the Frenet coordinate system. The Frenet coordinate system is a local motion coordinate system established with the road centerline as the reference, used to describe the vehicle's driving state along the road. The Frenet coordinate system decomposes the vehicle's motion components along the reference path with the road extension direction as the longitudinal direction and the road transverse direction as the lateral direction.

[0128] Let's take the application scenario of a tree structure as an example. For the vehicle state vector at time t corresponding to the k-th level, let St=[ s t , d t , v t , a t ,θ t , k t ],in, s t This represents the vehicle's position coordinates at time t; d t It represents the lateral offset distance at time t, which characterizes the left and right deviation of the vehicle's current position from the lane baseline (center line / boundary line). Positive numbers indicate leftward deviation, and negative numbers indicate rightward deviation, which helps determine whether the vehicle has crossed the line or deviated from the lane. v t This represents the vehicle's speed at time t. a t Let θ represent the vehicle's acceleration at time t. t This represents the vehicle's heading angle at time t. k t The curvature of the road at time t represents the location of the vehicle.

[0129] Optionally, during the hierarchical progression simulation, the Frenet coordinate system is first determined. Based on the lateral driving actions in the current driving action sequence, the corresponding road physical reference line is extracted from the global map, and the Frenet coordinate system at the current simulation moment is constructed using this physical reference line as the benchmark.

[0130] For example, when the lateral driving action is lane-changing 1, the center line of the left-hand opposite lane is smoothly extracted as the physical reference benchmark and the corresponding Frenet coordinate system is established; when the lateral driving action is lane-changing 2, the dividing line between the two lanes is extracted as the physical reference benchmark and the corresponding Frenet coordinate system is established.

[0131] In step B12, a trajectory segment is generated using the vehicle state vector corresponding to the driving action.

[0132] In the exemplary embodiment of this application, for any two adjacent driving actions in the driving action sequence, the vehicle state vectors corresponding to each adjacent driving action are obtained; the preset lateral target offset corresponding to the two driving actions is matched according to the two driving actions; the temporal continuous fitting processing is completed based on the vehicle state vectors of the adjacent driving actions, so that the vehicle driving state smoothly transitions to the target offset state of the corresponding driving action according to the temporal relationship of the driving action sequence, and the trajectory segment between adjacent driving actions is generated.

[0133] For example, adjacent driving actions are the driving actions corresponding to effective nodes at adjacent levels in the tree structure. Taking the vehicle state of the previous level in the adjacent level as the initial starting point, a pure path tracking algorithm is used to calculate the expected front wheel angle required for the vehicle to smoothly enter and conform to the target reference path, thereby achieving precise vehicle heading control; the longitudinal driving action corresponding to the current branch is executed, and the corresponding physical acceleration is output. Within the time step corresponding to a single level, numerical integration is performed using time resolution. Using the predicted trajectories of obstacles such as pedestrians and vehicles, the future coordinates of obstacles at the level are obtained, and the longitudinal driving speed of the vehicle is determined by combining it with the Intelligent Driver Model (IDM) model. The solved lateral and longitudinal control parameters are substituted into the vehicle's two-degree-of-freedom kinematic model to deduce and generate a continuous and smooth set of microscopic trajectory points within the level interval, thus obtaining the basic shape of the complete driving path between adjacent nodes.

[0134] After completing the trajectory basic shape deduction, the preset lateral target offset corresponding to the driving action is matched according to the driving action between two effective nodes; the temporal continuous fitting process is completed based on the vehicle state vector of two adjacent effective nodes, so that the vehicle driving state smoothly transitions to the target offset state of the corresponding driving action with the hierarchical progression relationship, and finally the trajectory segment between adjacent effective nodes is generated.

[0135] For example, lateral driving actions include, but are not limited to, lane keeping, changing lanes to the left (lane borrowing 1), driving over the lane line (lane borrowing 2), and lane return. If the action between adjacent valid nodes is a straight-line keeping action or a lane return action, the matched lateral target offset is 0.0. By continuously fitting the vehicle state vectors of the preceding and following valid nodes, a trajectory segment in which the vehicle drives smoothly along the lane centerline is generated. If the action between adjacent valid nodes is lane borrowing 1, the matched lateral target offset is W (current lane width). By combining the vehicle state vectors of the preceding and following valid nodes, a state transition is completed, generating a trajectory segment in which the vehicle completely enters the adjacent lane on the left. If the action between adjacent valid nodes is lane borrowing 2, the matched lateral target offset is 0.5W. By smoothly connecting the driving states with the vehicle state vectors of the preceding and following valid nodes, a trajectory segment in which the vehicle drives along the two lane dividing line is generated.

[0136] In step B13, the various trajectory segments are sequentially spliced ​​together based on the driving action sequence indicated by the driving action sequence to obtain multiple basic driving trajectories.

[0137] For example, the sequence of driving actions forms a temporal deduction logic according to the order of driving actions, and two adjacent driving actions correspond to an independent and continuous trajectory segment. After obtaining the trajectory segments corresponding to all adjacent driving actions, according to the order of driving actions indicated by the driving action sequence, the multiple independent trajectory segments are aligned end to end and spliced ​​together step by step to obtain a complete and temporally coherent driving path, i.e., the basic driving trajectory.

[0138] This application embodiment can obtain the basic driving data corresponding to each driving action by using the vehicle state vector corresponding to the driving action in the driving action sequence; by generating trajectory segments using the vehicle state vector corresponding to the driving action, each path can be adapted to the preset driving behavior, ensuring the rationality of the local driving pattern; by splicing all trajectory segments according to the driving action sequence to form multiple basic driving trajectories, the smooth connection of segmented paths can be achieved, constructing a complete basic driving trajectory with temporal continuity, providing a data foundation for subsequent determination of vehicle driving trajectory.

[0139] In step B2, each basic driving trajectory is verified to obtain the verification results.

[0140] In the exemplary embodiment of this application, after obtaining the basic driving trajectory, a security check and a driving efficiency check can be performed on each basic driving trajectory to obtain the check results. The check results include the security check results and the driving efficiency check results.

[0141] The safety verification results are used to characterize whether there are safety issues such as collision risk, excessive lane departure, and abnormal driving posture during the entire simulation of the basic driving trajectory. By verifying the compliance of the driving boundary, the obstacle avoidance effect, and the driving status safety of the basic driving trajectory, it is determined whether a single basic driving trajectory meets the vehicle's safe driving constraints.

[0142] The driving efficiency verification results are used to characterize the overall driving time, driving smoothness and path efficiency corresponding to the basic driving trajectory. By verifying the driving rhythm, path redundancy and motion stability of the basic driving trajectory, it is determined whether a single basic driving trajectory meets the driving constraints for efficient vehicle passage.

[0143] In an exemplary embodiment of this application, a Responsibility Sensitive Safety (RSS) model is used to perform RSS safety verification on the basic driving trajectory to obtain safety verification results. The safety verification results include at least one of lateral safety distance verification results, longitudinal safety distance verification results, or braking distance verification results. The following will describe each type of safety verification result: (1) The lateral safety distance verification result is used to characterize the degree of safety matching between the vehicle and surrounding obstacles and lane boundaries in the lateral motion dimension. The lateral safety distance verification result is determined based on the vehicle driving posture corresponding to the effective nodes in the basic driving trajectory and the lateral relative distance between the vehicle and surrounding obstacles.

[0144] Vehicle driving attitude refers to the current heading angle, body yaw state, and other driving characteristics of the vehicle. By extracting relevant parameters from the vehicle state vector corresponding to each valid node, the vehicle driving attitude can be obtained.

[0145] The lateral relative distance between a vehicle and surrounding obstacles refers to the straight-line distance between the vehicle body and surrounding obstacles in the lateral direction of the lane in the Frenet coordinate system, which can be calculated from the position of the obstacle and the position of the vehicle.

[0146] The extracted vehicle driving posture and the lateral relative distance between the vehicle and surrounding obstacles are used as input data. The pre-set lateral safety threshold and lateral motion constraint rules in the RSS model are invoked. Each valid node corresponding to a driving action within the basic driving trajectory is traversed, and the actual lateral relative distance is compared with the safety threshold, and the vehicle driving posture is verified to meet the lateral driving constraints, completing the single-node safety compliance judgment. Finally, the judgment results of all valid nodes are summarized to obtain the lateral safety status of the entire trajectory, thus obtaining the lateral safety distance verification results for each basic driving trajectory. The safety threshold and lateral driving constraints can be determined based on the actual driving situation of the vehicle; this application does not impose any restrictions on this.

[0147] (2) The longitudinal safety distance verification result is used to characterize whether the distance between the vehicle and obstacles in front and behind in the direction of lane extension meets the driving safety requirements. The longitudinal safety distance verification result is determined based on the vehicle speed corresponding to the effective nodes in the basic driving trajectory and the longitudinal relative distance between the vehicle and the surrounding obstacles.

[0148] Vehicle speed can be directly extracted from the vehicle state vector corresponding to each valid node; the longitudinal relative distance between the vehicle and surrounding obstacles refers to the straight-line distance between the vehicle and surrounding obstacles along the lane extension direction in the Frenet coordinate system, calculated by combining the obstacle position and the vehicle position. These two types of data are input into the RSS model, and compliance judgments are made for each valid node corresponding to each driving action, against the model's preset longitudinal safety distance threshold and following constraint rules. The judgments of all nodes are summarized to obtain the longitudinal safety status of the entire trajectory, thereby generating the longitudinal safety distance verification result. The longitudinal safety distance threshold and following constraint rules can be determined based on the actual driving situation of the vehicle; this application does not impose any restrictions on this.

[0149] (3) The braking distance verification result is used to characterize whether the required sliding distance of the vehicle during emergency braking at the current driving speed meets the safety standards. The braking distance verification result is determined based on the vehicle speed corresponding to the valid nodes in the basic driving trajectory.

[0150] Vehicle speed is extracted from the vehicle state vector corresponding to each valid node. The speed parameter is input into the RSS model, and the theoretical braking distance is calculated by combining the vehicle braking performance parameters built into the model. This theoretical braking distance is then compared with a preset safe braking distance threshold to complete the verification and judgment of the nodes corresponding to the driving action. The verification results of all nodes are integrated to finally obtain the braking distance verification result. The safe braking distance threshold can be determined based on the actual driving conditions of the vehicle, and this application does not impose any restrictions on it.

[0151] This application embodiment generates at least one verification result among lateral safety distance verification result, longitudinal safety distance verification result, and braking distance verification result by combining multiple types of vehicle parameters corresponding to the effective nodes of the basic driving trajectory and the relative distance to obstacles. It can perform multi-dimensional safety verification on the basic driving trajectory from multiple dimensions such as vehicle lateral driving safety, longitudinal following safety, and braking driving safety, covering various safety risk scenarios during vehicle driving. This can improve the safety of determining the vehicle driving trajectory based on the verification results in subsequent processes.

[0152] The driving efficiency is verified on the basic driving trajectory to obtain the driving efficiency verification result. The driving efficiency verification result is determined based on at least one of the following: speed cost, acceleration cost, lateral offset cost, safety distance cost, heading angle change cost, and rule cost corresponding to the effective nodes in the basic driving trajectory.

[0153] Speed ​​cost measures the deviation between the actual vehicle speed and the ideal travel speed. The greater the deviation, the higher the speed cost, directly reflecting the overall speed level of the trajectory. The speed cost is calculated by subtracting the vehicle speed at the effective node from the preset ideal travel speed; the larger the difference, the higher the corresponding cost.

[0154] Acceleration cost measures the severity of a vehicle's acceleration and deceleration. Greater acceleration fluctuations and more frequent starts and stops result in a higher acceleration cost, reflecting the smoothness of the ride. Acceleration cost is calculated by extracting the vehicle acceleration parameters corresponding to effective nodes and combining them with the acceleration changes of adjacent nodes. Greater parameter fluctuations correspond to higher cost values.

[0155] Lateral offset cost measures the degree of deviation of a vehicle from the lane baseline. The greater the lateral offset, the higher the lateral offset cost, reflecting the regularity of the vehicle's lane keeping. The lateral offset cost is calculated by comparing the lateral offset distance of the vehicle at the effective node with the lane baseline position; the greater the offset distance, the higher the corresponding cost.

[0156] The safety distance cost measures the degree to which the actual distance between the vehicle and surrounding obstacles deviates from the standard safety distance. The greater the deviation and the lower the safety redundancy, the higher the safety distance cost, reflecting the safety adaptation level of trajectory driving. The safety distance cost is obtained by calculating the difference between the actual relative distance between the vehicle and surrounding obstacles and the preset standard safety distance. The larger the deviation value, the higher the corresponding cost.

[0157] The heading angle change cost measures the magnitude and frequency of changes in the vehicle's direction of travel. The more pronounced the abrupt changes in the heading angle and the more frequent the angle changes, the higher the heading angle change cost, reflecting the stability of the trajectory. The heading angle change cost is obtained by extracting the heading angle parameters of adjacent valid nodes and calculating the angle change amplitude. The larger the angle change, the higher the corresponding cost.

[0158] Rule cost measures the degree to which a basic driving trajectory conforms to preset driving rules and road traffic regulations. The more violations a trajectory contains and the worse its compliance, the higher the rule cost, reflecting the compliance level of the trajectory's driving. The rule cost is determined based on the driving behavior corresponding to the valid nodes to determine whether preset traffic rules have been violated; the more violations, the higher the corresponding cost.

[0159] After obtaining the costs for each type, the total cost can be calculated. The total cost is calculated as follows: Total Cost = w1 × Velocity Cost + w2 × Acceleration Cost + w3 × Lateral Offset Cost + w4 × Safety Distance Cost + w5 × Heading Angle Change Cost + w6 × Rule Cost. Here, w1-w6 are the weighting coefficients corresponding to each type of cost. These weighting coefficients can be set based on parameters such as the actual driving scenario and road environment; this application does not impose any restrictions on this.

[0160] The calculated total cost is compared with a preset efficiency threshold. The lower the total cost, the better the overall trajectory performance and the higher the driving efficiency; the higher the total cost, the worse the overall trajectory performance and the lower the driving efficiency. Based on the comparison results, efficiency levels are assigned or whether the trajectory meets the standard, ultimately forming the corresponding driving efficiency verification result.

[0161] This application embodiment evaluates the performance based on multiple cost indicators of effective nodes corresponding to driving actions in the basic driving trajectory. It can comprehensively evaluate the trajectory performance from multiple dimensions such as driving speed, driving smoothness, lane keeping, safe distance, directional stability and traffic compliance, fully reflect the actual driving state of the basic driving trajectory, and thus obtain the driving efficiency verification result. This improves the completeness and objectivity of the driving efficiency evaluation, thereby improving the safety of determining the vehicle driving trajectory based on the verification result in subsequent processes.

[0162] In step B3, the vehicle's driving trajectory is determined from multiple basic driving trajectories based on the verification results.

[0163] In the exemplary embodiment of this application, all basic driving trajectories are comprehensively screened by combining the safety verification results and driving efficiency verification results corresponding to each basic driving trajectory. First, based on the safety verification results, basic driving trajectories that have safety hazards or do not meet driving safety constraints are eliminated. Then, for the remaining qualified trajectories, the total cost of each trajectory is compared with the driving efficiency verification results, and the basic driving trajectory with the lowest total cost is selected as the final vehicle driving trajectory.

[0164] This application embodiment generates multiple basic driving trajectories based on effective hierarchical branches, which can provide a rich selection of alternative driving paths. Then, a comprehensive verification is carried out on each basic driving trajectory to form a verification result. The safety status and driving performance of each path can be accurately screened. Based on the verification result, the vehicle driving trajectory is selected from multiple basic driving trajectories, ensuring that the final determined vehicle driving trajectory meets both safety requirements and driving performance requirements.

[0165] This application embodiment generates a tree structure with multiple levels based on continuous time steps. The nodes in each level of the tree structure correspond one-to-one with the vehicle's driving actions. Then, the first driving action of the first level is used to determine the second driving action of the second level. Combining the first and second driving actions, a valid sequence of driving actions corresponding to valid nodes across multiple levels is determined. This avoids redundant calculations caused by driving actions with invalid nodes in the valid sequence. The vehicle's driving trajectory with continuous time steps is determined based on the driving action sequence, improving the accuracy of vehicle trajectory generation, reducing invalid calculations in the trajectory generation process, and reducing the generation time of the vehicle's driving trajectory, thereby improving the real-time performance of the vehicle's driving trajectory and the safety of vehicle driving.

[0166] In the exemplary embodiments of this application, after determining the vehicle's driving trajectory, the vehicle can be controlled to drive according to the driving trajectory. During the vehicle's driving process, the vehicle's driving trajectory and driving parameters can also be optimized.

[0167] For example, the Augmented Lagrange Iterative Linear Quadratic Regulator (AL-ILQR) algorithm is used to iteratively correct the initially selected driving trajectory using vehicle kinematic constraints and driving safety constraints as optimization boundary conditions. During the optimization process, the trajectory points and driving parameters are continuously adjusted to eliminate problems such as local abrupt changes in the trajectory and unreasonable curvature, so that the entire trajectory fully conforms to the vehicle's motion characteristics and safe driving requirements, thus obtaining the final vehicle driving trajectory.

[0168] For example, a PID (Proportional-Integral-Differential) algorithm is used to optimize the vehicle's driving parameters. This algorithm employs incremental speed control logic, relying on the proportional element to quickly respond to speed deviations, the integral element to eliminate static speed drift caused by hill driving, and the derivative element to suppress speed overshoot caused by rapid acceleration. The current vehicle acceleration is superimposed with the acceleration compensation value calculated by the algorithm as the final output, smoothly tracking the desired speed corresponding to the trajectory, avoiding sudden acceleration and deceleration, and ensuring stable vehicle driving.

[0169] In an exemplary embodiment of this application, while controlling the vehicle to travel according to the generated vehicle trajectory, the vehicle trajectory can also be updated. The process of updating the vehicle trajectory includes: determining an execution trajectory segment in the vehicle trajectory, wherein the time period corresponding to the execution trajectory segment is shorter than the planned time period; and, after the vehicle completes the execution trajectory segment, obtaining the vehicle trajectory corresponding to the updated planned time period, wherein the updated planned time period is determined based on the planned time period and the time period corresponding to the execution trajectory segment.

[0170] For example, an execution trajectory segment refers to the segment of trajectory that a vehicle actually executes from the current moment, following a planned trajectory. For instance, if the planned time period is 7 seconds, the time period corresponding to the execution trajectory segment is 1 second.

[0171] Once the vehicle has fully executed the current execution trajectory segment, the original planned time period is updated by rolling translation. The completed time periods within the original planned time period are discarded, and the current driving time is taken as the new planning start time. The planned time period is then extended forward to obtain the updated planned time period. Based on the updated planned time period, steps 201 to 204 are reused to iteratively generate a vehicle driving trajectory that is adapted to the current driving conditions, i.e., the updated driving trajectory.

[0172] For example, when the execution trajectory segment corresponding to 1 second of vehicle travel is completed, the planned time period is updated by backward translation, the 1-second period of completed travel is removed, and the updated planned time period is adjusted to 1 to 8 seconds. Steps 201 to 204 are reused to generate vehicle travel trajectories of 1 to 8 seconds.

[0173] This application embodiment executes the driving trajectory step by step by selecting execution trajectory segments with a duration shorter than the planned time period. This ensures the stable execution of the vehicle's phased driving trajectory. After the vehicle completes the execution trajectory segment, the planned time period is updated. This allows for continuous adaptation to real-time changes in the driving environment during the vehicle's journey. The updated planned time period always matches the vehicle's current driving sequence, avoiding the problem that a fixed-duration planned time period cannot adapt to dynamic driving scenarios. This continuously outputs a driving trajectory that fits the real-time operating conditions, improving the accuracy and adaptability of the vehicle's entire driving trajectory and ensuring the stability and safety of the vehicle's driving process.

[0174] This application also provides a device for generating vehicle driving trajectory. Figure 3 This is a schematic diagram of a vehicle trajectory generation device provided in an embodiment of this application, as shown below. Figure 3 As shown, the device includes: The generation module 301 is used to generate a tree structure with multiple levels based on continuous time steps. One time step corresponds to one level, and each level contains multiple nodes. Each node corresponds to the driving action of the vehicle at the current level. The determining module 302 is used to determine the second driving action of the second level based on the first driving action of the first level in the tree structure, wherein the second level is located after the first level; The determining module 302 is also used to determine multiple driving action sequences based on the second driving action and the first driving action, the driving action sequence including the driving actions corresponding to the valid nodes at each level; The determination module 302 is also used to determine the vehicle's driving trajectory with a continuous time step based on multiple driving action sequences.

[0175] In one possible implementation, the first level includes a first sub-level and a second sub-level that are temporally adjacent, with the first sub-level preceding the second sub-level. In the first level, the first driving action of any sub-level includes a first lateral driving action, and the second driving action of the second level includes a second lateral driving action. The second lateral driving action is consistent with the first lateral driving action of the second sub-level.

[0176] In one possible implementation, the first driving action includes a first longitudinal driving action, and the second driving action includes a second longitudinal driving action, which is the same as the first longitudinal driving action.

[0177] In one possible implementation, the driving action corresponding to each node includes the vehicle's lateral driving action and longitudinal driving action. The effective nodes of the second level are the nodes that match the second lateral driving action and the second longitudinal driving action in the driving action corresponding to each node of the second level.

[0178] In one possible implementation, the determining module 302 is used to determine an abnormal action sequence that includes abnormal driving actions from multiple driving action sequences. The abnormal driving action is the driving action corresponding to an abnormal node. The abnormal node includes at least one of risk nodes or invalid action nodes at any level. Risk nodes are nodes where the driving action poses a driving safety hazard, and invalid action nodes are nodes where the driving action cannot be executed normally. The abnormal action sequence is removed from the multiple driving action sequences to obtain a valid action sequence. The vehicle driving trajectory is determined based on the valid action sequence.

[0179] In one possible implementation, the device further includes a prediction module for predicting the obstacle position corresponding to a second-level reference time based on the vehicle's environmental data, wherein the reference time is determined based on the current time and a time step. The determination module 302 is also used to determine the risk coefficient of each effective node in the second level using the obstacle position, the expected collision time, and the vehicle position, where the vehicle position is the position indicated by the effective node corresponding to the second driving action; and to determine the effective nodes with a risk coefficient greater than or equal to the risk threshold as risk nodes.

[0180] In one possible implementation, the determining module 302 is further configured to determine a reference cost using the actual cost and predicted cost corresponding to multiple driving action sequences. The actual cost is the driving cost generated by extrapolating the first driving action along the driving action sequence to the target driving action, and the predicted cost is the expected driving cost extrapolated from the driving action after the target driving action in the driving action sequence. Valid nodes with a reference cost greater than or equal to a cost threshold are determined as invalid action nodes.

[0181] In one possible implementation, the determining module 302 is used to determine multiple basic driving trajectories based on multiple driving action sequences; to perform verification on each basic driving trajectory to obtain verification results; and to determine the vehicle driving trajectory from the multiple basic driving trajectories based on the verification results.

[0182] In one possible implementation, the verification result includes at least one of the following: lateral safety distance verification result, longitudinal safety distance verification result, or braking distance verification result; The lateral safety distance verification result is determined based on the vehicle's driving posture corresponding to the valid nodes in the basic driving trajectory and the lateral relative distance between the vehicle and surrounding obstacles. The longitudinal safety distance verification result is determined based on the vehicle speed corresponding to the valid nodes in the basic driving trajectory and the longitudinal relative distance between the vehicle and surrounding obstacles. The braking distance verification result is determined based on the vehicle speed corresponding to the valid nodes in the basic driving trajectory.

[0183] In one possible implementation, the verification result includes a driving efficiency verification result, which is determined based on at least one of the following: speed cost, acceleration cost, lateral offset cost, safety distance cost, heading angle change cost, and rule cost corresponding to the valid nodes in the basic driving trajectory.

[0184] In one possible implementation, the determining module 302 is used to obtain the vehicle state vector corresponding to each driving action in multiple driving action sequences; generate trajectory segments using the vehicle state vector corresponding to the driving action; and sequentially splice the trajectory segments according to the driving action sequence to obtain multiple basic driving trajectories.

[0185] The vehicle trajectory generation device of this application generates a tree structure with multiple levels based on continuous time steps. The nodes in the tree structure correspond one-to-one with the vehicle driving actions. Then, the first driving action of the first level determines the second driving action of the second level. Combining the first and second driving actions, a valid sequence of driving actions corresponding to the valid nodes of multiple levels is determined. This avoids redundant calculations caused by driving actions with invalid nodes in the valid sequence of driving actions. The vehicle trajectory with continuous time steps is determined based on the sequence of driving actions, which improves the accuracy of vehicle trajectory generation, reduces invalid calculations in the trajectory generation process, and reduces the generation time of vehicle trajectory, thereby improving the real-time performance of vehicle trajectory and vehicle driving safety.

[0186] It should be understood that the above-described apparatus is only illustrated by the division of the functional modules described above when implementing its functions. In practical applications, the functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0187] This application embodiment also provides a vehicle, which includes a vehicle body and a vehicle control system. The vehicle control system is used to generate a vehicle driving trajectory based on the above method and control the vehicle body to drive based on the vehicle driving trajectory. The method for generating the vehicle driving trajectory has been described in steps 201 to 204 and will not be repeated here.

[0188] Figure 4This is a schematic diagram of the server structure provided in the embodiments of this application. The server 400 can vary considerably due to different configurations or performance. It may include one or more processors 401 and one or more memories 402. The one or more memories 402 store at least one piece of program code, which is loaded and executed by the one or more processors 401 to implement the vehicle trajectory generation method provided in the various method embodiments described above. Of course, the server 400 may also have wired or wireless network interfaces, a keyboard, and input / output interfaces for input and output. The server 400 may also include other components for implementing device functions, which will not be elaborated here.

[0189] In an exemplary embodiment, a computer device is also provided, the computer device including a processor and a memory, the memory storing at least one piece of program code, the at least one piece of program code being loaded and executed by the processor, so that the computer device implements a method for generating a vehicle driving trajectory.

[0190] In an exemplary embodiment, a computer-readable storage medium is also provided, which stores at least one piece of program code, which is loaded and executed by a processor to enable a computer to implement any of the above-described methods for generating vehicle driving trajectories.

[0191] Optionally, the aforementioned computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc.

[0192] In an exemplary embodiment, a computer program or computer program product is also provided, which stores at least one computer instruction, which is loaded and executed by a processor to enable the computer to implement any of the above-described methods for generating vehicle driving trajectories.

[0193] It should be noted that all relevant content involved in this application (such as tree structure, time step, first driving action, second driving action, vehicle driving trajectory, etc.) is authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0194] It should be understood that "multiple" as used in this article refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0195] The above description is merely an exemplary embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.

Claims

1. A method for generating a vehicle's driving trajectory, characterized in that, The method includes: A tree structure with multiple levels is generated based on continuous time steps. One time step corresponds to one level, and each level contains multiple nodes. Each node corresponds to the vehicle's driving action at the current level. The second driving action of the second level is determined based on the first driving action of the first level in the tree structure, and the second level is located after the first level. Multiple driving action sequences are determined based on the second driving action and the first driving action, and the driving action sequences include driving actions corresponding to valid nodes at each level; The vehicle trajectory with the continuous time step is determined based on the multiple driving action sequences.

2. The method according to claim 1, characterized in that, The first level includes a first sub-level and a second sub-level that are temporally adjacent, and the first sub-level is located before the second sub-level. In the first level, the first driving action of any sub-level includes a first lateral driving action, and the second driving action of the second level includes a second lateral driving action. The second lateral movement is consistent with the first lateral movement of the second sub-level.

3. The method according to claim 1, characterized in that, The first driving action includes a first longitudinal driving action, and the second driving action includes a second longitudinal driving action, which is the same as the first longitudinal driving action.

4. The method according to claim 1, characterized in that, The driving action corresponding to each node includes the lateral driving action and the longitudinal driving action of the vehicle. The effective node of the second level is the node that matches the second lateral driving action and the second longitudinal driving action in the driving action corresponding to each node of the second level.

5. The method according to claim 1, characterized in that, Determining the vehicle trajectory with consecutive time steps based on the multiple driving action sequences includes: Among the multiple driving action sequences, an abnormal action sequence including abnormal driving actions is determined. The abnormal driving action is the driving action corresponding to an abnormal node. The abnormal node includes at least one of risk nodes or invalid action nodes at any level. The risk node is a node where the driving action poses a driving safety hazard, and the invalid action node is a node where the driving action cannot be executed normally. Remove the abnormal action sequences from the plurality of driving action sequences to obtain the valid action sequences; The vehicle's trajectory is determined based on the effective action sequence.

6. The method according to claim 5, characterized in that, The method further includes: The obstacle position corresponding to the reference time at the second level is predicted based on the vehicle's environmental data, wherein the reference time is determined based on the current time and the time step. The risk coefficient of each effective node in the second level is determined using the obstacle location, the expected collision time, and the vehicle location, where the vehicle location is the location indicated by the effective node corresponding to the second driving action. Valid nodes whose risk coefficient is greater than or equal to the risk threshold are identified as risk nodes.

7. The method according to claim 5, characterized in that, The method further includes: A reference cost is determined by using the actual cost and predicted cost corresponding to the multiple driving action sequences. The actual cost is the driving cost generated by extrapolating the first driving action to the target driving action along the driving action sequence. The predicted cost is the expected driving cost extrapolated from the driving actions after the target driving action in the driving action sequence. Valid nodes whose reference cost is greater than or equal to the cost threshold are identified as invalid action nodes.

8. The method according to any one of claims 1 to 7, characterized in that, Determining the vehicle trajectory with consecutive time steps based on the multiple driving action sequences includes: Multiple basic driving trajectories are determined based on the multiple driving action sequences; Perform verification on each basic driving trajectory and obtain the verification results; The vehicle's driving trajectory is determined from the multiple basic driving trajectories based on the verification results.

9. The method according to claim 8, characterized in that, The verification result includes at least one of the following: lateral safety distance verification result, longitudinal safety distance verification result, or braking distance verification result; The lateral safety distance verification result is determined based on the vehicle driving posture corresponding to the valid nodes in the basic driving trajectory and the lateral relative distance between the vehicle and surrounding obstacles. The longitudinal safety distance verification result is determined based on the vehicle speed corresponding to the valid nodes in the basic driving trajectory and the longitudinal relative distance between the vehicle and the surrounding obstacles. The braking distance verification result is determined based on the vehicle speed corresponding to the valid nodes in the basic driving trajectory.

10. The method according to claim 8, characterized in that, The verification results include driving efficiency verification results, which are determined based on at least one of the following: speed cost, acceleration cost, lateral offset cost, safety distance cost, heading angle change cost, and rule cost corresponding to the effective nodes in the basic driving trajectory.

11. The method according to claim 8, characterized in that, The determination of multiple basic driving trajectories based on the multiple driving action sequences includes: Obtain the vehicle state vector corresponding to each driving action from the multiple driving action sequences; A trajectory segment is generated using the vehicle state vector corresponding to the driving action; The various trajectory segments are sequentially spliced ​​together based on the driving action sequence indicated by the driving action sequence to obtain the multiple basic driving trajectories.

12. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one piece of program code, the at least one piece of program code being loaded and executed by the processor to enable the computer device to implement the method for generating vehicle driving trajectories as described in any one of claims 1 to 11.

13. A vehicle, characterized in that, The vehicle includes: a vehicle body and a vehicle control system; The vehicle control system is used to generate a vehicle driving trajectory based on any one of the methods of claims 1 to 11, and to control the vehicle body to drive based on the vehicle driving trajectory.

Images