Vehicle automatic turning method, device, medium and product

Abstract

The application relates to the technical field of vehicle control, in particular to a vehicle automatic turning method, device, medium and product. According to a current multi-attribute grid map and a preset action primitive library, at least one candidate action sequence is generated. A reversible space-time occupation envelope corresponding to each candidate action sequence is calculated. The reversible space-time occupation envelope corresponding to each candidate action sequence is intersected with a dynamic obstacle prediction track for analysis. A target action sequence is determined based on the intersection analysis result, and automatic turning is performed according to the target action sequence. The scheme breaks through the limitations of traditional static space feasibility judgment and fixed action sequence, and improves the safety and availability of the automatic turning function.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and in particular to a method, device, medium and product for automatic vehicle turning. Background Technology

[0002] In automatic U-turn scenarios, vehicles need to complete large-angle attitude changes within a short distance. Existing automatic U-turn implementation methods mainly include static space feasibility assessment, preset path tracking based on high-precision maps, fixed U-turn action sequences, and real-time obstacle avoidance. These methods rely on sensors such as surround view and ultrasonic sensors to collect environmental information, complete spatial determination and path generation, and achieve vehicle attitude reversal.

[0003] However, using static drivable space or road boundaries as criteria to determine the feasibility of a U-turn does not take into account the differences in the area swept by the vehicle during forward and reverse phases, nor does it consider the temporal conflicts of the predicted trajectory of dynamic obstacles, which can easily lead to misjudgments or conservative exits in path planning. Summary of the Invention

[0004] This application provides a method, device, medium, and product for automatic vehicle U-turns to solve the aforementioned technical problems. This solution addresses the misjudgment problem of traditional static spatial judgment in complex scenarios by constructing a reversible spatiotemporal occupancy envelope and performing temporal conflict analysis on the predicted trajectory of dynamic obstacles.

[0005] In a first aspect, embodiments of this application provide a method for automatic vehicle U-turn, the method comprising:

[0006] Generate at least one candidate action sequence based on the current multi-attribute raster map and the preset action primitive library;

[0007] Calculate the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence;

[0008] The intersection analysis is performed between the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence and the predicted trajectory of dynamic obstacles.

[0009] The target action sequence is determined based on the intersection analysis results, and an automatic U-turn is executed according to the target action sequence.

[0010] In one possible embodiment, generating a candidate action sequence includes:

[0011] A pre-trained action graph model is used to search for at least one candidate action sequence by starting with the current vehicle pose as the graph node and using action primitives as edges. The graph nodes include vehicle gear, blind spot level, and recoverable markers. The candidate action sequence includes at least one of the following: single U-turn sequence, three-point U-turn sequence, and multi-point U-turn sequence.

[0012] In one possible embodiment, it also includes:

[0013] Generate a driveable area for U-turns based on the current multi-attribute raster map;

[0014] At least one candidate target pose is generated within the drivable area for U-turns according to a preset adaptation rule, so that when searching, the candidate target pose is used as the termination graph node and the drivable area for U-turns is used as the boundary to generate a candidate action sequence.

[0015] In one possible embodiment, calculating the reversible spatiotemporal occupancy envelope includes:

[0016] Predict the vehicle pose at each time step for each action in the candidate action sequence;

[0017] The sweep envelope of the vehicle pose at each moment is calculated using a preset dilation rule corresponding to the action type. The sweep envelope includes the safety dilation distance.

[0018] In one possible embodiment, the preset dilation rule includes the region to be dilated corresponding to each action type, and the calculation of the sweep envelope includes:

[0019] Calculate the safe expansion distance based on the confidence level of the grid where the current pose is located, the grid blind zone level, the speed of dynamic obstacles, and the action type;

[0020] The region to be expanded is expanded according to the action type and the safe expansion distance to generate a sweep envelope.

[0021] In one possible embodiment, the intersection analysis includes:

[0022] Calculate the spatiotemporal overlap point between the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence and the predicted trajectory of the dynamic obstacle at the same moment;

[0023] The temporal conflict cost of generating each candidate action sequence based on spatiotemporal overlap points.

[0024] In one possible embodiment, generating the timing conflict cost includes:

[0025] If the spatiotemporal overlap point leaves the envelope region before the vehicle arrives in the predicted trajectory, the cost of temporal conflict is reduced.

[0026] If there are no spatiotemporal overlap points at present, but the predicted trajectory overlaps with the sweep envelope at a future time, the cost of temporal conflict will increase.

[0027] In one possible embodiment, determining the target action sequence includes:

[0028] The total cost corresponding to each candidate action sequence is calculated based on static collision cost, temporal conflict cost, blind zone cost, boundary cost, number of gear shifts, curvature change, and target pose error.

[0029] The target action sequence is determined based on the total cost corresponding to each candidate action sequence.

[0030] In one possible embodiment, it also includes:

[0031] If the total cost corresponding to each candidate action sequence does not meet the threshold of the target action sequence, and the recoverable flag in any candidate action sequence is true, then at least one probe action is generated.

[0032] The reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence is updated based on the perception results corresponding to the detection action.

[0033] In one possible embodiment, it also includes:

[0034] If the blind zone cost of the target action sequence is greater than a preset cost threshold, then at least one detection action is generated.

[0035] When the perception result corresponding to the detection action does not meet the execution conditions, the target action sequence is adjusted.

[0036] In one possible embodiment, performing an automatic U-turn includes:

[0037] The target action sequence is divided into multiple phases to be executed, and each phase to be executed includes constraints.

[0038] During the execution of each phase to be executed, the real-time pose of the vehicle is acquired.

[0039] In one possible embodiment, it also includes:

[0040] Based on the vehicle's real-time pose, the current multi-attribute raster map and reversible spatiotemporal occupancy envelope are periodically updated;

[0041] When a preset abnormal condition is triggered and the vehicle comes to a stop, at least one candidate action sequence is regenerated based on the updated multi-attribute raster map.

[0042] In one possible embodiment, it also includes:

[0043] In response to receiving an automatic U-turn command, it acquires multi-source sensor data, including visual perception data, ultrasonic perception data, and current vehicle status data;

[0044] The current multi-attribute raster map is generated based on multi-source sensor data. The current multi-attribute raster map includes raster coordinates, obstacle occupancy probability, obstacle height, obstacle speed, confidence level, and blind spot level.

[0045] In one possible embodiment, the preset action primitive library includes basic action primitives and probe action primitives;

[0046] Each action primitive includes direction, curvature range, steering angular velocity threshold, speed threshold, driving distance range, shift requirements, abort conditions, and recoverable conditions.

[0047] In one possible embodiment, it also includes:

[0048] When the vehicle travels to the target area corresponding to the target pose, check the vehicle center position error, heading angle error, distance between the vehicle body and the road boundary, driving space, gear status and obstacle conditions.

[0049] If the test fails, a corrective action sequence is generated and execution continues.

[0050] Secondly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0051] The memory stores the instructions that the computer executes;

[0052] The processor executes computer execution instructions stored in memory, causing the processor to perform the methods described above.

[0053] Thirdly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided above.

[0054] Fourthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method provided above.

[0055] The automatic vehicle U-turn method, device, medium, and product provided in this application utilize a multi-attribute raster map, which differs from traditional binary drivable areas. By incorporating environmental input, it reduces planning errors caused by missed or false detections. The action primitive library provides standardized motion primitives, enabling automatic generation of various U-turn forms based on the environment, improving adaptability to narrow and dead-end road scenarios. Calculating the reversible spatiotemporal occupancy envelope represents the spatial and temporal occupancy of vehicles during the execution of candidate action sequences. This eliminates reliance on a single static vehicle position for feasibility assessment, reflecting the vehicle's spatial occupancy throughout the entire action execution process and improving the accuracy of U-turn feasibility judgment. Intersecting the spatiotemporal occupancy of candidate vehicle actions with the predicted future trajectories of dynamic obstacles identifies whether vehicles and dynamic obstacles will spatially overlap at the same moment during the execution of candidate action sequences, providing a dynamic risk dimension assessment basis for subsequent target action sequence selection. Using the spatiotemporal intersection analysis results of dynamic obstacles as a crucial basis, the target action sequence is determined from multiple candidate action sequences. This ensures that the selected target sequence avoids future conflict risks from dynamic obstacles while reducing the likelihood of conservative withdrawal. Attached Figure Description

[0056] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0057] Figure 1 An application scenario diagram of an automatic vehicle turning method provided in this application;

[0058] Figure 2 A flowchart (I) illustrating an embodiment of the automatic U-turn method for vehicles provided in this application.

[0059] Figure 3 A schematic flowchart (II) of an embodiment of the automatic vehicle turning method provided in this application;

[0060] Figure 4 A flowchart (III) illustrating an embodiment of the automatic vehicle turning method provided in this application.

[0061] Figure 5 This is a schematic diagram of the structure of an automatic vehicle turning device provided in an embodiment of this application;

[0062] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0063] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0064] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0065] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, they do not mean that the applicant has used or necessarily used the solution.

[0066] To clearly understand the technical solution of this application, the solutions of the prior art will be described in detail first.

[0067] In automatic U-turn applications, vehicles need to perform significant attitude changes within a limited space. Current mainstream methods for achieving automatic U-turns include static drivable space determination, pre-set trajectory tracking based on high-precision maps, fixed-pattern U-turn choreography, and real-time obstacle avoidance. The system primarily uses surround-view cameras and ultrasonic sensors to collect surrounding environmental information, complete spatial condition assessment and path generation, thereby achieving vehicle attitude reversal. However, most existing solutions rely on static drivable areas and road boundaries to determine the feasibility of a U-turn, failing to fully consider the significant difference in the area swept by the vehicle during forward and reverse maneuvers, and lacking temporal conflict verification based on predicted trajectories of dynamic obstacles. This can easily lead to planning deviations, either generating paths with potential collision risks or directly determining that a U-turn is unfeasible due to slight uncertainty in scene conditions, resulting in a conservative exit strategy.

[0068] Therefore, when facing the technical problems in existing technologies, since the U-turn process involves complex movements involving alternating forward and reverse motions, a fixed single U-turn action is difficult to adapt to different spaces and obstacle distributions. Therefore, a multi-attribute raster map carrying rich environmental information is used to represent the surrounding environment. Simultaneously, action primitives are introduced as motion units to generate multiple candidate action sequences, resulting in various feasible U-turn alternatives and overcoming the limitations of fixed action sequences. Secondly, since the sweep range of the vehicle body differs during forward and reverse movements, static positions cannot reflect the risks throughout the entire movement process. By constructing a reversible spatiotemporal occupancy envelope, for each candidate action sequence, the time and space occupancy of the vehicle during the entire action execution process is represented, fully reflecting the spatial occupancy characteristics under different driving directions (forward and reverse). To identify whether dynamic obstacles will collide with the vehicle in the future, instead of using only the current position of the obstacle, an intersection analysis is performed between the reversible spatiotemporal occupancy envelope and the predicted trajectory of the dynamic obstacle to achieve temporal-dimensional conflict identification and assess the potential collision risk between each candidate sequence and the dynamic target. Finally, based on the conflict assessment results obtained from the intersection analysis, the target action sequence is selected from multiple candidate action sequences and the vehicle is controlled to execute it, thereby completing the automatic U-turn, avoiding safety hazards caused by dynamic obstacles, and reducing planning misjudgments and unnecessary conservative exit phenomena.

[0069] Figure 1 This is an application scenario diagram illustrating the automatic vehicle turning method provided in this application, such as... Figure 1 As shown in the diagram, the scenario corresponding to the automatic vehicle turning method provided in this application includes: a target vehicle 101. The target vehicle 101 integrates an automatic vehicle turning device.

[0070] Understandably, the target vehicle 101 also integrates multi-source sensors, such as ultrasonic probes, surround-view cameras, and forward-view cameras; millimeter-wave radar, lidar, inertial measurement units, etc., can also be replaced or added depending on the vehicle model.

[0071] For example, a user can initiate an automatic U-turn command by clicking "Automatic U-turn" on the vehicle's infotainment interface, or by voice or physical button. Upon receiving the command, the automatic U-turn device in the target vehicle 101 constructs a current multi-attribute grid map based on multi-source sensor data. It then generates at least one candidate action sequence based on the current multi-attribute grid map and a preset action primitive library. Next, the automatic U-turn device in the target vehicle 101 calculates the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence; it then performs an intersection analysis between the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence and the predicted trajectory of the dynamic obstacle. Finally, based on the intersection analysis results, it determines the target action sequence and controls the target vehicle 101 to perform an automatic U-turn according to the target action sequence.

[0072] The technical solution of this application and how it solves the above-mentioned technical problems will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0073] Figure 2 A flowchart (I) illustrating an embodiment of the automatic vehicle U-turn method provided in this application is shown below. Figure 2 As shown, the executing entity in this embodiment is a vehicle automatic U-turn device. This device can be implemented through a computer program, or through a medium storing the relevant computer program, such as a USB flash drive, optical disc, or chip. Alternatively, it can be implemented through an electronic device that integrates or installs the relevant computer program, such as a vehicle controller, domain controller, or vehicle edge processor. The vehicle automatic U-turn method provided in this embodiment includes the following steps:

[0074] S201. Generate at least one candidate action sequence based on the current multi-attribute raster map and the preset action primitive library.

[0075] Multi-attribute raster maps refer to environmental maps obtained by discretizing the local environment surrounding a vehicle using a grid. Unlike binary maps that only mark passable and impassable areas, each raster cell, in addition to indicating the presence of obstacles at that location, also carries multiple environmental attribute information, recording environmental dimensions such as obstacle-related features, perception credibility, and environmental risks. This map establishes a local coordinate system with the vehicle itself as a reference and updates in real time as the vehicle moves, without relying on high-precision maps to provide environmental information.

[0076] The preset action primitive library refers to a collection of pre-configured basic motion units used to combine and splice together a complete U-turn action. An action primitive is the smallest indivisible motion unit that completes a U-turn. Each action primitive defines a type of basic vehicle motion behavior, such as forward steering, reverse steering, straight movement, and waiting in place, which are various basic actions used in the U-turn process.

[0077] In this context, a candidate action sequence refers to a complete set of continuous actions obtained by sequentially concatenating action primitives in chronological order. It represents a complete U-turn behavior scheme that can be used to achieve vehicle posture reversal. A candidate action sequence consists of one or more action primitives in sequence. Different candidate action sequences represent different U-turn execution strategies, which can correspond to different U-turn forms such as a direct U-turn, a three-point U-turn, or a multi-point reciprocating U-turn.

[0078] Understandably, users can initiate a U-turn by clicking the automatic U-turn command on the vehicle's infotainment system interface, or by voice, physical buttons, navigation routes, recognition of the road ahead ending, driver's turn signal, or steering wheel movements.

[0079] Optionally, upon receiving the instruction, the system enters a pre-activation state to check prerequisites such as the status of the four doors and two hoods, gear position, braking system, steering system, power system, sensor status, vehicle speed, gradient, driver's seat belt, and surrounding traffic regulations. If any condition is not met, the system will indicate that the function is unavailable.

[0080] Optionally, the current multi-attribute raster map is formed by fusing multi-source sensing data. This sensing data can include semantic segmentation results, object detection results, road boundaries, walls, curbs, lane lines, drivable areas, pedestrians, vehicles, cones, and pillars from surround-view and front-view cameras. Ultrasonic probes output nearby echo points, echo intensity, echo duration, reflection angle, and suspicious noise markers. It can also include vehicle speed, wheel speed, yaw rate, steering angle, gear position, acceleration, and braking status.

[0081] Specifically, a real-time constructed multi-attribute raster map is acquired, which delineates the feasible space for vehicle movement, filtering out restricted areas containing obstacles or where vehicles are not permitted. Simultaneously, a pre-configured motion primitive library is read, retrieving all available basic motion units. Using the vehicle's current pose as the starting state for the search, and with motion primitives as connecting edges between states, a state expansion search is performed. For each motion primitive invoked, the new state the vehicle will reach after executing the action is deduced based on the vehicle's kinematic characteristics.

[0082] It should be noted that during the process of expanding and generating new states, constraint verification is performed using a multi-attribute grid map. If the derived vehicle state falls into a grid-marked impassable restricted area, the action expansion path is discarded and the process is stopped. Through continuous iterative expansion, when the derived vehicle state meets the requirements for the U-turn endpoint posture, the process backtracks along the state propagation relationship, organizing the sequentially passed action primitives in order of execution to obtain a candidate action sequence. For example, forward left turn to the first stop point, reverse right turn to the second stop point, forward left turn towards the target posture. If the road width is smaller, a reverse or forward phase can be automatically added to form a multi-point U-turn.

[0083] Understandably, the search process does not stop immediately after obtaining the first valid sequence. Instead, it continues to iterate and deduce, outputting multiple distinct candidate action sequences while satisfying spatial and vehicle kinematic constraints. Each candidate action sequence represents a theoretically feasible U-turn option for attitude reversal, which is then used for further risk assessment and screening in subsequent steps.

[0084] S202. Calculate the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence.

[0085] The reversible spatiotemporal occupancy envelope is a set generated for an entire candidate action sequence. This set contains the total spatial area occupied by the vehicle's movement at each discrete time point during the execution of the sequence, representing the physical area occupied by the vehicle at different times throughout the entire U-turn maneuver from start to finish. "Reversible" means that the calculation logic of this envelope is compatible with both forward and reverse driving directions in a U-turn scenario, handling the alternating forward and reverse motion. The spatiotemporal representation encompasses both time and space dimensions, representing not only the vehicle's occupied area at a single instant but also the spatial changes during the vehicle's movement over time.

[0086] Specifically, candidate action sequences are extracted one by one. Each candidate action sequence consists of multiple sequentially executed actions. For a single candidate action sequence, the following processing is performed: Each action within the sequence is decomposed individually. For a given action, the total execution time of the action is determined, and the complete execution process is divided into multiple discrete time points. Based on vehicle kinematics and considering the driving direction and steering characteristics of the action, the vehicle's posture at each discrete time point is calculated sequentially, i.e., the vehicle's position and orientation at that moment.

[0087] Furthermore, for the vehicle pose at each discrete time point, the outer boundary contour of the vehicle body at that moment is determined by combining the vehicle's own external outline, the outline of the exterior rearview mirrors, the steering angle, the vehicle width, the wheelbase, and the front and rear overhang lengths. Combining the driving direction attribute (forward or reverse) of the action, a sweep envelope corresponding to that moment is generated. The sweep envelope represents the space occupied by the vehicle when performing the action at that time point.

[0088] It is understandable that an action corresponds to a set of sweep envelopes arranged chronologically. All sweep envelopes generated by all actions in the candidate action sequence, which contain corresponding time information, are aggregated together to form the complete reversible spatiotemporal occupancy envelope of the candidate action sequence.

[0089] Furthermore, the above processing logic is repeated, and the above operation is performed independently for each remaining candidate action sequence to obtain the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence. The generated reversible spatiotemporal occupancy envelope carries both temporal and spatial information.

[0090] S203. Perform intersection analysis between the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence and the predicted trajectory of the dynamic obstacle.

[0091] Among them, the dynamic obstacle prediction trajectory refers to the future motion result of objects in the environment whose positions will change, including pedestrians, non-motorized vehicles, and moving vehicles, which is derived from the historical motion state of the obstacle and used to describe the expected spatial position of the obstacle at various future moments.

[0092] Intersection analysis involves comparing whether spatial regions overlap and verifying whether the times of overlap are the same. Only when spatial regions intersect at the same time are potential conflicts identified, thus distinguishing between risk-free scenarios where spatial regions overlap but are staggered in time.

[0093] Specifically, the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence is extracted sequentially. This reversible spatiotemporal occupancy envelope contains several sets of time-stamped sweep envelopes. Simultaneously, the predicted trajectories of all dynamic obstacles within the environment are acquired. For a currently processed candidate action sequence, each time-stamped sweep envelope within it is traversed, and the execution time corresponding to that sweep envelope and the vehicle space occupancy area it represents are read. For each predicted dynamic obstacle trajectory, the predicted spatial location of the obstacle that matches the sweep envelope time is found. That is, the spatial range in which the dynamic obstacle is expected to be located at the same future time as the sweep envelope is found.

[0094] Furthermore, the swept envelope space of the vehicle at that moment is geometrically intersected with the predicted space occupied by the dynamic obstacle at the same moment. If the two spatial ranges intersect, it indicates that the candidate action sequence has a potential risk of conflict with the dynamic obstacle at that moment. If the two spatial ranges do not intersect at the same moment, it indicates that there is no such conflict at that moment. After completing the comparison of a single time point, the swept envelopes of all other time points under the candidate sequence are processed, and the spatiotemporal matching comparison of all time points and all predicted trajectories of dynamic obstacles is completed iteratively.

[0095] Understandably, the comparison results for all moments in the candidate action sequence are compiled, and the timing and frequency of conflicts are statistically analyzed to form a complete intersection analysis result for the candidate action sequence. This process is repeated to perform the same intersection analysis on the remaining candidate action sequences, yielding the intersection analysis result for each candidate action sequence.

[0096] S204. Determine the target action sequence based on the intersection analysis results, and perform automatic U-turn according to the target action sequence.

[0097] The intersection analysis result refers to the comprehensive conclusion obtained after completing the spatiotemporal coupling comparison between the reversible spatiotemporal occupancy envelope and the predicted trajectory of dynamic obstacles. It includes risk-related information such as whether each candidate action sequence has spatiotemporal overlap with dynamic obstacles at each discrete moment, the time of conflict, and the frequency of conflict, which is used to characterize the potential collision risk of dynamic obstacles faced by each candidate action sequence.

[0098] The target action sequence refers to a set of action sequences that are selected from multiple candidate action sequences, meet the constraints, and are actually delivered to the vehicle for execution. It consists of multiple action primitives arranged in chronological order and is the final execution plan used to complete the vehicle's attitude reversal and achieve automatic turning.

[0099] Among them, performing an automatic U-turn refers to controlling the relevant components of the vehicle chassis to complete a series of actual movements such as steering, driving, and gear shifting in sequence according to the motion requirements of each action contained in the target action sequence, so that the vehicle can complete the attitude adjustment and realize the complete process of turning around and driving out.

[0100] Specifically, based on the intersection analysis results, candidate action sequences with unavoidable dynamic obstacle spatiotemporal conflicts are first eliminated. Among the remaining candidate action sequences without unavoidable dynamic conflicts, a further comparison is made based on the risk level reflected in the intersection analysis, and the sequence with the lowest dynamic conflict risk is selected as the target action sequence. After determining the target action sequence, the entire sequence is decomposed into multiple action units to be executed sequentially, and the driving direction, steering state, movement amplitude, and constraints to be followed during the movement process for each action unit are clearly defined.

[0101] Further, in the actual execution phase, each action unit within the target action sequence is executed sequentially. Before initiating each action unit, it is confirmed that the corresponding pre-execution conditions are met. Once the conditions are met, the corresponding control command is issued to the vehicle chassis actuators, driving the vehicle to complete the motion behavior corresponding to that action unit. After completing the current action unit, the process switches to the next action unit, and all action units within the sequence are executed in a loop.

[0102] Optionally, the vehicle's actual position and pose are continuously acquired during the action execution, and motion tracking is performed to enable the vehicle to gradually complete the attitude reversal according to the motion logic of the target action sequence. When the entire target action sequence is completed and the vehicle reaches the target pose corresponding to the U-turn, the automatic U-turn process is completed. If an abnormal situation occurs during execution, the execution of the current sequence is terminated, and the upper-level logic handles the subsequent processing.

[0103] The automatic U-turn method for vehicles provided in this application utilizes a multi-attribute raster map, which differs from traditional binary drivable areas. By incorporating environmental input, it reduces planning errors caused by missed or false detections. The action primitive library provides standardized motion primitives, enabling the automatic generation of various U-turn forms based on the environment, improving adaptability to narrow and dead-end road scenarios. The calculation of a reversible spatiotemporal occupancy envelope represents the spatial and temporal occupancy of vehicles during the execution of candidate action sequences. This eliminates reliance on a single static vehicle position for feasibility assessment, reflecting the vehicle's spatial occupancy throughout the entire action execution process and improving the accuracy of U-turn feasibility judgments. Intersecting the spatiotemporal occupancy of candidate vehicle actions with the predicted future trajectories of dynamic obstacles identifies whether vehicles and dynamic obstacles will spatially overlap at the same moment during the execution of candidate action sequences, providing a dynamic risk dimension assessment basis for subsequent selection of target action sequences. Using the spatiotemporal intersection analysis results of dynamic obstacles as a crucial basis, the target action sequence is determined from multiple candidate action sequences. This ensures that the selected target sequence avoids future conflict risks from dynamic obstacles while reducing the likelihood of conservative exits.

[0104] As an optional implementation, based on the above embodiments, a candidate action sequence is generated, including:

[0105] A pre-trained action graph model is used to search for at least one candidate action sequence by starting with the current vehicle pose as the graph node and using action primitives as edges. The graph nodes include vehicle gear, blind spot level, and recoverable markers. The candidate action sequence includes at least one of the following: single U-turn sequence, three-point U-turn sequence, and multi-point U-turn sequence.

[0106] The action graph model refers to the state graph model used for U-turn path search, which consists of a large number of graph nodes and action edges connecting the nodes. Graph nodes are used to record various state information of the vehicle, and the edges connecting the nodes correspond to various action primitives. By traversing and searching the graph, the complete motion process from the initial state to the target U-turn posture is deduced.

[0107] In this context, a graph node refers to a state node in the motion graph model, representing the complete state of the vehicle at a specific moment. Besides the vehicle's current pose, a graph node also carries information such as the vehicle's gear position, blind spot level, recoverable markers, and may also include gear position, speed limit, current phase number, remaining target attitude error, and safe distance from the nearest obstacle.

[0108] Among them, vehicle gear refers to whether the vehicle is in forward or reverse gear in the current state; blind spot level is used to characterize the level of occlusion and blind spot risk in the surrounding environment of the vehicle position corresponding to the node; recoverable marker is used to characterize whether there is still a feasible movement path from the vehicle state corresponding to the node, and whether the U-turn target can be completed; marked as unrecoverable means that it is difficult to adjust to the target U-turn posture in this state.

[0109] Specifically, a starting graph node is first constructed, using the vehicle's real-time pose as the node's basic information. The vehicle's current gear and the surrounding blind spot level obtained from multi-attribute raster map statistics are simultaneously filled in. An initial recoverable marker is set, and this starting node is added to the search set. All preset action primitives are read, and each action primitive is used as an expandable edge of the action graph. The graph node to be processed is retrieved from the search set, and all available action edges are traversed. The action primitive corresponding to that edge is simulated and executed, and the new vehicle pose after the action is derived using vehicle kinematics. Then, the gear and blind spot level of the new node are updated using the environment raster, and it is evaluated whether there is still a feasible path to the U-turn endpoint in the new state. The recoverable marker of the new node is set, generating a new graph node to be verified.

[0110] Optionally, the newly generated graph nodes are validated. If a node falls into a traversable region, it is discarded and no further expansion is made. Validated new graph nodes are stored in the search set, along with a record of which parent node and which action edge the node migrated from.

[0111] Understandably, the process involves continuously retrieving nodes from the search set, expanding action edges, and generating and validating new nodes. When the vehicle pose of a newly generated graph node meets the target pose requirement for turning around, the process backtracks along the parent-child relationships between nodes, organizing all action edges along the path in order of execution to obtain a candidate action sequence.

[0112] It should be noted that all candidate action sequences satisfy vehicle kinematic constraints, collision constraints, gear shifting constraints, and endpoint posture constraints. The automatic U-turn method provided in this application constructs the starting node of the action graph based on the vehicle's current real state. The node integrates multi-dimensional information such as gear position, blind spot level, and recoverable markers, eliminating deadlock-type invalid nodes during the search phase and reducing unnecessary computation. Using action primitives as the migration edges of the graph for searching, there is no need to pre-set fixed U-turn logic; various candidate action sequences such as single, three-point, or multi-point U-turns can be automatically generated based on the actual environment. Through the output of multiple candidate solution sets, sufficient alternative solutions are provided for subsequent envelope calculation and dynamic obstacle conflict analysis, improving the probability of finding solutions in complex open U-turn scenarios such as narrow roads and blind spots, and enhancing scenario adaptability.

[0113] As an optional implementation, based on the above embodiments, it further includes:

[0114] Generate a driveable area for U-turns based on the current multi-attribute raster map;

[0115] At least one candidate target pose is generated within the drivable area for U-turns according to a preset adaptation rule, so that when searching, the candidate target pose is used as the termination graph node and the drivable area for U-turns is used as the boundary to generate a candidate action sequence.

[0116] The U-turn drivable area refers to a connected, locally open space calculated from a multi-attribute raster map. This area is where vehicles can reverse their posture by moving forward and backward. This area excludes restricted zones such as walls, curbs, and static obstacles. As a spatial boundary constraint for action graph search, all graph nodes generated by the vehicle search cannot exceed the boundaries of this area.

[0117] The candidate target pose refers to the desired state after the vehicle completes an automatic U-turn, including the target position and target heading angle. All candidate target poses are confined to the drivable area for the U-turn and serve as the termination node of the motion graph search, representing the desired endpoint state, such as being opposite to the direction of oncoming traffic, aligned with the exit direction, aligned with the direction of the next navigation path, or aligned with the direction specified by the user on the interface.

[0118] Among them, the preset adaptation rules refer to multiple pre-set pose generation logics, which are used to combine road direction, navigation path and user intent to generate reasonable candidate target poses in the driving area where U-turns are possible, and filter out invalid poses that are close to obstacles or close to the boundary.

[0119] Specifically, the existing multi-attribute raster map is read, and information on obstacles and insurmountable boundaries within the map is extracted. After safety expansion, walls, insurmountable curbs, and obstacles are designated as restricted areas. Fragmented small areas that are too small to allow for vehicle reversal are removed, and the remaining connected open space becomes the driving area for U-turns.

[0120] Furthermore, within the drivable area for U-turns, candidate target poses are generated according to preset adaptation rules. One or more candidate target poses are generated based on logic such as road exit direction, navigation path direction, and opposite direction to oncoming traffic, and unqualified poses that fall in restricted areas or are too close to obstacles are eliminated.

[0121] Understandably, the obtained set of candidate target poses is used as the terminal graph node for the action graph search, while the drivable area for turning around is used as the spatial boundary constraint for the search. During the subsequent action graph search to expand the graph nodes, if a newly derived graph node exceeds the drivable area for turning around, that node is discarded. When the expanded graph node satisfies the pose matching condition with any candidate target pose, the corresponding candidate action sequence is backtracked and generated.

[0122] The automatic U-turn method for vehicles provided in this application generates a drivable U-turn area based on a multi-attribute grid map, sets spatial boundaries for action graph search, filters fragmented and invalid space, reduces invalid search calculations, and avoids generating action schemes that lead into restricted areas. Within this area, multiple candidate target poses are generated as search termination graph nodes according to adaptation rules, eliminating the limitation of a single endpoint. During the search, the drivable area is used as a boundary constraint, and the candidate target poses are used as termination conditions, adapting to scenarios such as dead-end roads and narrow roads without fixed U-turn endpoints, improving the rationality and scenario adaptability of the generated candidate action sequences.

[0123] As an optional implementation, based on the above embodiments, the calculation of the reversible spatiotemporal occupancy envelope includes:

[0124] Predict the vehicle pose at each time step for each action in the candidate action sequence;

[0125] The sweep envelope of the vehicle pose at each moment is calculated using a preset dilation rule corresponding to the action type. The sweep envelope includes the safety dilation distance.

[0126] Among them, action type refers to the driving attribute of a single action in the candidate action sequence, mainly distinguishing between forward actions and reverse actions.

[0127] Among them, the preset expansion rule refers to a set of pre-defined constraint logic for generating the sweep envelope, which configures differentiated expansion processing logic for different action types and specifies the calculation method and effective area of ​​the safe expansion distance under different actions.

[0128] The safety expansion distance refers to the additional buffer distance that is added outward based on the original outer contour of the vehicle. It is used to reserve a safety margin and avoid the risk of scraping and colliding with surrounding obstacles.

[0129] Specifically, for each action in the sequence, the complete execution process of that action is determined, and the complete execution time of the action is divided into several discrete time points. Based on the motion deduction method of vehicle kinematics, combined with the steering and driving direction information of the action, the vehicle pose corresponding to each discrete time point under the action is predicted one by one.

[0130] Furthermore, the action type corresponding to the current action is identified, and the preset expansion rule matching that action type is retrieved. Based on the vehicle pose at the current moment, the original outer contour of the vehicle body at that moment is reconstructed. According to the corresponding preset expansion rule, the safe expansion distance to be used at this moment is determined, and the safe expansion distance is applied to the original outer contour of the vehicle body to perform spatial expansion processing outward, generating the sweep envelope corresponding to this moment.

[0131] For example, the formula for calculating the safety expansion distance dsafe can be given by equation (1):

[0132] dsafe = d0 + k1×(1-c) + k2×o + k3×vrel + k4×Ireverse (1)

[0133] Where d0 is the basic safety distance, c is the confidence level of the grid where the current pose is located, o is the grid blind zone level, vrel is the dynamic obstacle speed, Ireverse is whether it is a reversing action, and k1, k2, k3, and k4 are the weights corresponding to the confidence level of the grid where the current pose is located, the grid blind zone level, the dynamic obstacle speed, and whether it is a reversing action, respectively.

[0134] Optionally, forward driving focuses on setting a safe expansion distance at the front and sides of the vehicle, while reverse driving focuses on setting a safe expansion distance at the rear and sides of the vehicle.

[0135] Understandably, the above process is repeated for all discrete moments of an action to obtain the next set of swept envelopes with time information for that action. The swept envelopes for all moments corresponding to all actions in the candidate sequence are then aggregated and combined to obtain the complete reversible spatiotemporal occupancy envelope of the candidate action sequence. The entire above processing procedure is then repeated for the remaining candidate action sequences.

[0136] The automatic U-turn method for vehicles provided in this application discretizes the action execution process over time, predicts the vehicle's pose at each moment, and reconstructs the position and orientation changes of the vehicle throughout its movement, avoiding information gaps caused by relying solely on start and end point assessments. It calls corresponding preset expansion rules based on the action type to adapt to the alternating motion characteristics of U-turns and reversing. A sweep envelope is obtained by expanding the original vehicle body contour using a safe expansion distance, introducing a safety buffer margin and improving the completeness and safety of space occupancy assessment.

[0137] As an optional implementation, based on the above embodiments, the preset dilation rule includes the region to be dilated corresponding to each action type, and the calculation of the sweep envelope includes:

[0138] Calculate the safe expansion distance based on the confidence level of the grid where the current pose is located, the grid blind zone level, the speed of dynamic obstacles, and the action type;

[0139] The region to be expanded is expanded according to the action type and the safe expansion distance to generate a sweep envelope.

[0140] The area to be expanded refers to the local contour area of ​​the vehicle body defined according to the type of action. The area to be expanded differs for different types of actions. For example, reversing focuses on the rear and rear side contour areas, while moving forward focuses on the front and front side contour areas. These are the targets for applying the safety expansion distance.

[0141] Among them, grid confidence refers to the degree of credibility of the environmental perception results representing the corresponding grid location in a multi-attribute grid map. The lower the confidence, the higher the uncertainty of the sensor observation results at that location.

[0142] Among them, the grid blind spot level represents the risk of blind spots being formed by the grid around the current vehicle's position and pose being blocked. The higher the level, the more limited the perception field of view in that area.

[0143] Among them, dynamic obstacle speed refers to the speed at which surrounding dynamic obstacles move, and is used to reflect the degree of collision risk posed by dynamic targets.

[0144] Specifically, the system acquires the vehicle's current pose, locates the corresponding grid cell in the multi-attribute grid map, reads the grid confidence and blind spot level at that location, and simultaneously acquires the speed information of surrounding dynamic obstacles. Combined with the action type of the currently executed action, the system calculates the corresponding safe expansion distance for that moment. It then consults preset expansion rules and determines the expansion area to be processed based on the current action type.

[0145] Furthermore, for the original outer contour of the vehicle body, only the area to be expanded as specified by the rules is subjected to spatial expansion processing outward using the calculated safe expansion distance; the remaining areas of the vehicle body are not subjected to this expansion processing. After the area expansion is completed, a closed polygonal space is formed, which is the sweep envelope corresponding to this moment.

[0146] Understandably, the above process is repeated across all discrete moments of an action to obtain the sweep envelope for all moments under that action. After all actions within a candidate sequence have been processed, the entire temporal sweep envelope is aggregated to form the reversible spatiotemporal occupancy envelope. The same process is performed on the remaining candidate action sequences.

[0147] The automatic U-turn method for vehicles provided in this application dynamically calculates the safe expansion distance by comprehensively considering grid confidence, blind spot level, dynamic obstacle speed, and action type. It abandons the fixed expansion distance mode and can adaptively adjust the buffer margin according to environmental perception uncertainty and external risks. Based on the action type, the corresponding expansion area is determined, and expansion is performed only on a specified local area of ​​the vehicle body to achieve differentiated envelope generation, matching the risk distribution of different directions of forward and reverse movement. This provides sufficient safety protection for high-risk areas while preventing excessive expansion of the overall envelope and mis-screening of feasible paths, balancing safety and traffic capacity, and providing a more realistic assessment basis for subsequent spatiotemporal conflict analysis.

[0148] As an optional implementation, based on the above embodiments, intersection analysis is performed, including:

[0149] Calculate the spatiotemporal overlap point between the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence and the predicted trajectory of the dynamic obstacle at the same moment;

[0150] The temporal conflict cost of generating each candidate action sequence based on spatiotemporal overlap points.

[0151] Spatiotemporal overlap points refer to the locations where the spatial regions of the vehicle's sweep envelope and the predicted areas occupied by dynamic obstacles intersect at the same moment. Both temporal alignment and spatial intersection must be met simultaneously; spatial overlap without temporal mismatch is not considered a spatiotemporal overlap point.

[0152] Among them, the temporal conflict cost refers to the evaluation index that quantifies the degree of potential conflict risk between a single candidate action sequence and a dynamic obstacle. It is generated by statistical analysis of the identified spatiotemporal overlap points and is used to measure the risk that the dynamic obstacle brings to the candidate sequence.

[0153] Specifically, for each sweep envelope of the current candidate sequence, the corresponding time and the spatial region occupied by the vehicle are extracted. In the dynamic obstacle prediction trajectory, the predicted spatial position of the obstacle at exactly the same time is retrieved. A geometric comparison is performed between the vehicle sweep envelope and the predicted spatial region of the dynamic obstacle at the same time. If their spatial ranges intersect, the corresponding intersection position is recorded as a spatiotemporal overlap point. If their spatial ranges do not intersect at the same time, there is no spatiotemporal overlap point.

[0154] Furthermore, the swept envelope of the candidate action sequence is traversed across all time points, and the predicted trajectories of all dynamic obstacles are also traversed to collect all spatiotemporal overlap points. Combining the number of spatiotemporal overlap points and the corresponding time of their occurrence, the temporal conflict cost of the candidate action sequence is calculated.

[0155] The automatic vehicle U-turn method provided in this application adopts dual constraints of time and space to identify spatiotemporal overlap points. Unlike simple static spatial comparison, it can filter out false conflicts where spaces intersect but time is staggered, avoiding misjudging the path as infeasible. It transforms the potential conflict between the vehicle and dynamic obstacles into a quantifiable temporal conflict cost, and converts dynamic risk into a numerical evaluation index, which facilitates horizontal comparison of risk levels among multiple candidate action sequences.

[0156] As an optional implementation, based on the above embodiments, the timing conflict cost is generated, including:

[0157] If the spatiotemporal overlap point leaves the envelope region before the vehicle arrives in the predicted trajectory, the cost of temporal conflict is reduced.

[0158] If there are no spatiotemporal overlap points at present, but the predicted trajectory overlaps with the sweep envelope at a future time, the cost of temporal conflict will increase.

[0159] Optionally, for cases where spatiotemporal overlap points have been detected, the time when the dynamic obstacle leaves the overlapping region is compared with the time when the vehicle's sweep envelope occupies that region. It is determined whether the dynamic obstacle can leave the space before the vehicle's corresponding sweep envelope reaches the overlapping region. If the condition for the obstacle to leave early is met, it indicates that the overlap is only an instantaneous spatial intersection and no actual collision will occur; therefore, the temporal conflict cost of this candidate action sequence is reduced.

[0160] Optionally, if no spatiotemporal overlap point is detected at the current moment, the complete predicted trajectory of the dynamic obstacle is checked further to see if the predicted position of the dynamic obstacle will spatially intersect with the sweep envelope at the corresponding moment in the future execution time of the candidate action sequence. If an intersection is predicted to occur in the future, even if there is no overlap at present, a potential intrusion risk is still determined, and the temporal conflict cost of the candidate action sequence is increased.

[0161] The automatic vehicle turning method provided in this application adjusts the timing conflict cost differently based on the different movement trends of dynamic obstacles. For scenarios where the obstacle can leave the overlapping area before the vehicle arrives, the timing conflict cost is reduced, identifying situations where avoidance is possible in time, and reducing the false exclusion of feasible solutions. For situations where there is currently no overlap but future intrusion is possible, the timing conflict cost is increased to identify potential collision risks that occur later in advance.

[0162] As an optional implementation, based on the above embodiments, determining the target action sequence includes:

[0163] The total cost corresponding to each candidate action sequence is calculated based on static collision cost, temporal conflict cost, blind zone cost, boundary cost, number of gear shifts, curvature change, and target pose error.

[0164] The target action sequence is determined based on the total cost corresponding to each candidate action sequence.

[0165] Among them, static collision cost refers to the risk quantification value brought about by the intersection of the reversible spatiotemporal occupancy envelope and static obstacles in the environment during the execution of the candidate action sequence. The higher the degree of intersection, the greater the static collision cost.

[0166] Among them, blind zone cost refers to the risk quantification value generated when the sweep envelope of the candidate action sequence passes through a high blind zone risk area. The higher the blind zone level the path passes through, the higher the blind zone cost.

[0167] Among them, boundary cost refers to the proximity of the candidate action sequence to insurmountable boundaries such as roads, walls, and curbs. The closer the path is to an insurmountable boundary, the higher the boundary cost.

[0168] Among them, the number of gear shifts refers to the total number of times the vehicle switches between forward and reverse gears during the execution of a candidate action sequence. An increase in the number of gear shifts will affect the smoothness of driving.

[0169] Among them, the curvature change refers to the change in steering curvature between adjacent actions in the sequence. The greater the curvature change, the more violent the counter-steering, which directly affects the ride comfort during the U-turn.

[0170] Among them, the target pose error refers to the position and heading angle deviation between the pose of the candidate action sequence termination node and the candidate target pose. The larger the deviation, the higher the error value.

[0171] Specifically, for each candidate action sequence, various sub-indicators corresponding to that sequence are collected. The reversible spatiotemporal occupancy envelope and static collision cost calculated from static obstacles are obtained; the temporal conflict cost obtained from dynamic obstacle analysis is obtained; the blind zone cost obtained by the path passing through blind zone areas is calculated; and the boundary cost is obtained by combining the proximity of the path to the insurmountable boundary. Simultaneously, the total number of gear shifts during the complete execution of the sequence is counted; the cumulative change in steering curvature between each action is calculated to obtain the curvature change; and the target pose error of the sequence endpoint relative to the candidate target pose is obtained.

[0172] Furthermore, the total cost of all the above-mentioned component costs and indicators are comprehensively calculated to obtain the total cost corresponding to the candidate action sequence. After the total cost of all candidate action sequences has been calculated, the total costs of all candidate sequences are compared, and the candidate sequence with the lowest total cost that meets the preset constraints is selected and determined as the target action sequence.

[0173] It is understandable that each candidate action sequence satisfies collision constraints, minimum safe distance, minimum turning radius of the vehicle, maximum steering angular velocity, shift safe distance, target attitude error, and execution time limit.

[0174] For example, the total cost J can be calculated using equation (2):

[0175] J(M)= ws·Cs+ wd·Cd+ wo·Co+ wb·Cb+ wn·Nshift + wk·Δκ|+ we·Egoal (2)

[0176] Where M is the candidate turnaround sequence identifier, Cs is the static collision cost, Cd is the temporal conflict cost, Co is the blind zone cost, Cb is the boundary cost, Nshift is the number of gear shifts, Δκ is the curvature change, Egoal is the target pose error, and ws, wd, wo, wb, wn, wk, and we are the weights corresponding to the static collision cost, temporal conflict cost, blind zone cost, boundary cost, number of gear shifts, curvature change, and target pose error, respectively.

[0177] The automatic U-turn method for vehicles provided in this application comprehensively considers static collisions, temporal conflicts, blind spots, boundary risks, as well as gear shifts, curvature changes, and target pose errors to calculate the total cost of each candidate sequence, constructing a multi-dimensional unified evaluation system. It takes into account various safety risks arising from static obstacles, dynamic targets, perception blind spots, and road boundaries, while also incorporating gear shifts and curvature changes to consider driving comfort, and relying on target pose errors to ensure the accuracy of the final U-turn. By comparing and filtering the total cost to obtain the target action sequence, it avoids relying on a single indicator to select a solution, and can select a U-turn solution with low safety risk, smooth action, and high final pose accuracy.

[0178] As an optional implementation, based on the above embodiments, it further includes:

[0179] If the total cost corresponding to each candidate action sequence does not meet the threshold of the target action sequence, and the recoverable flag in any candidate action sequence is true, then at least one probe action is generated.

[0180] The reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence is updated based on the perception results corresponding to the detection action.

[0181] Among them, the detection action refers to a special type of short-distance, low-speed movement action that is not intended to directly complete a U-turn. Its core function is to change the vehicle's current position and adjust the sensor's observation angle to perceive and detect high-risk blind spots in the surrounding area. After the action is completed, the vehicle still has the ability to return to its original state and continue making a U-turn.

[0182] Among them, the perception result corresponding to the detection action refers to the surrounding environment perception information collected by the sensor in the new observation pose during the vehicle's detection action.

[0183] Specifically, after calculating the total cost of all candidate action sequences, the total cost of each candidate sequence is compared with a preset total cost threshold. If the total cost of all candidate action sequences exceeds the threshold, it means that there is no target action sequence that can be directly executed under the current conditions. Further checks are made on the recoverable markers of the graph nodes corresponding to each candidate action sequence. If any candidate sequence has a node recoverable marker that is true, it means that the vehicle's current state is not in a deadlock and meets the basic conditions for conducting detection; therefore, at least one detection action is generated.

[0184] It should be noted that when generating the detection action, the action itself is ensured to have a short movement distance and low operating speed, and after the detection action is executed, the vehicle still has the ability to return to the current position and continue to carry out the U-turn planning.

[0185] Furthermore, the vehicle is controlled to perform the detection action. During the detection process, the sensors follow the vehicle's new observation position to collect information about the surrounding environment and obtain new perception results. The newly acquired perception results are used to refresh the multi-attribute raster map. Based on the updated environmental information, the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence is revised, reflecting the newly acquired environmental conditions in the spatiotemporal envelope. After the update is complete, the process returns to the overall cost evaluation stage to re-evaluate each candidate action sequence and determine whether a qualified target action sequence can be selected.

[0186] Optionally, if all candidate sequences are marked as false, no further detection action will be generated, and the system will enter the abnormal function handling process, indicating that the function is unavailable and maintaining vehicle braking.

[0187] Optionally, based on the severity of the blind spot, such as obstruction by wall corners or multiple vehicles, and the current recoverability of the vehicle, detection actions can be generated in stages, such as low-speed forward probing, steering fine-tuning, and multi-stage micro-movement, to gradually eliminate the risk of the blind spot.

[0188] For example, in low-level blind spots, sensor coverage information is quickly obtained by moving forward at low speed (0.2-0.5m) or making slight steering adjustments. In medium-level blind spots, multi-stage micro-movements are generated, such as moving forward 0.3m, reversing 0.1m, and turning 5°, to gradually cover the blind spot. In high-level blind spots, detour paths are planned based on environmental topology information, such as road boundaries and obstacle distribution, to obtain the sensor's field of view.

[0189] The automatic U-turn method for vehicles provided in this application does not directly exit when the total cost of all candidate action sequences does not meet the threshold. Instead, it determines whether adjustment conditions are met based on recoverable markers. By changing the sensor's observation perspective through short-range, low-speed detection actions, it supplements blind spot environmental information. Then, based on the newly obtained perception results, it updates the reversible spatiotemporal occupancy envelope to reflect the latest environmental conditions, and re-evaluates the proposed solution. This avoids directly exiting the function due to insufficient perception information, increasing the probability of finding feasible U-turn solutions in occluded and blind spot scenarios while ensuring safety, thus balancing the function's safety and usability.

[0190] As an optional implementation, based on the above embodiments, it further includes:

[0191] If the blind zone cost of the target action sequence is greater than a preset cost threshold, then at least one detection action is generated.

[0192] When the perception result corresponding to the detection action does not meet the execution conditions, the target action sequence is adjusted.

[0193] The blind zone cost threshold is a pre-set risk assessment threshold used to measure whether the risk posed by the perceived blind zone to the target action sequence is within a range that can be directly executed. If the blind zone cost exceeds this threshold, it means that the path passes through many highly uncertain blind zones, and it is not advisable to directly execute a U-turn.

[0194] Among them, the execution conditions refer to the risk constraints that the perception results obtained after detection need to meet, such as whether the blind spots around the original target path are eliminated, whether new obstacles are detected, and whether the perception confidence is improved.

[0195] Specifically, after determining the target action sequence, the blind zone cost corresponding to the sequence is read and compared with a preset blind zone cost threshold. If the blind zone cost is greater than the preset cost threshold, it means that although the overall cost of the target action sequence meets the standard, the path still has a high blind zone risk and is not suitable for direct execution. In this case, at least one probe action is generated.

[0196] Furthermore, the vehicle is controlled to perform detection actions. During the detection process, new perception results are acquired by sensors, and these results are verified to see if they meet preset execution conditions. If the perception results meet the execution conditions, it means that the blind spot risk corresponding to the original path has been eliminated, and the original target action sequence can continue to be executed.

[0197] Optionally, if the perception results returned by the detection action do not meet the execution conditions, it indicates that a high blind spot risk still exists after detection, or a new obstacle has been identified, and the original target action sequence is no longer safe. In this case, the target action sequence can be adjusted. This can involve modifying the driving distance, steering curvature, and gear shifting positions of each action within the original sequence, or the current target action sequence can be abandoned, and a suitable sequence can be selected from other candidate action sequences to complete the adjustment process. After the adjustment is completed, a risk assessment is conducted again to determine whether the adjusted sequence can be put into execution.

[0198] Optionally, the detection action should meet the following requirements: the vehicle can safely and feasiblely move short distances from its current position to the detection position; the detection action should allow for immediate stopping at low speeds; the detection position should improve visibility or ultrasonic coverage in critical blind spots; and the vehicle should be able to turn around or return to its original position after detection. By calculating the visibility gain of each detection action and comprehensively evaluating it in conjunction with the risk of the detection action, the detection action with the highest benefit and lowest risk is selected.

[0199] The automatic U-turn method provided in this application verifies the blind spot cost separately after selecting the target action sequence, identifying situations where the overall cost is acceptable but the blind spot risk is high. When the blind spot cost exceeds the limit, a detection action is generated to verify the actual environment inside the blind spot using a new sensor observation perspective. If the perception results obtained from the detection do not meet the execution conditions, the target action sequence is adjusted, such as modifying motion parameters or changing the alternative sequence. Instead of simply exiting the function, this method handles the perception blind spot risk through detection verification and sequence adjustment, improving the usability of automatic U-turns in occluded scenarios while ensuring safety.

[0200] As an optional implementation, based on the above embodiments, automatic U-turn is performed, including:

[0201] The target action sequence is divided into multiple phases to be executed, and each phase to be executed includes constraints.

[0202] During the execution of each phase to be executed, the real-time pose of the vehicle is acquired.

[0203] The phase to be executed refers to the segmented execution unit obtained by decomposing the target action sequence. A target action sequence is composed of several phases to be executed in sequence, and each phase corresponds to a continuous segment of vehicle movement. For example, a segment of forward turning movement or a segment of reverse turning movement, the complete U-turn action is segmented and controlled.

[0204] Among them, the constraints refer to the execution rules corresponding to each phase to be executed, such as start conditions, tracking targets, speed limits, termination conditions, and fault exit conditions.

[0205] Among them, the real-time vehicle pose refers to the state information calculated in real time during the operation of the vehicle, including the vehicle's current planar position and body heading angle, reflecting the vehicle's true position and body orientation.

[0206] Specifically, the predetermined target action sequence is obtained, and based on the boundaries of each action within the sequence, the entire sequence is broken down into multiple consecutive execution phases. Corresponding constraints are configured for each execution phase, including the phase-specific start condition, tracking target, speed limit, termination condition, and fault exit condition.

[0207] Furthermore, each phase to be executed is processed sequentially according to its order. Before starting the current phase, the start conditions for that phase are verified. Only when all start conditions are met is the phase execution process allowed. Throughout the phase operation, the vehicle's real-time pose is continuously acquired and compared with the tracking target set for that phase to support the chassis controller in completing motion tracking control. The execution process continuously monitors whether a fault exit condition is triggered. Once triggered, the turnaround execution is terminated.

[0208] Understandably, the termination conditions for continuously comparing the vehicle's real-time pose and phase are met. When the termination conditions are met, the current phase to be executed ends, and the next phase to be executed begins. The entire process of condition verification, obtaining real-time pose, and status monitoring is repeated until all phases are executed.

[0209] The automatic U-turn method for vehicles provided in this application divides the target action sequence into multiple phases to be executed, enabling segmented control of complex U-turn movements. Each phase is equipped with complete constraints, covering start conditions, tracking targets, speed limits, termination conditions, and fault exit conditions, achieving full-process safety constraints for a single movement segment. During phase execution, the vehicle's real-time pose is continuously acquired for motion tracking, phase completion determination, and fault monitoring. The phase-segmented mode allows for state verification and anomaly handling at segment boundaries, improving the controllability and safety of the automatic U-turn execution phase.

[0210] As an optional implementation, based on the above embodiments, it further includes:

[0211] Based on the vehicle's real-time pose, the current multi-attribute raster map and reversible spatiotemporal occupancy envelope are periodically updated;

[0212] When a preset abnormal condition is triggered and the vehicle comes to a stop, at least one candidate action sequence is regenerated based on the updated multi-attribute raster map.

[0213] The preset abnormal conditions refer to predefined abnormal trigger criteria, such as excessive tracking deviation, sudden changes in the surrounding environment, abnormal sensor status, or close-range intrusion of dynamic obstacles. Once the condition is met, the vehicle is controlled to brake to a stop, suspending the continued execution of the current phase.

[0214] Specifically, during the execution of the phase to be executed, the vehicle's real-time pose is periodically acquired. Based on the vehicle's latest real-time pose and combined with the latest environmental perception information collected by the sensors, the current multi-attribute grid map is periodically updated, correcting various attributes such as obstacles, confidence levels, and blind spot levels within the grid. Based on the updated results of the multi-attribute grid map, the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence is updated synchronously, reflecting the latest changes in the environment in the spatiotemporal envelope and keeping the evaluation data synchronized with the real environment.

[0215] Furthermore, during execution, the system continuously monitors for the triggering of preset abnormal conditions. If a preset abnormal condition is triggered, the vehicle is immediately brought to a stop, ceasing execution of the current pending phase. Once the vehicle has come to a stable stop, the old candidate action sequence is no longer used. Based on the newly updated multi-attribute raster map, a new graph search is performed to generate at least one new candidate action sequence, re-entering the subsequent cost evaluation and target action sequence selection process.

[0216] Optionally, if a new safe sequence is obtained through local replanning, execution continues from the current pose; if it cannot continue, the driver is prompted to take over.

[0217] The automatic U-turn method provided in this application periodically updates the multi-attribute grid map and reversible spatiotemporal occupancy envelope based on the vehicle's real-time pose during phase execution. This ensures that the risk assessment data is synchronized with the latest environment, preventing misjudgments caused by using outdated environmental information. When a preset abnormal condition is triggered, a new sequence of candidate actions is regenerated based on the updated map, and a new round of planning is initiated, rather than directly terminating the function. This forms a closed loop of execution, perception update, abnormal braking, and replanning, addressing sudden environmental changes during the U-turn process and improving the robustness of automatic U-turns in dynamic and complex scenarios.

[0218] As an optional implementation, based on the above embodiments, it further includes:

[0219] In response to receiving an automatic U-turn command, it acquires multi-source sensor data, including visual perception data, ultrasonic perception data, and current vehicle status data;

[0220] The current multi-attribute raster map is generated based on multi-source sensor data. The current multi-attribute raster map includes raster coordinates, obstacle occupancy probability, obstacle height, obstacle speed, confidence level, and blind spot level.

[0221] The automatic U-turn command refers to the control command that triggers the automatic U-turn function. It can come from vehicle-machine interaction, voice commands, navigation logic, or driver operation signals, and serve as the trigger signal to start the entire U-turn process.

[0222] Multi-source sensor data refers to the raw information from multiple sensors collected at the moment the function is activated, including visual perception data, ultrasonic perception data, and current vehicle status data. Visual perception data comes from surround-view cameras and front-view cameras, outputting semantic segmentation, object detection, and road boundary information. Ultrasonic perception data comes from ultrasonic probes, outputting information such as near-range echo and echo intensity. Vehicle status data includes chassis status information such as vehicle speed, wheel speed, gear position, and steering angle.

[0223] The confidence level can be calculated based on obstacle source, distance, observation angle, historical continuity, obstacle height, ultrasonic echo intensity, visual semantic category, and multi-sensor consistency.

[0224] Specifically, upon receiving an automatic U-turn command, the system immediately collects raw information from multiple sensors, acquiring visual perception data, ultrasonic perception data, and simultaneously reading the current vehicle status data output by the vehicle chassis. Perception fusion processing is performed on the multi-source sensor data to align the temporal and spatial references between different sensors, eliminating coordinate and temporal deviations between them. Combining the pose information from the vehicle status data, all perception and detection results are uniformly projected onto a local coordinate system centered on the vehicle.

[0225] Furthermore, the surrounding environment is divided into grids according to fixed dimensions to generate a grid array. For each grid, visual and ultrasonic observations are fused to fill in the grid coordinates. The probability of an obstacle being present at that location is calculated to obtain the obstacle occupancy probability. The obstacle height is obtained from the visual output, and the obstacle velocity is calculated based on continuous multi-frame observations. The confidence level of the grid is calculated by combining sensor observation quality and multi-sensor result consistency, and the blind zone level corresponding to that grid is calculated by combining obstacle occlusion relationships. The last observation time can also be marked.

[0226] Optionally, when there is a conflict between visual and ultrasonic results for the same grid, confidence fusion is performed based on the sensor's visible area, obstacle material, distance, historical continuity, and vehicle movement direction.

[0227] Optionally, when constructing a local occupancy grid, the weights of each sensor on the grid occupancy status can be dynamically adjusted based on sensor type, obstacle attributes such as height and material, environmental conditions such as lighting and occlusion, and consistency of historical observations. For example, in low-light scenes, the weight of cameras can be reduced while the weight of ultrasonic sensors can be increased. In corner areas, the weight of millimeter-wave radar can be increased to cover camera blind spots.

[0228] The automatic vehicle U-turn method provided in this application integrates multi-dimensional information such as obstacle occupancy probability, height, speed, confidence level, and blind spot level into the grid, in addition to location information, taking into account both physical environment information and perception uncertainty. Multi-sensor fusion leverages the strengths of each sensor to compensate for their weaknesses, avoiding the shortcomings of single sensors. The entire planning process uses the same grid map, ensuring consistent data standards. This provides fine-grained environmental input for subsequent candidate sequence search, spatiotemporal envelope calculation, and multi-dimensional cost evaluation, improving the processing capability for occlusion-related U-turn scenarios.

[0229] As an optional implementation, based on the above embodiments, the preset action primitive library includes basic action primitives and detection action primitives;

[0230] Each action primitive includes direction, curvature range, steering angular velocity threshold, speed threshold, driving distance range, shift requirements, abort conditions, and recoverable conditions.

[0231] Among them, basic motion primitives refer to the basic motion units used to directly complete the turning posture, mainly including forward steering, reverse steering, forward straight driving, and reverse straight driving.

[0232] Among them, the detection action primitive refers to a special motion unit used for perception and detection. It is not directly aimed at completing a turnaround or reversal. The motion is aimed at changing the vehicle's observation perspective and exploring the blind spot environment. The motion amplitude is small and the speed is low. It is used to carry out environmental detection when there is insufficient perception information, such as detecting forward movement or detection reversal.

[0233] Specifically, during the development phase, a pre-defined action primitive library is constructed, categorizing action primitives into two types: basic action primitives and detection action primitives. For each action primitive, eight attributes are fully configured: direction, curvature range, steering angular velocity threshold, speed threshold, travel distance range, shift requirements, abort conditions, and recoverability conditions. Basic action primitives focus on the motion of the U-turn itself, configuring various constraints according to common forward, reverse, straight, and turning scenarios during the U-turn process. Detection action primitives are configured with a smaller travel distance range and a lower speed threshold, while strictly setting recoverability conditions to ensure that the vehicle can return to its original state and continue the U-turn planning after the detection is executed.

[0234] Understandably, when executing any action primitive, the constraints configured within the primitive must be followed throughout the process; the curvature, rate of change of direction, speed, and distance traveled must not exceed the set ranges. The execution process continuously monitors abort conditions; if triggered, the current action can be terminated midway. Simultaneously, recoverability conditions are checked to determine whether there is a possibility of replanning after executing the action, providing a basis for determining the recoverability markers of graph nodes.

[0235] The automatic vehicle U-turn method provided in this application's embodiments configures multi-dimensional attributes for each primitive, directly calls the primitive library during the search phase to ensure that the generated action sequence is within the chassis's executable range, and considers gear shifting logic in advance. It relies on abortable conditions to support mid-motion stopping and recoverable conditions to assist in deadlock judgment. It decouples the U-turn motion from the detection motion, reduces runtime computation, improves the executability of candidate action sequences, and adapts to the entire workflow of multi-point U-turns and detection replanning.

[0236] As an optional implementation, based on the above embodiments, it further includes:

[0237] When the vehicle travels to the target area corresponding to the target pose, check the vehicle center position error, heading angle error, distance between the vehicle body and the road boundary, driving space, gear status and obstacle conditions.

[0238] If the test fails, a corrective action sequence is generated and execution continues.

[0239] The target area refers to the allowable error range defined outward from the target pose. After the vehicle completes the U-turn, it should enter this area, which represents the approximate completion range of the U-turn task, not a single point with zero error.

[0240] Among them, vehicle center position error refers to the positional deviation between the actual center coordinates of the vehicle and the center coordinates of the target pose, which is used to measure the degree of deviation of the vehicle's parking position.

[0241] Among them, the heading angle error refers to the angle difference between the actual vehicle body facing angle and the expected heading angle of the target pose, which represents whether the vehicle body turning angle meets the requirements for turning around.

[0242] Among them, the distance between the vehicle body and the road boundary refers to the shortest distance from the outer contour of the vehicle body to the curb, wall, or other road boundaries that cannot be crossed, and is used to determine whether the vehicle is too close to the boundary.

[0243] The exit space refers to the amount of space in front of the vehicle that needs to move forward from its current position to exit the U-turn area. It is used to verify whether the vehicle can leave smoothly after the U-turn is completed.

[0244] Among them, the correction action sequence refers to a small-scale motion sequence generated when the target area is reached but multiple verification conditions are not all met. It consists of short-distance forward, reverse, and turning actions, and is used to fine-tune the vehicle's position and attitude without restarting a large-scale global search.

[0245] Specifically, the vehicle executes each phase to be executed, continuously tracking the planned trajectory. When it detects that the vehicle has entered the target area corresponding to the target pose, it does not immediately determine that the U-turn is complete, but initiates a multi-dimensional verification process. This involves collecting the vehicle's center position error and heading angle error, calculating the actual distance between the vehicle body and the road boundary, measuring the available exit space ahead using a multi-attribute grid map, reading the vehicle's current gear status, and simultaneously reading the occupancy status of surrounding obstacles. Each of these indicators is then compared to its respective preset pass threshold.

[0246] Optionally, if all verification items meet the passing thresholds, the automatic U-turn task is considered complete, and the entire U-turn control process ends. If any verification fails, it indicates that although the vehicle has entered the target area, its position, angle, and surrounding environment have not yet met the final usage requirements, and a correction action sequence is generated. The correction action sequence starts with the current real-time vehicle pose as the starting state and aims to meet all verification indicators as the adjustment target, generating a combination of minor vehicle movement and steering adjustments. The correction action sequence is then broken down into phases to be executed, continuing the original process to fine-tune the vehicle's attitude and position.

[0247] The automatic U-turn method for vehicles provided in this application generates a correction action sequence to perform local attitude fine-tuning if the verification fails, eliminating the need for a global search and reducing computational overhead. The existing phase-separated execution framework is reused to complete the correction motion, and all conditions are re-verified after correction. This approach avoids the vehicle remaining in a state of excessive deviation, touching the edge, or being unable to leave, ensuring the safety of its position and attitude and its ability to leave after the U-turn, thus improving the final output effect of the U-turn.

[0248] Figure 3 A schematic flowchart (II) of an embodiment of the automatic vehicle turning method provided in this application is shown below. Figure 3 As shown, the automatic vehicle turning method provided in this embodiment includes the following steps:

[0249] S301, in response to receiving an automatic U-turn command, acquires multi-source sensor data, including visual perception data, ultrasonic perception data, and current vehicle status data.

[0250] S302. Generate a current multi-attribute grid map based on multi-source sensor data. The current multi-attribute grid map includes grid coordinates, obstacle occupancy probability, obstacle height, obstacle speed, confidence level, and blind spot level.

[0251] S303. Generate a drivable area for turning around based on the current multi-attribute raster map.

[0252] S304. Generate at least one candidate target pose within the drivable area for U-turns according to the preset adaptation rules.

[0253] S305. A pre-trained action graph model is used to search for at least one candidate action sequence by starting the current vehicle pose as the graph node, ending the candidate target pose as the graph node, using action primitives as edges, and using the drivable area for U-turns as the boundary. The graph nodes include vehicle gear, blind spot level, and recoverable markers. The candidate action sequence includes at least one of the following: a single U-turn sequence, a three-point U-turn sequence, and a multi-point U-turn sequence.

[0254] S306. Predict the vehicle pose at each time step corresponding to each action in the candidate action sequence.

[0255] S307. Calculate the safe expansion distance based on the confidence level of the grid where the current pose is located, the grid blind zone level, the speed of dynamic obstacles, and the action type.

[0256] S308. Expand the region to be expanded according to the action type and the safe expansion distance to generate the sweep envelope of the vehicle pose at each moment.

[0257] S309. Calculate the spatiotemporal overlap point between the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence and the predicted trajectory of the dynamic obstacle at the same moment.

[0258] S310. Temporal conflict cost of generating each candidate action sequence based on spatiotemporal overlap points.

[0259] S311. If the spatiotemporal overlap point will leave the envelope area before the vehicle arrives in the predicted trajectory, the temporal conflict cost will be reduced.

[0260] S312. If there is no spatiotemporal overlap point at present, but the predicted trajectory overlaps with the sweep envelope at a future time, then increase the temporal conflict cost.

[0261] S313. Calculate the total cost corresponding to each candidate action sequence based on static collision cost, temporal conflict cost, blind zone cost, boundary cost, number of gear shifts, curvature change, and target pose error.

[0262] S314. Determine the target action sequence based on the total cost corresponding to each candidate action sequence.

[0263] S315. Divide the target action sequence into multiple phases to be executed, each phase to be executed including constraints.

[0264] S316. When executing each phase to be executed, obtain the real-time position and pose of the vehicle.

[0265] S317. Based on the real-time vehicle pose, periodically update the current multi-attribute grid map and reversible spatiotemporal occupancy envelope.

[0266] S318. When a preset abnormal condition is triggered and the vehicle comes to a stop, at least one candidate action sequence is regenerated based on the updated multi-attribute raster map.

[0267] S319. When the vehicle travels to the target area corresponding to the target pose, verify the vehicle center position error, heading angle error, distance between the vehicle body and the road boundary, exit space, gear status, and obstacle conditions.

[0268] S320. If the test fails, a corrective action sequence is generated and execution continues.

[0269] It should be noted that the execution order of S311 and S312 is not important.

[0270] In this embodiment, the implementation method and technical effect of S301-S320 are similar to those of the corresponding solutions in the above embodiments, and will not be repeated here.

[0271] Figure 4 A flowchart (III) illustrating an embodiment of the automatic vehicle turning method provided in this application is shown below. Figure 4 As shown, the automatic vehicle turning method provided in this embodiment includes the following steps:

[0272] S401. Determine the target action sequence based on the total cost corresponding to each candidate action sequence.

[0273] S402. If the total cost corresponding to each candidate action sequence does not meet the threshold of the target action sequence, and the recoverable flag in any candidate action sequence is true, then at least one probe action is generated.

[0274] S403. Update the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence based on the perception results corresponding to the detection action.

[0275] S404. If the blind zone cost of the target action sequence is greater than the preset cost threshold, then at least one detection action is generated.

[0276] S405. When the perception result corresponding to the detection action does not meet the execution conditions, adjust the target action sequence.

[0277] It should be noted that the execution order of S402-S403 and S404-S405 is not important.

[0278] In this embodiment, the implementation method and technical effect of S401-S405 are similar to those of the corresponding solutions in the above embodiments, and will not be repeated here.

[0279] Figure 5 This is a structural schematic diagram of the automatic vehicle turning device provided in this application, as shown below. Figure 5 As shown, the automatic vehicle turning device 50 provided in this embodiment includes: a generation module 51, a calculation module 52, an analysis module 53, a determination module 54, and an execution module 55.

[0280] The system includes a generation module 51, which generates at least one candidate action sequence based on the current multi-attribute raster map and a preset action primitive library; a calculation module 52, which calculates the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence; an analysis module 53, which performs an intersection analysis between the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence and the predicted trajectory of the dynamic obstacle; a determination module 54, which determines the target action sequence based on the intersection analysis results; and an execution module 55, which performs an automatic turn based on the target action sequence.

[0281] The automatic vehicle turning device provided in this embodiment can perform... Figures 2-4 The implementation principles and technical effects of the methods shown are similar, and will not be repeated here.

[0282] Optionally, the generation module 51, when generating candidate action sequences, is specifically used to: use a pre-trained action graph model with the current vehicle pose as the starting graph node and action primitives as edges to search for at least one candidate action sequence; the graph node includes vehicle gear, blind spot level, and recoverable marker; the candidate action sequence includes at least one of a single U-turn sequence, a three-point U-turn sequence, and a multi-point U-turn sequence.

[0283] Optionally, the generation module 51 is also used to generate a U-turn drivable area based on the current multi-attribute raster map; and to generate at least one candidate target pose within the U-turn drivable area according to a preset adaptation rule, so that when searching, the candidate target pose is used as the termination graph node and the U-turn drivable area is used as the boundary to generate a candidate action sequence.

[0284] Optionally, the calculation module 52, when calculating the reversible spatiotemporal occupancy envelope, is specifically used to: predict the vehicle pose at each time corresponding to each action in the candidate action sequence; and calculate the sweep envelope of the vehicle pose at each time using a preset dilation rule corresponding to the action type, wherein the sweep envelope includes a safety dilation distance.

[0285] Optionally, the preset expansion rules include the expansion region corresponding to each action type. The calculation module 52 is also used to calculate the safe expansion distance based on the confidence level of the grid where the current pose is located, the grid blind zone level, the dynamic obstacle speed, and the action type; and to expand the expansion region according to the action type and the safe expansion distance to generate a sweep envelope.

[0286] Optionally, the analysis module 53, when performing intersection analysis between the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence and the predicted trajectory of the dynamic obstacle, is specifically used to: calculate the spatiotemporal overlap point between the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence and the predicted trajectory of the dynamic obstacle at the same moment; and generate the temporal conflict cost of each candidate action sequence based on the spatiotemporal overlap point.

[0287] Optionally, the analysis module 53 is further configured to reduce the temporal conflict cost if the spatiotemporal overlap point in the predicted trajectory will leave the envelope region before the vehicle arrives; and to increase the temporal conflict cost if there is no spatiotemporal overlap point at present, but the predicted trajectory overlaps with the sweep envelope at a future time.

[0288] Optionally, the determining module 54, when determining the target action sequence, is specifically used to: calculate the total cost corresponding to each candidate action sequence based on static collision cost, temporal conflict cost, blind zone cost, boundary cost, number of gear shifts, curvature change, and target pose error; and determine the target action sequence based on the total cost corresponding to each candidate action sequence.

[0289] Optionally, the automatic vehicle turning device provided in this embodiment also includes an update module.

[0290] Correspondingly, the generation module 51 is also used to generate at least one detection action if the total cost corresponding to each candidate action sequence does not meet the threshold of the target action sequence and the recoverable flag in any candidate action sequence is true; the update module is used to update the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence based on the perception result corresponding to the detection action.

[0291] Optionally, the automatic vehicle turning device provided in this embodiment also includes an adjustment module.

[0292] Correspondingly, the generation module 51 is also used to generate at least one detection action if the blind zone cost of the target action sequence is greater than a preset cost threshold; the adjustment module is used to adjust the target action sequence when the perception result corresponding to the detection action does not meet the execution conditions.

[0293] Optionally, when performing an automatic U-turn, the execution module 55 is specifically used to: divide the target action sequence into multiple phases to be executed, each phase to be executed including constraints; and when executing each phase to be executed, acquire the real-time pose of the vehicle.

[0294] Optionally, the update module is also used to periodically update the current multi-attribute grid map and reversible spatiotemporal occupancy envelope based on the real-time pose of the vehicle; the generation module 51 is also used to regenerate at least one candidate action sequence based on the updated multi-attribute grid map when the vehicle stops due to a preset abnormal condition.

[0295] Optionally, the automatic vehicle turning device provided in this embodiment also includes an acquisition module.

[0296] Correspondingly, the acquisition module is used to acquire multi-source sensor data in response to receiving an automatic U-turn command, including visual perception data, ultrasonic perception data and current vehicle status data; the generation module 51 is also used to generate a current multi-attribute grid map based on the multi-source sensor data. The current multi-attribute grid map includes grid coordinates, obstacle occupancy probability, obstacle height, obstacle speed, confidence level, and blind spot level.

[0297] Optionally, the preset action primitive library includes basic action primitives and detection action primitives; each action primitive includes direction, curvature range, steering angular velocity threshold, speed threshold, driving distance range, shift requirements, abort conditions, and recoverable conditions.

[0298] Optionally, the automatic vehicle turning device provided in this embodiment also includes a verification module.

[0299] Correspondingly, the verification module is used to verify the vehicle center position error, heading angle error, distance between the vehicle body and the road boundary, driving space, gear status and obstacle status when the vehicle drives to the target area corresponding to the target pose; the generation module 51 is also used to generate a correction action sequence to continue execution if the verification fails.

[0300] Figure 6 A schematic diagram of the structure of the electronic device provided in this application. Figure 6 As shown, the electronic device 60 provided in this embodiment includes a processor 61 and a memory 62. The processor 61 and the memory 62 are connected via a bus and communicate with each other.

[0301] In the specific implementation process, the processor 61 executes the computer execution instructions stored in the memory 62, causing the processor 61 to perform the above-described method.

[0302] The specific implementation process of processor 61 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0303] In the above embodiments, it should be understood that the processor 61 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0304] The memory 62 may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0305] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0306] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0307] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0308] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0309] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0310] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0311] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0312] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0313] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0314] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0315] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for automatic vehicle U-turn, characterized in that, The method includes: Generate at least one candidate action sequence based on the current multi-attribute raster map and the preset action primitive library; Calculate the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence; The intersection analysis is performed between the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence and the predicted trajectory of dynamic obstacles. The target action sequence is determined based on the intersection analysis results, and an automatic U-turn is executed according to the target action sequence.

2. The method according to claim 1, characterized in that, Generate candidate action sequences, including: A pre-trained action graph model is used to search for at least one candidate action sequence by starting with the current vehicle pose as the graph node and using action primitives as edges. The graph nodes include vehicle gear, blind spot level, and recoverable markers. The candidate action sequence includes at least one of the following: single U-turn sequence, three-point U-turn sequence, and multi-point U-turn sequence.

3. The method according to claim 2, characterized in that, Also includes: Generate a driveable area for U-turns based on the current multi-attribute raster map; At least one candidate target pose is generated within the drivable area for U-turns according to a preset adaptation rule, so that when searching, the candidate target pose is used as the termination graph node and the drivable area for U-turns is used as the boundary to generate a candidate action sequence.

4. The method according to claim 1, characterized in that, Calculating the reversible spacetime occupancy envelope includes: Predict the vehicle pose at each time step for each action in the candidate action sequence; The sweep envelope of the vehicle pose at each moment is calculated using a preset dilation rule corresponding to the action type. The sweep envelope includes the safety dilation distance.

5. The method according to claim 4, characterized in that, The preset expansion rules include the regions to be expanded corresponding to each action type, and the sweep envelope is calculated, including: Calculate the safe expansion distance based on the confidence level of the grid where the current pose is located, the grid blind zone level, the speed of dynamic obstacles, and the action type; The region to be expanded is expanded according to the action type and the safe expansion distance to generate a sweep envelope.

6. The method according to claim 2, characterized in that, Intersection analysis, including: Calculate the spatiotemporal overlap point between the reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence and the predicted trajectory of the dynamic obstacle at the same moment; The temporal conflict cost of generating each candidate action sequence based on spatiotemporal overlap points.

7. The method according to claim 6, characterized in that, The cost of generating timing conflicts includes: If the spatiotemporal overlap point leaves the envelope region before the vehicle arrives in the predicted trajectory, the cost of temporal conflict is reduced. If there are no spatiotemporal overlap points at present, but the predicted trajectory overlaps with the sweep envelope at a future time, the cost of temporal conflict will increase.

8. The method according to claim 6, characterized in that, Determine the target action sequence, including: The total cost corresponding to each candidate action sequence is calculated based on static collision cost, temporal conflict cost, blind zone cost, boundary cost, number of gear shifts, curvature change, and target pose error. The target action sequence is determined based on the total cost corresponding to each candidate action sequence.

9. The method according to claim 8, characterized in that, Also includes: If the total cost corresponding to each candidate action sequence does not meet the threshold of the target action sequence, and the recoverable flag in any candidate action sequence is true, then at least one probe action is generated. The reversible spatiotemporal occupancy envelope corresponding to each candidate action sequence is updated based on the perception results corresponding to the detection action.

10. The method according to claim 8, characterized in that, Also includes: If the blind zone cost of the target action sequence is greater than a preset cost threshold, then at least one detection action is generated. When the perception result corresponding to the detection action does not meet the execution conditions, the target action sequence is adjusted.

11. The method according to claim 1, characterized in that, Perform automatic U-turn, including: The target action sequence is divided into multiple phases to be executed, and each phase to be executed includes constraints. During the execution of each phase to be executed, the real-time pose of the vehicle is acquired.

12. The method according to claim 11, characterized in that, Also includes: Based on the vehicle's real-time pose, the current multi-attribute raster map and reversible spatiotemporal occupancy envelope are periodically updated; When a preset abnormal condition is triggered and the vehicle comes to a stop, at least one candidate action sequence is regenerated based on the updated multi-attribute raster map.

13. The method according to claim 1, characterized in that, Also includes: In response to receiving an automatic U-turn command, it acquires multi-source sensor data, including visual perception data, ultrasonic perception data, and current vehicle status data; The current multi-attribute raster map is generated based on multi-source sensor data. The current multi-attribute raster map includes raster coordinates, obstacle occupancy probability, obstacle height, obstacle speed, confidence level, and blind spot level.

14. The method according to claim 1, characterized in that, The preset action primitive library includes basic action primitives and detection action primitives; Each action primitive includes direction, curvature range, steering angular velocity threshold, speed threshold, driving distance range, shift requirements, abort conditions, and recoverable conditions.

15. The method according to claim 1, characterized in that, Also includes: When the vehicle travels to the target area corresponding to the target pose, check the vehicle center position error, heading angle error, distance between the vehicle body and the road boundary, driving space, gear status and obstacle conditions. If the test fails, a corrective action sequence is generated and execution continues.

16. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-15.

17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-15.

18. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method according to any one of claims 1-15.

Images