Decision-making methods, devices, and mixed-traffic control systems for autonomous vehicles

Abstract

This disclosure provides a decision-making method, device, and vehicle mixed-traffic control system for autonomous vehicles. The method includes: detecting obstacles in the autonomous vehicle's driving direction; identifying the type of key obstacles when they are detected; wherein key obstacles include obstacles that may interact with the autonomous vehicle; generating sampling paths within a preset area in the autonomous vehicle's driving direction, and performing collision detection on the sampling paths based on key obstacles of a first preset type to select a first sampling path; determining the interaction intention between the autonomous vehicle and key obstacles of a second preset type based on a first dynamic intention of key obstacles of a second preset type and a second dynamic intention of the autonomous vehicle while driving on the first sampling path; evaluating the interaction intention, and generating a driving strategy for the autonomous vehicle based on the evaluation results. This solves the problem of safe and efficient decision-making for autonomous vehicles facing multiple obstacles in complex dynamic traffic environments.

Description

Technical Field

[0001] This disclosure relates to the field of autonomous driving technology, and in particular to a decision-making method, device and mixed-traffic control system for an autonomous vehicle. Background Technology

[0002] In the field of intelligent transportation, the development of autonomous driving technology is of great significance for improving traffic safety and efficiency. For autonomous driving scenarios, safety is a basic requirement for protecting public life and property, while efficiency directly affects the practicality and promotional value of the technology.

[0003] Currently, autonomous driving operation scenarios generally face challenges of complexity. Regarding road conditions, in addition to structured urban roads, there are numerous unstructured or semi-structured roads, with some sections lacking clear lane markings, making it difficult for vehicles to regulate their driving trajectories using traditional lane markings. Simultaneously, traffic intersections are diverse in form, with irregular roundabouts and multi-directional intersections being common, and the open road environment further increases the uncertainty of driving decisions.

[0004] Furthermore, in actual operation, autonomous vehicles need to frequently interact with a large number of manned vehicles, non-motorized vehicles, and pedestrians. In scenarios such as logistics and public transportation, autonomous vehicles have frequent contact with auxiliary equipment and other vehicles, resulting in complex and varied interactions. Against this backdrop, improving the intelligence of autonomous vehicles' interactions with other dynamic traffic participants while ensuring safety has become a core requirement for the development of autonomous driving technology. This not only requires more advanced perception technologies to accurately identify various targets but also relies on efficient decision-making algorithms to achieve safe and efficient driving in complex scenarios. Summary of the Invention

[0005] This disclosure provides a decision-making method, device, and mixed-traffic control system for unmanned vehicles to address the problems existing in current technical solutions.

[0006] Based on the above problems, firstly, a decision-making method for autonomous vehicles is provided, including:

[0007] Detect obstacles in the direction the autonomous vehicle is traveling;

[0008] If a critical obstacle is detected, the type of the critical obstacle is identified; wherein, the critical obstacle includes obstacles that may interact with the autonomous vehicle.

[0009] A sampling path is generated within a preset area in the direction of travel of the unmanned vehicle, and collision detection is performed on the sampling path based on key obstacles of a first preset type to select the first sampling path;

[0010] Based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the unmanned vehicle when it travels on the first sampling path, the interaction intent between the unmanned vehicle and the key obstacle of the second preset type is determined.

[0011] The interaction intent is evaluated, and a driving strategy for the autonomous vehicle is generated based on the evaluation results.

[0012] In conjunction with the first aspect, in one possible implementation, the obstacle includes: non-interactive obstacles and key obstacles; the key obstacles include: key static obstacles, abnormal obstacles, and key dynamic obstacles; the key static obstacles include: constant static obstacles and active static obstacles.

[0013] In conjunction with the first aspect, in one possible implementation, generating a sampling path within a preset area along the autonomous vehicle's driving direction, and performing collision detection on the sampling path based on key obstacles of a first preset type to select a first sampling path includes:

[0014] Sampling points are generated within a preset area in the direction of travel of the unmanned vehicle, and the sampling points are subjected to a first collision detection based on key obstacles of a first preset type to filter out the first sampling points;

[0015] Multiple sampling paths are generated based on the first sampling point, and a second collision detection is performed on the multiple sampling paths based on key obstacles of a first preset type to filter out the first sampling path.

[0016] In conjunction with the first aspect, in one possible implementation, the preset area range includes: longitudinal length and lateral width;

[0017] Sampling points are generated within a preset area along the driving direction of the autonomous vehicle, including:

[0018] Based on the longitudinal remaining length of the predicted trajectory of the unmanned vehicle and / or the longitudinal position of the key obstacle, determine the longitudinal sampling interval over the longitudinal length;

[0019] According to the longitudinal sampling interval, longitudinal sampling points are generated along the longitudinal length in the direction of travel of the unmanned vehicle;

[0020] Based on the lateral position of the key obstacle and the lateral width of the drivable space in the lateral direction of each longitudinal sampling point, the lateral sampling interval on the corresponding lateral width of each longitudinal sampling point is determined.

[0021] In the horizontal direction at the location of each vertical sampling point, horizontal sampling points are generated according to the corresponding horizontal sampling interval;

[0022] The longitudinal length is determined based on the driving speed of the autonomous vehicle; the lateral width is determined based on the lateral drivable boundary of the autonomous vehicle, and satisfies the condition that the lateral width is narrower the closer to the autonomous vehicle.

[0023] In conjunction with the first aspect, in one possible implementation, the key obstacles of the first preset type include abnormal obstacles and key static obstacles;

[0024] The first collision detection of the sampling points based on key obstacles of a first preset type includes:

[0025] Determine whether the sampling point overlaps with the first preset type of key obstacle. If there is an overlap, remove the sampling point to obtain a first sampling point that does not overlap with the first preset type of key obstacle.

[0026] The second collision detection is performed on the multiple sampling paths based on key obstacles of a first preset type to filter out the first sampling path, including:

[0027] Determine whether the multiple sampling paths intersect with the key obstacles of the first preset type. If they intersect, remove the sampling path to obtain a first sampling path that does not intersect with the key obstacles of the first preset type.

[0028] In conjunction with the first aspect, in one possible implementation, after obtaining the first sampling point, the method further includes:

[0029] The first sampling points that exceed the steering capability of the unmanned vehicle are excluded from the first sampling points to obtain the first sampling points after further filtering.

[0030] Multiple sampling paths are generated based on the first sampling point, including:

[0031] Multiple sampling paths are generated based on the first sampling point after the second filtering.

[0032] In conjunction with the first aspect, in one possible implementation, the key obstacle of the second preset type includes a key dynamic obstacle;

[0033] The step of determining the interaction intent between the autonomous vehicle and the second preset type of key obstacle based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the autonomous vehicle while driving on the first sampling path includes:

[0034] The predicted trajectory of the key dynamic obstacle is obtained, and various accelerations of the key dynamic obstacle as it travels on the predicted trajectory are sampled to obtain the first dynamic intent;

[0035] For each first sampling path, the various accelerations of the unmanned vehicle traveling on that first sampling path are sampled to obtain the second dynamic intent;

[0036] By combining the first dynamic intent and the second dynamic intent with the first sampling path, an interaction intent is determined for the unmanned vehicle to avoid interacting with the key dynamic obstacle during its journey along the first sampling path.

[0037] In conjunction with the first aspect, in one possible implementation, evaluating the interaction intent and generating the driving strategy of the autonomous vehicle based on the evaluation result includes:

[0038] Each interaction intent is evaluated based on preset metrics; wherein the preset metrics include at least one of the following: security, human-likeness, and accessibility;

[0039] The interaction intentions are evaluated and ranked according to the evaluation results of each interaction intention;

[0040] Based on the requirement of multiple intents, the sorted interaction intents are classified and coded, and duplicate intents at the end of the sort are removed for each classification code;

[0041] The driving strategy of the autonomous vehicle is generated based on the remaining interaction intentions after removing duplicate intentions and the evaluation and ranking results of the remaining interaction intentions.

[0042] In conjunction with the first aspect, in one possible implementation, before determining the interaction intention of the autonomous vehicle with the second preset type of key obstacle while traveling on the first sampling path based on the first dynamic intention of the key obstacle of the second preset type and the second dynamic intention of the autonomous vehicle, the method further includes:

[0043] The first sampling path is evaluated and ranked based on at least one of the following: path shape change variance, distance from reference line, distance to active static obstacle, and end point distance, to obtain a preset number of first sampling paths ranked first.

[0044] Among them, the smaller the variance of the path shape change, the higher the corresponding evaluation value; the smaller the distance from the reference line, the higher the corresponding evaluation value; the larger the distance to the active static obstacle, the higher the corresponding evaluation value; and the smaller the distance to the endpoint, the higher the corresponding evaluation value.

[0045] The step of determining the interaction intent between the autonomous vehicle and the second preset type of key obstacle based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the autonomous vehicle while driving on the first sampling path includes:

[0046] Based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the unmanned vehicle when driving on the first sampled path of the first preset number, the interaction intent between the unmanned vehicle and the key obstacle of the second preset type is determined.

[0047] In conjunction with the first aspect, in one possible implementation, the step of identifying the type of the key obstacle upon detection includes:

[0048] If a critical obstacle is detected, the type of the obstacle is identified based on the reference line projection of the critical obstacle;

[0049] The method further includes:

[0050] When the obstacle is an abnormal obstacle, the obstacle is expanded to form an abnormal driving area as the corresponding area of ​​the obstacle;

[0051] When the obstacle is one that allows normal interaction with the unmanned vehicle, the specified area near the predicted trajectory of the obstacle is obtained as the corresponding area of ​​the obstacle.

[0052] SL maps are constructed based on the corresponding regions of different types of obstacles.

[0053] Secondly, a decision-making device for an autonomous vehicle is provided, comprising:

[0054] The obstacle detection module is used to detect obstacles in the direction the autonomous vehicle is traveling.

[0055] An obstacle recognition module is used to identify the type of a key obstacle when it is detected; wherein the key obstacle includes obstacles that may interact with the unmanned vehicle.

[0056] The collision test module is used to generate a sampling path within a preset area in the direction of travel of the unmanned vehicle, and to perform collision detection on the sampling path based on key obstacles of a first preset type, and to select the first sampling path.

[0057] The interaction intent module is used to determine the interaction intent between the unmanned vehicle and the key obstacle of the second preset type based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the unmanned vehicle when it travels on the first sampling path.

[0058] The strategy generation module is used to evaluate the interaction intent and generate the driving strategy of the autonomous vehicle based on the evaluation results.

[0059] Thirdly, a vehicle mixed-traffic control system is provided, including a decision-making device for an unmanned vehicle as described in the second aspect, or in any possible embodiment in conjunction with the second aspect.

[0060] The beneficial effects of the embodiments disclosed herein include:

[0061] This disclosure provides a decision-making method, device, and vehicle mixed-traffic control system for autonomous vehicles, applicable to interactions between autonomous vehicles and numerous other vehicles in various complex scenarios, as well as responses to irregular obstacles. The specific decision-making method includes: First, detecting obstacles in the autonomous vehicle's driving direction and acquiring their position, speed, and other information to provide a data foundation for subsequent decisions. Next, identifying the type of key obstacles that may lead to interaction, distinguishing between dynamic objects such as pedestrians and vehicles and static objects such as guardrails, and handling different obstacles in different ways. Then, generating sampling paths in a preset area, performing collision detection based on first preset type key obstacles, and selecting first sampling paths with no direct collision risk to reduce decision complexity. Next, determining the autonomous vehicle's interaction intent towards the second preset type of obstacle based on the first dynamic intent of the second preset type key obstacle and the second dynamic intent of the autonomous vehicle while driving on the first sampling path, ensuring the safety and accuracy of the interaction intent from two aspects. Finally, performing multi-dimensional evaluation of the interaction intent to generate the final driving strategy. In summary, this method solves the problem of safe and efficient decision-making for unmanned vehicles facing obstacles in dynamic traffic environments. It processes different types of obstacles in a hierarchical manner, combines intent analysis and multi-dimensional evaluation to reduce accident risks, and realizes driving decisions for unmanned vehicles in complex road conditions. Attached Figure Description

[0062] Figure 1 One of the flowcharts for a decision-making method for an unmanned vehicle provided in this disclosure embodiment;

[0063] Figure 2 Schematic diagrams illustrating various interaction types provided in embodiments of this disclosure;

[0064] Figure 3 This is a schematic diagram of a collision test provided in an embodiment of the present disclosure;

[0065] Figure 4 This is one of the acceleration sampling diagrams provided in the embodiments of this disclosure;

[0066] Figure 5 This is the second schematic diagram of acceleration sampling provided in the embodiments of this disclosure;

[0067] Figure 6 A flowchart for obstacle type identification provided in this embodiment of the disclosure;

[0068] Figure 7A second flowchart illustrating a decision-making method for an unmanned vehicle provided in this embodiment of the disclosure;

[0069] Figure 8 This is a schematic diagram of a decision-making device for an unmanned vehicle provided in an embodiment of this disclosure. Detailed Implementation

[0070] This disclosure provides a decision-making method, apparatus, and vehicle mixed-traffic control system for unmanned vehicles. Preferred embodiments of this disclosure are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit this disclosure. Furthermore, the embodiments and features described in this application can be combined with each other unless otherwise specified.

[0071] This disclosure provides a decision-making method for autonomous vehicles, such as... Figure 1 As shown, it can be implemented as follows:

[0072] S101. Detect obstacles in the direction of travel of the unmanned vehicle;

[0073] S102. If a critical obstacle is detected, the type of the critical obstacle is identified; whereby critical obstacles include obstacles that may interact with the autonomous vehicle.

[0074] S103. Generate a sampling path in a preset area within the driving direction of the unmanned vehicle, and perform collision detection on the sampling path based on key obstacles of the first preset type to select the first sampling path.

[0075] S104. Based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the unmanned vehicle when driving on the first sampling path, determine the interaction intent between the unmanned vehicle and the key obstacle of the second preset type.

[0076] S105. Evaluate the interaction intent and generate the driving strategy for the autonomous vehicle based on the evaluation results.

[0077] When autonomous vehicles navigate complex road environments, they need to generate accurate driving decisions in real time to ensure safety and traffic efficiency. To address this, this disclosure provides a decision-making method for autonomous vehicles that addresses different types of obstacles and generates appropriate driving strategies.

[0078] In this embodiment of the disclosure, firstly, obstacles in the direction of travel of the unmanned vehicle are detected. This step uses sensors such as lidar, camera, and millimeter-wave radar deployed on the unmanned vehicle to obtain basic information such as the position, speed, and size of obstacles in its direction of travel, laying the data foundation for subsequent determination of the type of obstacle and generation of the final driving strategy of the unmanned vehicle.

[0079] Next, it is determined whether the detected obstacles are likely to interact with the autonomous vehicle. Obstacles that are likely to interact with the autonomous vehicle are considered key obstacles. After detecting key obstacles that may interact with the autonomous vehicle, their types are identified. The obstacle type identification can utilize technologies such as image recognition and deep learning algorithms, which are not limited here.

[0080] Next, multiple sampling paths are generated in a preset area along the autonomous vehicle's driving direction. Collision detection is then performed on key obstacles of the first preset type to eliminate paths with collision risks and select the theoretically safe first sampling path. This narrows down the scope of subsequent decisions and improves decision-making efficiency. The key obstacles of the first preset type are described in detail later and are static obstacles such as stationary vehicles and fixed barriers, and will not be elaborated here.

[0081] Subsequently, based on the dynamic intentions of pedestrians suddenly crossing the road, vehicles changing lanes, and other key obstacles of the second preset type, combined with the intentions of the unmanned vehicle when driving on the first sampling path, the interaction intentions of both parties are determined, and it is determined whether the unmanned vehicle needs to slow down to yield or accelerate to pass.

[0082] Finally, the interaction intent is evaluated from dimensions such as safety, traffic efficiency, and comfort, and the evaluation results are transformed into a series of specific driving strategies with execution priorities.

[0083] In summary, the autonomous vehicle decision-making method disclosed herein, by handling different types of obstacles in a hierarchical manner and combining dynamic intent analysis and multi-dimensional evaluation, not only achieves basic collision avoidance but also improves the rationality and diversity of decision-making in complex interaction scenarios. Compared to traditional decision-making methods, this method classifies and processes key obstacles, optimizes paths and strategies step by step, effectively reduces the risk of traffic accidents, and improves the efficiency and ride comfort of autonomous vehicles in complex road conditions.

[0084] In another embodiment provided in this disclosure, obstacles include: non-interactive obstacles and critical obstacles; critical obstacles include: critical static obstacles, abnormal obstacles and critical dynamic obstacles; critical static obstacles include: normal static obstacles and active static obstacles.

[0085] In the decision-making system of autonomous vehicles, the scientific classification of obstacles is the foundation for accurate decision-making. A reasonable obstacle classification framework enables autonomous vehicles to handle different types of objects in a targeted manner, thereby improving decision-making efficiency and safety. The following provides a detailed explanation of the obstacle classification system adopted in the solution provided in this disclosure and its function.

[0086] In this embodiment, obstacles are first categorized into non-interactive obstacles and critical obstacles. Non-interactive obstacles refer to objects that do not affect the autonomous vehicle's path, such as trees in the distance or fixed road signs that do not obstruct the lane. These obstacles do not require special attention, reducing computational resource consumption. Critical obstacles, on the other hand, are objects that may interact with the autonomous vehicle and are core processing targets in the decision-making process. For example, a method for identifying critical obstacles is provided, such as... Figure 2 As shown, based on the spatial geometric relationship between obstacles and the autonomous vehicle's trajectory, obstacles can be divided into overlapping obstacles that interact with the autonomous vehicle and non-overlapping obstacles that have potential interactions with the autonomous vehicle (such as...). Figure 2 (a) Overlapping obstacles can be further classified into point-overlapping obstacles (e.g., ...) based on the geometric characteristics of the overlapping area. Figure 2 (b) and line-overlapping obstacles (such as...) Figure 2 (c) The above three types of obstacles are all key obstacles that may interact with the autonomous vehicle.

[0087] Key obstacles are further subdivided into key static obstacles, abnormal obstacles, and key dynamic obstacles. Key static obstacles include constant static obstacles and active static obstacles. Constant static obstacles refer to objects that are fixed in position and stable in state, such as road barriers or vehicles waiting stationary in front of stop lines; active static obstacles, although stationary, may have potential dynamic changes, such as vehicles temporarily parked on the roadside with their doors open, where passengers may suddenly get off later, requiring autonomous vehicles to maintain a certain level of vigilance and allow sufficient space to respond.

[0088] Abnormal obstacles are objects that do not conform to normal movement patterns or whose behavior is difficult to predict, such as animals suddenly running into the road or dropped goods. These obstacles require autonomous vehicles to react quickly and replan their routes. Critical dynamic obstacles are objects with obvious movement characteristics that may interact with the autonomous vehicle, such as vehicles changing lanes or pedestrians crossing the road. The autonomous vehicle needs to track their dynamics in real time and adjust its driving strategy to ensure safe interaction.

[0089] In this embodiment, various objects that an autonomous vehicle may encounter during its operation are systematically classified. Through differentiated processing logic, precise responses to obstacles of different risk levels are achieved. Compared to simple obstacle differentiation methods, the classification provided in this disclosure helps autonomous vehicles quickly identify potential threats in complex environments, rationally allocate computing resources, effectively reduce collision risks, and significantly improve the decision-making accuracy and driving safety of autonomous vehicles in complex road conditions.

[0090] In another embodiment provided in this disclosure, step S103, which involves generating a sampling path within a preset area along the driving direction of the autonomous vehicle and performing collision detection on the sampling path based on key obstacles of a first preset type to select a first sampling path, can be implemented as follows:

[0091] Step 1: Generate sampling points in a preset area along the driving direction of the unmanned vehicle, and perform a first collision detection on the sampling points based on key obstacles of a first preset type to select the first sampling points;

[0092] Step 2: Generate multiple sampling paths based on the first sampling point, and perform a second collision detection on the multiple sampling paths based on key obstacles of the first preset type to select the first sampling path.

[0093] In the decision-making process of autonomous vehicles, path planning is a crucial step in ensuring safe driving. Selecting sampling paths based on key obstacles of a first preset type is a core step in optimizing the driving route while ensuring safety. To improve the accuracy and efficiency of path selection, this step can be further divided into two stages: sampling point selection and secondary path selection.

[0094] In this embodiment, the first stage involves generating sampling points within a preset area along the autonomous vehicle's driving direction. First, collision detection is performed on these sampling points based on key obstacles of a first preset type (such as stationary vehicles, fixed barriers, etc.) to filter out the first sampling points. This process is similar to marking potential feasible locations on a map. A collision detection algorithm quickly determines whether each sampling point will collide with a key obstacle of the first preset type, eliminating points with collision risk. For example, at a parking lot entrance / exit, if a stationary vehicle occupies part of the passage, this step can quickly eliminate sampling points within the vehicle-occupied area, retaining points within the safe area as the first sampling points. This step significantly narrows the range of subsequent path generation, reduces computational load, and lays the foundation for efficient decision-making.

[0095] In the second stage, a second collision detection is performed on multiple sampling paths generated from the first sampling points based on key obstacles of the first preset type, filtering out the first sampling path. Using the first sampling points as a basis, all first sampling points are interconnected to generate sampling paths, such as straight-line driving paths and paths with turns at different angles. Collision detection is then performed again on key obstacles of the first preset type to ensure that there is no collision risk throughout the path. For example, when driving on a narrow street, all possible sampling paths are generated based on the filtered first sampling points. After the second collision detection, a safe and reasonable detour path is retained as the first sampling path. This secondary detection further optimizes the path scheme and improves the reliability of path planning.

[0096] In summary, the method disclosed herein breaks down sampling path selection into sampling point selection and secondary path selection, employing a hierarchical and progressive processing logic. Compared to a single path selection method, this approach reduces overall computational complexity and improves the accuracy and safety of path planning by eliminating risk points and paths in stages. This enables autonomous vehicles to quickly find safe and feasible driving paths in complex environments, effectively reducing the risk of collisions with static or low-speed obstacles and enhancing the autonomous vehicle's ability to navigate and make decisions in complex road conditions.

[0097] In another embodiment provided in this disclosure, the preset area range includes: longitudinal length and transverse width;

[0098] In step 1 above, generating sampling points within a preset area along the autonomous vehicle's driving direction can be implemented as follows:

[0099] Step (1) Based on the longitudinal remaining length of the predicted trajectory of the unmanned vehicle and / or the longitudinal position of the key obstacle, determine the longitudinal sampling interval in the longitudinal length;

[0100] Step (2) Generate longitudinal sampling points along the longitudinal length of the unmanned vehicle's driving direction according to the longitudinal sampling interval;

[0101] Step (3) Based on the lateral position of the key obstacle and the lateral width of the drivable space in the lateral direction of each longitudinal sampling point, determine the lateral sampling interval on the lateral width corresponding to each longitudinal sampling point.

[0102] Step (4) Generate horizontal sampling points in the horizontal direction at the location of each vertical sampling point, according to the corresponding horizontal sampling interval;

[0103] The longitudinal length is determined based on the autonomous vehicle's driving speed; the lateral width is determined based on the autonomous vehicle's lateral drivable boundary, and satisfies the condition that the lateral width is narrower the closer to the autonomous vehicle.

[0104] In the process of autonomous vehicles planning paths based on key obstacles, the generation of sampling points within a pre-defined area is a core and fundamental step. The setting of the longitudinal length and lateral width of this area directly affects the rationality of the sampling point distribution and the accuracy of subsequent path planning. Through a systematic sampling point generation process, autonomous vehicles can more accurately capture feasible spaces, providing a reliable guarantee for safe driving.

[0105] In this embodiment of the disclosure, firstly, based on the remaining longitudinal length of the autonomous vehicle's predicted trajectory and / or the longitudinal position of key obstacles, a longitudinal sampling interval is determined along the longitudinal length. If there are no key obstacles a considerable distance ahead of the autonomous vehicle's predicted trajectory, a larger longitudinal sampling interval can be set; if approaching a key obstacle, such as a convoy of vehicles waiting to pass at an intersection ahead, the sampling interval is reduced to capture potential waypoints more densely. For example, on a highway, where the vehicle has a wide field of vision ahead, the longitudinal sampling interval can be set to 5 meters; while on a congested road section, when approaching the vehicle in front, the interval may be reduced to 1 meter, thus balancing computational efficiency and path planning accuracy.

[0106] Subsequently, longitudinal sampling points are generated within a preset longitudinal length range along the direction of travel of the autonomous vehicle, according to the longitudinal sampling interval, forming preliminary longitudinal path nodes to provide a benchmark for subsequent lateral sampling.

[0107] Next, based on the lateral position of the key obstacles and the lateral width of the drivable space in the lateral direction of each longitudinal sampling point, the lateral sampling interval on the corresponding lateral width is determined. For example, when the autonomous vehicle needs to go around a parked vehicle, the lateral drivable space close to the vehicle is narrow, so the lateral sampling interval is set to be small; the space far away from the vehicle is wide, so the interval is set to be large, thus accurately adapting to the traffic conditions of different locations.

[0108] Finally, in the lateral direction at the location of each longitudinal sampling point, lateral sampling points are generated according to the corresponding lateral sampling interval, completing the comprehensive layout of sampling points within the preset area. The longitudinal length is determined based on the autonomous vehicle's driving speed; a higher speed results in a longer preset longitudinal range. The lateral width is determined based on the autonomous vehicle's lateral drivable boundary, following the rule that the lateral width decreases as it approaches the autonomous vehicle, thus resulting in a smaller lateral sampling interval closer to the vehicle, simulating the spatial constraints of real-world driving scenarios.

[0109] In summary, by dynamically adjusting the longitudinal and lateral sampling intervals and considering factors such as the autonomous vehicle's speed, obstacle positions, and drivable boundaries, an adaptive distribution of sampling points is achieved. Compared to fixed-interval sampling, this method improves sampling flexibility and environmental adaptability, more accurately captures feasible path spaces, reduces invalid sampling, improves path planning efficiency and accuracy, and effectively ensures the safe and efficient operation of autonomous vehicles in complex road conditions.

[0110] In another embodiment provided in this disclosure, the key obstacles of the first preset type include abnormal obstacles and key static obstacles;

[0111] In step 1 above, performing the first collision detection on the sampling points based on the key obstacles of the first preset type can be implemented as follows:

[0112] Step a: Determine whether the sampling point overlaps with the key obstacle of the first preset type. If there is an overlap, remove the sampling point to obtain the first sampling point that does not overlap with the key obstacle of the first preset type.

[0113] Step b: Perform second collision detection on multiple sampling paths based on key obstacles of the first preset type, and filter out the first sampling path, including:

[0114] Step c: Determine whether multiple sampling paths intersect with key obstacles of the first preset type. If they intersect, remove the sampling path to obtain a first sampling path that does not intersect with key obstacles of the first preset type.

[0115] In the path planning process of autonomous vehicles, collision detection is a core component to ensure driving safety. Collision detection of critical obstacles of the first preset type (including abnormal obstacles and critical static obstacles) can effectively eliminate dangerous paths and ensure the safe passage of autonomous vehicles through complex road conditions. This detection process forms a rigorous layered screening mechanism through screening at two levels: sampling points and sampling paths.

[0116] In this embodiment of the disclosure, firstly, at the sampling point level, it is necessary to determine whether the sampling point overlaps with a key obstacle of the first preset type. If there is an overlap, it means that the sampling point is in the space occupied by the obstacle, and it will be removed, thus obtaining a first sampling point that does not overlap with the key obstacle of the first preset type. Figure 3 As shown, sampling points marked with an "X" are those that overlap with obstacles and should be removed. For example, when there are fallen goods (abnormal obstacles) on the road or parked vehicles (critical static obstacles) on the roadside, this step can quickly eliminate sampling points located at the locations of the goods or vehicles. This operation can initially filter out obviously infeasible path nodes, reducing the computational load of subsequent path planning.

[0117] Next, after generating multiple sampling paths based on the selected first sampling points, the collision detection phase of the sampling paths begins. This phase requires determining whether any of the multiple sampling paths intersect with key obstacles of a first preset type. If an intersection exists (e.g., ...), the collision detection will then proceed. Figure 3 (As shown by the dashed line), this indicates that the path will collide with obstacles during travel. These obstacles will be removed, resulting in the first sampled path that does not intersect with the first preset type of key obstacle. For example, in a narrow alley, among the multiple detour paths generated based on the first sample point, if a path is planned to pass through an area where vehicles are parked, that path will be excluded, and the path that safely avoids the obstacle will be retained.

[0118] This disclosure employs a layered collision detection method, first performing preliminary screening at the sampling point level, and then conducting in-depth investigation of the sampling path. Compared to a single collision detection method, it achieves avoidance of abnormal obstacles and critical static obstacles, improving the safety and reliability of path planning. By efficiently filtering dangerous paths, it can help autonomous vehicles quickly find safe and feasible driving routes in complex environments, effectively reducing the risk of collisions.

[0119] In another embodiment provided in this disclosure, after obtaining the first sampling point by performing step 1, the method further includes:

[0120] The first sampling points that exceed the steering capability of the autonomous vehicle are excluded from the first sampling points, resulting in a second set of first sampling points after further filtering.

[0121] Therefore, in step 2, generating multiple sampling paths based on the first sampling point can be implemented as follows:

[0122] Multiple sampling paths are generated based on the first sampling point after further filtering.

[0123] In the path planning process of autonomous vehicles, to ensure the practical feasibility of the generated sampling paths, further screening can be performed based on the autonomous vehicle's own performance. Therefore, in this embodiment, after obtaining the first sampling points based on collision detection of key obstacles of the first preset type, the steering capability of the autonomous vehicle is then considered. By excluding sampling points that exceed the steering capability, and generating sampling paths based on the results of the second screening, the matching between path planning and the actual handling performance of the vehicle is achieved.

[0124] In this embodiment, after obtaining the first sampling points, these sampling points are further filtered to exclude those that exceed the steering capability of the autonomous vehicle. The steering capability of an autonomous vehicle is limited by factors such as its mechanical structure and power system. If the spacing between sampling points is too large or the angle deviation exceeds the vehicle's steering limits, even if the path has no spatial obstacle collision risk, it cannot be practically executed. For example, in a narrow alley scenario, if the steering angle set for a certain first sampling point exceeds the maximum steering angle of the autonomous vehicle, the vehicle cannot complete the corresponding steering action, and that sampling point will be excluded. This step, by incorporating the vehicle's physical performance parameters into the filtering conditions, ensures that the remaining sampling points meet the practical operational feasibility of the autonomous vehicle.

[0125] Subsequently, multiple sampling paths are generated based on the first set of sampled points after further filtering. These sampled points, verified by steering capability, serve as the basis for path generation, ensuring that the generated paths can be executed by the autonomous vehicle in actual driving. For example, in a scenario where the vehicle is reversing out of a parking lot, the sampled points filtered by steering capability will guide the generation of a reversing path that conforms to the vehicle's steering characteristics, avoiding the planning of sharp turns or large-amplitude turns that the vehicle cannot achieve.

[0126] In this embodiment, the constraint of the autonomous vehicle's steering capability is incorporated into the process of selecting sampling points. This not only ensures a safe and collision-free path but also guarantees the feasibility of the path from the perspective of actual vehicle operation, improving the practicality and reliability of the autonomous vehicle's path planning. By accurately matching environmental feasibility with vehicle performance, decision-making errors caused by mismatch between path planning and vehicle capabilities are effectively reduced, enhancing the autonomous vehicle's ability to navigate complex road conditions and its operational stability.

[0127] In another embodiment provided in this disclosure, the key obstacle of the second preset type includes a key dynamic obstacle;

[0128] Step S104 above, determining the interaction intent between the autonomous vehicle and the key obstacle of the second preset type based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the autonomous vehicle when traveling on the first sampling path, can be implemented as follows:

[0129] Step 1: Obtain the predicted trajectory of the key dynamic obstacle and sample various accelerations of the key dynamic obstacle as it travels along the predicted trajectory to obtain the first dynamic intent.

[0130] Step 2: For each first sampling path, sample the various accelerations of the unmanned vehicle as it travels along that first sampling path to obtain the second dynamic intent;

[0131] Step 3: Combine the first dynamic intent and the second dynamic intent with the first sampling path to determine the interaction intent of the unmanned vehicle to avoid interacting with key dynamic obstacles during the first sampling path.

[0132] After the autonomous vehicle completes the initial path selection based on static and abnormal obstacles, it needs to further refine its driving strategy when facing dynamic interaction scenarios with key dynamic obstacles. For key obstacles of the second preset type (i.e., key dynamic obstacles), by sampling and combining their dynamic intentions with those of the autonomous vehicle, potential interaction risks can be effectively predicted, reasonable interaction intentions can be formulated, and the safe passage of the autonomous vehicle in dynamic traffic environments can be ensured.

[0133] In this embodiment of the disclosure, firstly, the predicted trajectory of the key dynamic obstacle is obtained, and then... Figure 4 and Figure 5As shown, various accelerations of a vehicle traveling along a predicted trajectory are sampled to obtain the first dynamic intent. For example, at an intersection, a vehicle approaching laterally acts as a key dynamic obstacle. Sensor data and a motion prediction model are used to obtain its trajectory over a future period, and various acceleration variations, such as acceleration, deceleration, and constant speed, are sampled. This step comprehensively captures the potential motion changes of the key dynamic obstacle, providing rich data support for subsequent interactive analysis and avoiding the risk of misjudgment due to considering only a single motion state.

[0134] Next, for each first sampling path, various accelerations of the autonomous vehicle traveling on that path are sampled to obtain the second dynamic intent. Taking the autonomous vehicle's plan to go straight through an intersection as an example, different acceleration states such as acceleration, deceleration, and maintaining a constant speed are sampled on the first sampling path to clarify the possible dynamic behaviors of the autonomous vehicle on that path and ensure that the autonomous vehicle's driving intent analysis meets the needs of actual driving scenarios.

[0135] Finally, the first and second dynamic intentions are combined with the first sampled path to determine the interaction intentions of the autonomous vehicle to avoid collisions with key dynamic obstacles during its journey along the first sampled path. By analyzing the spatiotemporal relationship between the key dynamic obstacles and the autonomous vehicle under different acceleration states, solutions that do not generate trajectory intersections or collision risks are selected. For example, in the intersection scenario mentioned above, if it is found that the autonomous vehicle maintaining a constant speed and traveling straight would pose a collision risk with an accelerating vehicle, then through combined analysis, reasonable driving strategies to avoid interaction, such as the autonomous vehicle slowing down to yield or accelerating in advance to pass, are determined.

[0136] In this embodiment, by performing multi-dimensional sampling and combined analysis of the dynamic intentions of the autonomous vehicle and key dynamic obstacles, and incorporating dynamic behavior changes into the interactive decision-making system, potential risks in complex traffic scenarios can be predicted more accurately, and realistic interactive intentions can be formulated. This effectively improves the scientific nature and safety of the autonomous vehicle's decision-making in dynamic traffic environments, reduces collision accidents caused by the uncertainty of dynamic obstacle behavior, and enhances the efficiency and reliability of the autonomous vehicle in complex road conditions.

[0137] In another embodiment provided in this disclosure, step S105, evaluating the interaction intent and generating a driving strategy for the autonomous vehicle based on the evaluation result, can be implemented as follows:

[0138] Step 1: Evaluate the preset metrics for each interaction intent; the preset metrics include at least one of the following: security, human-likeness, and accessibility;

[0139] Step 2: Rank the interaction intents based on the evaluation results of each intent.

[0140] Step 3: Based on the multi-intent requirements, classify and encode the sorted interaction intents, and remove duplicate intents that are ranked lower for each category code;

[0141] Step 4: Generate the driving strategy for the autonomous vehicle based on the remaining interaction intentions after removing duplicate intentions and the evaluation and ranking results of the remaining interaction intentions.

[0142] After determining the interaction intentions of the autonomous vehicle with key dynamic obstacles, these intentions need to be transformed into executable driving strategies. To this end, the interaction intentions are systematically evaluated and processed, their quality is measured from multiple dimensions, and they are filtered and optimized to ultimately generate driving strategies that meet actual needs, ensuring that the autonomous vehicle can drive safely and efficiently in complex traffic scenarios.

[0143] In this embodiment, firstly, preset indicators for each interaction intent are evaluated. These indicators cover aspects such as safety, human-likeness, and passability. The safety indicator assesses whether the interaction intent will lead to a collision between the autonomous vehicle and obstacles. For example, if the autonomous vehicle intends to slow down to avoid a pedestrian, it needs to be determined whether the driving trajectory after deceleration is safe. The human-likeness indicator considers whether the intent conforms to human driving habits, such as avoiding abrupt behaviors like sudden acceleration or braking. The passability indicator assesses whether the intent allows the autonomous vehicle to smoothly pass through the current road segment, such as the efficiency of passage through narrow passages. This step provides a quantitative basis for subsequent screening.

[0144] Next, based on the evaluation results of each interaction intent, the interaction intents are evaluated and ranked. For example, in the scenario of yielding to pedestrians at an intersection, interaction intents that are highly safe, conform to human driving habits, and have good passability are ranked first, while intents that pose a collision risk or are too abrupt are ranked later, thus forming a priority order to facilitate subsequent screening and decision-making.

[0145] Subsequently, based on the multi-intent requirement, the sorted interaction intents are classified and coded, and duplicate intents at the bottom of the sorting are removed for each category. For example, intents such as "slow down and give way" and "pass at a constant speed" are classified and coded. If there are multiple similar intents in a certain category, only the ones at the top of the sorting are retained, and duplicate and less effective intents are removed to reduce redundant decision information and improve decision efficiency.

[0146] Finally, a driving strategy for the autonomous vehicle is generated based on the remaining interaction intentions after removing duplicate intentions and the evaluation and ranking results of the remaining interaction intentions. The top-ranked interaction intentions are selected to generate a driving strategy with execution priority, guiding the autonomous vehicle's actual driving. The selection can be either an optimal interaction intention or a preset number of top-ranked interaction intentions; there is no restriction here.

[0147] In summary, the method provided in this disclosure selects superior autonomous vehicle driving decisions through multi-dimensional evaluation, ranking, and classification optimization. Compared to single-index evaluation, it considers multiple aspects such as safety, human-likeness, and maneuverability, balancing safety and traffic efficiency while reducing repetitive and ineffective decisions. This effectively improves the rationality and reliability of autonomous vehicle decisions, enabling them to make safer and more efficient driving decisions in complex traffic environments that are closer to human driving habits, thus enhancing the practicality and user acceptance of autonomous vehicles.

[0148] In another embodiment provided in this disclosure, before step S104 above, which determines the interaction intention of the unmanned vehicle with the key obstacle of the second preset type when traveling on the first sampling path based on the first dynamic intention of the key obstacle of the second preset type and the second dynamic intention of the unmanned vehicle, the method further includes:

[0149] The first sampling path is evaluated and ranked based on at least one of the following: path shape change variance, distance from reference line, distance to active static obstacle, and distance to endpoint, to obtain a preset number of first sampling paths ranked first.

[0150] Among them, the smaller the variance of path shape change, the higher the corresponding evaluation value; the smaller the distance from the reference line, the higher the corresponding evaluation value; the larger the distance to active and static obstacles, the higher the corresponding evaluation value; and the smaller the distance to the endpoint, the higher the corresponding evaluation value.

[0151] Furthermore, step S104 above, determining the interaction intent between the autonomous vehicle and the key obstacle of the second preset type based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the autonomous vehicle when driving on the first sampling path, can also be implemented as follows:

[0152] Based on the first dynamic intent of the key obstacles of the second preset type and the second dynamic intent of the autonomous vehicle when driving on the first sampled path of the first preset number, the interaction intent between the autonomous vehicle and the key obstacles of the second preset type is determined.

[0153] After the autonomous vehicle completes the initial path selection and obtains the first sampled paths, these paths differ in terms of safety, comfort, and feasibility. Directly using them for interaction intent analysis with key dynamic obstacles may lead to inefficient or unreasonable decision-making. Therefore, further quantitative evaluation and selection of the first sampled paths are needed to optimize the subsequent interaction decision-making process with key dynamic obstacles.

[0154] In this embodiment of the disclosure, firstly, the first sampled path is evaluated and ranked based on at least one of the following: path shape change variance, distance from the reference line, distance to active static obstacles, and endpoint distance. Path shape change variance reflects the smoothness of the path; a smaller variance means a smoother path, higher vehicle driving comfort, and a higher corresponding evaluation value. For example, on a straight road, a smooth straight driving path has a smaller shape change variance and a higher evaluation value compared to a path with frequent turns. Distance from the reference line measures the degree to which the path deviates from the ideal driving route; a smaller distance is closer to the expected driving direction and a higher evaluation value. A larger distance from active static obstacles indicates higher driving safety and a correspondingly higher evaluation value. For example, when there are temporarily parked vehicles on the roadside, paths further away from the vehicles are safer and have a higher evaluation value. Endpoint distance refers to the distance between the path's endpoint and the target point; a smaller distance indicates that the path is closer to the expected target and a higher evaluation value.

[0155] For calculating the evaluation value, taking the above four factors as an example, the weight calculation method can be used for accurate evaluation during implementation. For example, the following formula can be used for calculation.

[0156]

[0157] in, , , and Characteristic weight coefficients, Variance representing path morphology changes Characterizes the distance from the reference line. Characterizing the distance between endpoints, The distance between the character and the active static obstacle.

[0158] Using the aforementioned evaluation criteria, all first sampling paths are comprehensively scored and ranked according to their evaluation scores, resulting in a predetermined number of top-ranked first sampling paths. For example, in a complex urban street scenario, after screening, the top 5 first sampling paths with the highest evaluation scores are retained for the next stage of decision-making.

[0159] Subsequently, based on the first dynamic intent of the key obstacles of the second preset type and the second dynamic intent of the autonomous vehicle when traveling on the first sampled path ranked in the top preset number, the interaction intent between the autonomous vehicle and the key obstacles of the second preset type is determined. This step aims to generate the interaction intent between the autonomous vehicle and the key dynamic obstacles. Through the above implementation method, the interaction decision between the autonomous vehicle and the key dynamic obstacles is based on a safer and more reasonable path, reducing unnecessary calculation and decision-making errors.

[0160] In summary, the method provided in this disclosure effectively reduces the computational burden caused by redundant paths and improves decision-making efficiency by introducing multi-dimensional path evaluation metrics to screen and optimize the first sampled path. Furthermore, by selecting superior paths, the interaction decisions between the autonomous vehicle and key dynamic obstacles become more scientific and reasonable, enhancing the driving safety and comfort of the autonomous vehicle in complex traffic environments.

[0161] In another embodiment provided in this disclosure, step S102, identifying the type of the key obstacle when it is detected, can be implemented as follows:

[0162] When a critical obstacle is detected, the type of obstacle is identified based on the reference line projection of the critical obstacle;

[0163] Furthermore, the above methods also include:

[0164] In the case of an abnormal obstacle, the obstacle is expanded to form an abnormal driving area as the corresponding area of ​​the obstacle;

[0165] When the obstacle is one that allows normal interaction with the autonomous vehicle, the specified area near the predicted trajectory of the obstacle is obtained as the corresponding area of ​​the obstacle.

[0166] SL maps are constructed based on the corresponding regions of different types of obstacles.

[0167] In the decision-making process of autonomous vehicles, accurately identifying key obstacle types is fundamental for subsequent targeted processing. By using a reference line projection-based identification method, different types of key obstacles can be efficiently distinguished, and corresponding obstacle regions can be constructed based on these differences, ultimately forming a Level Layout (SL) map. This provides an intuitive basis for autonomous vehicles to plan safe paths. The SL map, specifically, is a map constructed using SLAM (Simultaneous Localization and Mapping) technology within a SLAM scenario. It is a digital representation of the surrounding environment by the autonomous vehicle. Autonomous vehicles use multiple sensors such as LiDAR and cameras to collect environmental information. SLAM algorithms process this data to determine the vehicle's real-time position while constructing a map containing information about roads, buildings, and obstacles. This map provides the foundation for the vehicle's path planning and obstacle avoidance functions, helping the vehicle achieve autonomous navigation in complex environments.

[0168] In this embodiment of the disclosure, such as Figure 6As shown, after detecting a key obstacle, its type is identified based on the obstacle's projection onto a reference line. The reference line can be set as the centerline of the autonomous vehicle's predetermined driving path or the road centerline. By analyzing the position, shape, and variation patterns of the obstacle's projection onto the reference line, its type is determined. For example, if the obstacle's projection is near the reference line and has an irregular shape, combined with sensor data, it can be identified as an abnormal obstacle that has suddenly entered the road; if the projection position is stable and conforms to the normal movement patterns of vehicles or pedestrians, it is likely an obstacle that allows normal interaction with the autonomous vehicle. This projection-based identification method utilizes geometric relationships to quickly locate features, reducing the computational complexity of identification.

[0169] Optionally, after identification, labels are output for each obstacle. When an obstacle is identified as an abnormal obstacle, it is expanded to form an abnormal driving area as the corresponding area for the obstacle. For example, if cargo suddenly falls onto the road, the space actually occupied by the cargo is expanded outward according to certain rules to generate a larger abnormal driving area, avoiding collisions caused by changes in cargo position or positioning errors of the autonomous vehicle, and reserving sufficient safe detour space for the autonomous vehicle.

[0170] Optionally, if the obstacle is one that allows normal interaction with the autonomous vehicle, the designated area near the predicted trajectory of the obstacle is obtained as the corresponding area of ​​the obstacle. Taking a vehicle traveling in the same direction as an example, the area a certain distance ahead of its predicted trajectory within the next three seconds is taken as the corresponding area. This focuses on the spatial range where interaction may occur in the short term, reducing unnecessary spatial calculations and focusing on the core area of ​​interaction decision-making.

[0171] Finally, a spatial mapping (SL) map is constructed based on the corresponding regions of different types of obstacles. By integrating the corresponding regions of various obstacles into a unified coordinate system, the relationship between obstacles and the autonomous vehicle's driving space is intuitively presented, providing a clear spatial reference for subsequent path planning, collision detection, and interaction intent analysis, and facilitating the autonomous vehicle to quickly determine feasible paths and risk areas.

[0172] In summary, the method disclosed in this publication combines reference line projection recognition with differentiated region construction. Compared to a single obstacle recognition method, it not only improves the efficiency of type recognition but also enhances the safety and flexibility of decision-making through targeted region processing. By constructing an SL map to achieve the visual integration of information on multiple types of obstacles, it provides strong support for the efficient decision-making of autonomous vehicles in complex environments, reducing the risk of accidents caused by misjudgment or improper handling of obstacles.

[0173] Figure 7 Another example of an autonomous vehicle decision-making process is also provided. For example... Figure 7 As shown, including:

[0174] First, obstacle decision-making begins, categorizing obstacles based on different interaction states (no interaction, normal dynamic interaction, normal static interaction, and abnormal). For normal dynamic obstacles, the predicted trajectory area within the first three seconds is sampled. For normal static and abnormal obstacles, operations such as expanding the coverage area of ​​the driving region are used to jointly construct the SL map. Subsequently, sampling of key obstacles at specific orientation positions and lateral sampling points based on driving intent are performed sequentially. Valid sampling points are filtered through collision detection, generating sampling trajectories and constructing corresponding obstacle situations. Then, intent encoding deduplication and regional dynamic obstacle acceleration sampling decisions are performed. Finally, a comprehensive evaluation of the sampling trajectories is conducted, ending the process. This completes the detection, sampling, and trajectory evaluation of various obstacles, ensuring driving safety and the rationality of decision-making.

[0175] Based on the same disclosed concept, embodiments of this disclosure also provide a decision-making device for an autonomous vehicle. For example... Figure 8 As shown, the device may include:

[0176] The obstacle detection module 801 is used to detect obstacles in the direction of travel of the unmanned vehicle;

[0177] The obstacle recognition module 802 is used to identify the type of key obstacles when they are detected; wherein, key obstacles include obstacles that may interact with the autonomous vehicle.

[0178] The collision test module 803 is used to generate a sampling path in a preset area in the direction of travel of the unmanned vehicle, and to perform collision detection on the sampling path based on key obstacles of a first preset type, and to select the first sampling path.

[0179] The interaction intent module 804 is used to determine the interaction intent between the unmanned vehicle and the key obstacle of the second preset type based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the unmanned vehicle when driving on the first sampling path.

[0180] The strategy generation module 805 evaluates the interaction intent and generates the driving strategy of the autonomous vehicle based on the evaluation results.

[0181] In another embodiment provided in this disclosure, obstacles include: non-interactive obstacles and critical obstacles; critical obstacles include: critical static obstacles, abnormal obstacles and critical dynamic obstacles; critical static obstacles include: normal static obstacles and active static obstacles.

[0182] In another embodiment provided in this disclosure, the collision test module 803 is used to generate sampling points in a preset area range in the driving direction of the unmanned vehicle, and to perform a first collision detection on the sampling points based on key obstacles of a first preset type to filter out the first sampling points; to generate multiple sampling paths based on the first sampling points, and to perform a second collision detection on the multiple sampling paths based on key obstacles of the first preset type to filter out the first sampling path.

[0183] In another embodiment provided in this disclosure, the preset area range includes: longitudinal length and transverse width;

[0184] The collision test module 803 is used to determine the longitudinal sampling interval along the longitudinal length based on the longitudinal remaining length of the predicted trajectory of the autonomous vehicle and / or the longitudinal position of the key obstacle; generate longitudinal sampling points along the longitudinal length in the direction of travel of the autonomous vehicle according to the longitudinal sampling interval; determine the lateral sampling interval along the lateral width corresponding to each longitudinal sampling point based on the lateral position of the key obstacle and the lateral width of the drivable space in the lateral direction of each longitudinal sampling point; and generate lateral sampling points in the lateral direction at the location of each longitudinal sampling point according to the corresponding lateral sampling interval; wherein, the longitudinal length is determined based on the driving speed of the autonomous vehicle; the lateral width is determined based on the lateral drivable boundary of the autonomous vehicle, and satisfies the condition that the lateral width is narrower the closer to the autonomous vehicle.

[0185] In another embodiment provided in this disclosure, the key obstacles of the first preset type include abnormal obstacles and key static obstacles;

[0186] The collision test module 803 is used to determine whether the sampling point overlaps with a key obstacle of the first preset type. If there is an overlap, the sampling point is removed to obtain a first sampling point that does not overlap with the key obstacle of the first preset type. It also determines whether multiple sampling paths intersect with the key obstacle of the first preset type. If there is an intersection, the sampling path is removed to obtain a first sampling path that does not intersect with the key obstacle of the first preset type.

[0187] In another embodiment provided in this disclosure, the collision test module 803 is further configured to exclude first sampling points that exceed the steering capability of the unmanned vehicle from the first sampling points, and obtain the first sampling points after further filtering; and generate multiple sampling paths based on the first sampling points after further filtering.

[0188] In another embodiment provided in this disclosure, the key obstacle of the second preset type includes a key dynamic obstacle;

[0189] The interaction intent module 804 is used to acquire the predicted trajectory of the key dynamic obstacle and sample the various accelerations of the key dynamic obstacle as it travels on the predicted trajectory to obtain a first dynamic intent; for each first sampling path, the module samples the various accelerations of the unmanned vehicle as it travels on the first sampling path to obtain a second dynamic intent; the module combines the first dynamic intent and the second dynamic intent with the first sampling path to determine the interaction intent of the unmanned vehicle to avoid interacting with the key dynamic obstacle during the travel of the unmanned vehicle on the first sampling path.

[0190] In another embodiment provided in this disclosure, the strategy generation module 805 is used to evaluate preset indicators for each interaction intent; wherein the preset indicators include at least one of the following: safety, human-likeness, and passability; according to the evaluation results of each interaction intent, the interaction intents are evaluated and ranked; the ranked interaction intents are classified and coded based on multi-intent requirements, and duplicate intents ranked lower are removed for each classification and coding; and a driving strategy for the autonomous vehicle is generated based on the remaining interaction intents after removing duplicate intents and the evaluation and ranking results of the remaining interaction intents.

[0191] In another embodiment provided in this disclosure, the interaction intent module 804 is further configured to evaluate and rank the first sampled path based on at least one of the path morphology change variance, the distance from the reference line, the distance to the active static obstacle, and the terminal distance, to obtain a first sampled path with a predetermined number of rankings; wherein, the smaller the path morphology change variance, the higher the corresponding evaluation value; the smaller the distance from the reference line, the higher the corresponding evaluation value; the larger the distance to the active static obstacle, the higher the corresponding evaluation value; the smaller the terminal distance, the higher the corresponding evaluation value; and the more the interaction intent between the autonomous vehicle and the second preset type of key obstacle is determined based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the autonomous vehicle when driving on the first sampled path with a predetermined number of rankings.

[0192] In another embodiment provided in this disclosure, the obstacle recognition module 802 is used to identify the type of obstacle based on the reference line projection of the key obstacle when a key obstacle is detected.

[0193] The aforementioned device further includes: an SL map construction module, used to expand the obstacle to form an abnormal driving area as the corresponding area of ​​the obstacle when the obstacle is an abnormal obstacle; to obtain the near-end specified area of ​​the predicted trajectory of the obstacle as the corresponding area of ​​the obstacle when the obstacle is an obstacle that interacts normally with the unmanned vehicle; and to construct an SL map based on the corresponding areas of the obstacles of different types of obstacles.

[0194] According to an embodiment of this application, a vehicle mixed-traffic control system is also provided, which may include the decision-making device for any of the aforementioned unmanned vehicles. Specific implementation details of this device can be found in the explanations of the above embodiments and will not be repeated here.

[0195] Through the above description of the embodiments, those skilled in the art can clearly understand that the embodiments of this disclosure can be implemented in hardware or by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this disclosure.

[0196] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes in the drawings are not necessarily essential for implementing this disclosure.

[0197] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0198] The sequence numbers of the embodiments disclosed above are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0199] Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, this disclosure is also intended to include such modifications and variations.

Claims

1. A decision-making method for an autonomous vehicle, characterized in that, include: Detect obstacles in the direction the autonomous vehicle is traveling; If a critical obstacle is detected, the type of the critical obstacle is identified; wherein, the critical obstacle includes obstacles that may interact with the autonomous vehicle. A sampling path is generated within a preset area along the autonomous vehicle's driving direction, and collision detection is performed on the sampling path based on key obstacles of a first preset type to filter out the first sampling path, including: Sampling points are generated within a preset area along the driving direction of the unmanned vehicle, and a first collision detection is performed on the sampling points based on key obstacles of a first preset type to filter out the first sampling points; multiple sampling paths are generated based on the first sampling points, and a second collision detection is performed on the multiple sampling paths based on key obstacles of a first preset type to filter out the first sampling path; Based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the unmanned vehicle when it travels on the first sampling path, the interaction intent between the unmanned vehicle and the key obstacle of the second preset type is determined. The interaction intent is evaluated, and a driving strategy for the autonomous vehicle is generated based on the evaluation results.

2. The method as described in claim 1, characterized in that, The obstacles include: non-interactive obstacles and key obstacles; the key obstacles include: key static obstacles, abnormal obstacles and key dynamic obstacles; the key static obstacles include: normal static obstacles and active static obstacles.

3. The method as described in claim 1, characterized in that, The preset area includes: longitudinal length and transverse width; Sampling points are generated within a preset area along the driving direction of the autonomous vehicle, including: Based on the longitudinal remaining length of the predicted trajectory of the unmanned vehicle and / or the longitudinal position of the key obstacle, determine the longitudinal sampling interval over the longitudinal length; According to the longitudinal sampling interval, longitudinal sampling points are generated along the longitudinal length in the direction of travel of the unmanned vehicle; Based on the lateral position of the key obstacle and the lateral width of the drivable space in the lateral direction of each longitudinal sampling point, the lateral sampling interval on the corresponding lateral width of each longitudinal sampling point is determined. In the horizontal direction at the location of each vertical sampling point, horizontal sampling points are generated according to the corresponding horizontal sampling interval; The longitudinal length is determined based on the driving speed of the autonomous vehicle; the lateral width is determined based on the lateral drivable boundary of the autonomous vehicle, and satisfies the condition that the lateral width is narrower the closer to the autonomous vehicle.

4. The method as described in claim 1, characterized in that, The first preset type of key obstacles includes abnormal obstacles and key static obstacles; The first collision detection of the sampling points based on key obstacles of a first preset type includes: Determine whether the sampling point overlaps with the first preset type of key obstacle. If there is an overlap, remove the sampling point to obtain a first sampling point that does not overlap with the first preset type of key obstacle. The second collision detection is performed on the multiple sampling paths based on key obstacles of a first preset type to filter out the first sampling path, including: Determine whether the multiple sampling paths intersect with the key obstacles of the first preset type. If they intersect, remove the sampling path to obtain a first sampling path that does not intersect with the key obstacles of the first preset type.

5. The method as described in claim 1, characterized in that, After obtaining the first sampling point, the process also includes: The first sampling points that exceed the steering capability of the unmanned vehicle are excluded from the first sampling points to obtain the first sampling points after further filtering. Multiple sampling paths are generated based on the first sampling point, including: Multiple sampling paths are generated based on the first sampling point after the second filtering.

6. The method as described in claim 1, characterized in that, The second preset type of key obstacles includes key dynamic obstacles; The step of determining the interaction intent between the autonomous vehicle and the second preset type of key obstacle based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the autonomous vehicle while driving on the first sampling path includes: The predicted trajectory of the key dynamic obstacle is obtained, and various accelerations of the key dynamic obstacle as it travels on the predicted trajectory are sampled to obtain the first dynamic intent. For each first sampling path, the various accelerations of the unmanned vehicle traveling on that first sampling path are sampled to obtain the second dynamic intent; By combining the first dynamic intent and the second dynamic intent with the first sampling path, an interaction intent is determined for the unmanned vehicle to avoid interacting with the key dynamic obstacle during its journey along the first sampling path.

7. The method as described in claim 1, characterized in that, The step of evaluating the interaction intent and generating the driving strategy of the autonomous vehicle based on the evaluation result includes: Each interaction intent is evaluated based on preset metrics; wherein the preset metrics include at least one of the following: security, human-likeness, and accessibility; The interaction intentions are evaluated and ranked according to the evaluation results of each interaction intention; Based on the requirement of multiple intents, the sorted interaction intents are classified and coded, and duplicate intents at the end of the sort are removed for each classification code. The driving strategy of the autonomous vehicle is generated based on the remaining interaction intentions after removing duplicate intentions and the evaluation and ranking results of the remaining interaction intentions.

8. The method as described in claim 1, characterized in that, Before determining the interaction intention of the autonomous vehicle with the second preset type of key obstacle when traveling on the first sampling path, based on the first dynamic intention of the key obstacle of the second preset type and the second dynamic intention of the autonomous vehicle, the method further includes: The first sampling path is evaluated and ranked based on at least one of the following: path shape change variance, distance from reference line, distance to active static obstacle, and end point distance, to obtain a preset number of first sampling paths ranked first. Among them, the smaller the variance of the path shape change, the higher the corresponding evaluation value; the smaller the distance from the reference line, the higher the corresponding evaluation value; the larger the distance to the active static obstacle, the higher the corresponding evaluation value; and the smaller the distance to the endpoint, the higher the corresponding evaluation value. The step of determining the interaction intent between the autonomous vehicle and the second preset type of key obstacle based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the autonomous vehicle while driving on the first sampling path includes: Based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the unmanned vehicle when driving on the first sampled path of the first preset number, the interaction intent between the unmanned vehicle and the key obstacle of the second preset type is determined.

9. The method as described in claim 1, characterized in that, The step of identifying the type of a key obstacle upon detection includes: If a critical obstacle is detected, the type of the obstacle is identified based on the reference line projection of the critical obstacle; The method further includes: When the obstacle is an abnormal obstacle, the obstacle is expanded to form an abnormal driving area as the corresponding area of ​​the obstacle; When the obstacle is one that allows normal interaction with the unmanned vehicle, the specified area near the predicted trajectory of the obstacle is obtained as the corresponding area of ​​the obstacle. SL maps are constructed based on the corresponding regions of different types of obstacles.

10. A decision-making device for an unmanned vehicle, characterized in that, include: The obstacle detection module is used to detect obstacles in the direction the autonomous vehicle is traveling. An obstacle recognition module is used to identify the type of a key obstacle when it is detected; wherein the key obstacle includes obstacles that may interact with the unmanned vehicle. A collision testing module is used to generate a sampling path within a preset area along the driving direction of the autonomous vehicle, and to perform collision detection on the sampling path based on key obstacles of a first preset type, thereby selecting a first sampling path, including: Sampling points are generated within a preset area along the driving direction of the unmanned vehicle, and a first collision detection is performed on the sampling points based on key obstacles of a first preset type to filter out the first sampling points; multiple sampling paths are generated based on the first sampling points, and a second collision detection is performed on the multiple sampling paths based on key obstacles of a first preset type to filter out the first sampling path; The interaction intent module is used to determine the interaction intent between the unmanned vehicle and the key obstacle of the second preset type based on the first dynamic intent of the key obstacle of the second preset type and the second dynamic intent of the unmanned vehicle when it travels on the first sampling path. The strategy generation module is used to evaluate the interaction intent and generate the driving strategy of the autonomous vehicle based on the evaluation results.

11. A vehicle mixed-traffic control system, characterized in that, Including the decision-making device for an unmanned vehicle as described in claim 10.

Images