Trajectory generation method, electronic device, and vehicle

Abstract

The application discloses a trajectory generation method, an electronic device and a vehicle, and relates to the technical field of trajectory generation. The method comprises the following steps: executing a first planning strategy in a first scene; executing a second planning strategy in a second scene; and outputting a driving trajectory according to the first planning strategy or the second planning strategy. The application analyzes the type of a scene in real time, executes different path planning strategies according to the type of the scene, and improves the safety and rationality of trajectory planning.

Description

Technical Field

[0001] This application relates to the field of driver assistance technology, and more specifically, to a trajectory generation method, electronic equipment, and vehicle. Background Technology

[0002] Trajectory generation for assisted driving vehicles is one of the core functions of autonomous driving systems, with the goal of generating a safe, efficient, and habit-friendly driving trajectory. However, existing trajectory generation methods suffer from low accuracy in trajectory planning. Summary of the Invention

[0003] This application provides a trajectory generation method, electronic device, and vehicle, which solves the problems of real-time scenario analysis and safe trajectory generation.

[0004] To achieve the above objectives, according to a first aspect of this application, a trajectory generation method is provided, applied to a vehicle, comprising:

[0005] Execute the first planning strategy in the first scenario;

[0006] Execute the second planning strategy in the second scenario;

[0007] Output the driving trajectory based on the first planning strategy or the second planning strategy.

[0008] Optionally, the first scenario and the second scenario include:

[0009] The first scenario includes general complex scenarios and / or moderately complex scenarios;

[0010] The second scenario includes extremely complex scenarios.

[0011] Optionally, the method further includes:

[0012] Scene analysis is performed based on the collected data to obtain scene classification results, wherein the collected data is surrounding environment information, and the scene classification results include the first scene and the second scene.

[0013] Optionally, the step of performing scene analysis based on the collected data to obtain scene classification results includes:

[0014] The collected data is analyzed based on the first scene classification model to obtain the first classification result;

[0015] The collected data is analyzed based on the second scene classification model to obtain the second classification result;

[0016] The scene classification result is obtained based on the first classification result and the second classification result.

[0017] Optionally, when the collected data is multiple, the method further includes:

[0018] Based on the first scene classification model, the second scene classification model, and the collected data, multiple preliminary scene classification results are obtained;

[0019] The multiple preliminary scene classification results are filtered according to preset scene division conditions to obtain the scene classification result.

[0020] Optionally, obtaining the scene classification result based on the first classification result and the second classification result includes:

[0021] The first classification result and the second classification result are weighted respectively;

[0022] The scene classification result is obtained based on the weighted first classification result and the weighted second classification result.

[0023] Optionally, the collected data includes data collected at the previous moment and data collected at the current moment, and the execution of the first planning strategy in the first scenario includes:

[0024] The target is identified and tracked based on the data collected at the previous moment and the data collected at the current moment, so as to generate the target's activity trajectory;

[0025] The driving trajectory is output based on the target's activity trajectory.

[0026] Optionally, the step of identifying and tracking the target based on the data collected at the previous moment and the data collected at the current moment to generate the target's activity trajectory includes:

[0027] Based on the data collected in the previous moment, the target is identified and tracked to generate a first activity trajectory;

[0028] Based on the data collected at the current moment, the target is identified and tracked to generate a second activity trajectory;

[0029] The target's activity trajectory is generated based on the first activity trajectory and the second activity trajectory.

[0030] Optionally, the step of identifying and tracking the target based on the data collected at the previous moment to generate a first activity trajectory includes:

[0031] A virtual scene is constructed based on the data collected in the previous moment;

[0032] The target is identified and tracked based on the virtual scene to generate the first activity trajectory.

[0033] Optionally, the step of constructing a virtual scene from the data collected at the previous moment includes:

[0034] When the data collected in the previous moment is a single data point, the virtual scene is generated based on the data collected in the previous moment.

[0035] When there are multiple data points collected in the previous moment, the virtual scene is generated by splicing the data points collected in the previous moment.

[0036] Optionally, executing the second planning strategy in the second scenario includes:

[0037] Multiple candidate trajectories are generated based on the collected data;

[0038] The driving trajectory is output based on the trajectory curvature and / or trajectory acceleration of the multiple candidate trajectories.

[0039] Optionally, when the vehicle is in the second scenario, the method further includes:

[0040] Upload the collected data and / or the driving trajectory in the second scenario to the cloud server to optimize the scenario analysis model on the cloud server.

[0041] According to a second aspect of this application, an electronic device is provided, comprising:

[0042] Memory, on which computer programs / instructions are stored;

[0043] A processor is configured to execute the computer program / instructions in the memory to implement the steps of the trajectory generation method described above.

[0044] According to a third aspect of this application, a computer-readable storage medium is provided, on which a computer program / instructions are stored, which, when executed by a processor, implement the steps of the trajectory generation method described above.

[0045] According to a fourth aspect of this application, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the steps of the trajectory generation method described above.

[0046] According to a fifth aspect of this application, a vehicle is provided, the vehicle including the electronic equipment as described above.

[0047] This application analyzes the scenario type in real time and executes different path planning strategies according to the scenario type to improve the safety and rationality of trajectory planning.

[0048] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This application provides a flowchart of a trajectory generation method in certain embodiments;

[0051] Figure 2 This application provides a schematic flowchart of a trajectory generation method in certain embodiments;

[0052] Figure 3 This application provides scene classification detectors in certain embodiments;

[0053] Figure 4 This application provides flowcharts for virtual scene construction and trajectory output in certain embodiments;

[0054] Figure 5 An electronic device is provided in certain embodiments of this application. Detailed Implementation

[0055] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Furthermore, it should be understood that the specific embodiments described herein are only for illustration and explanation of this application and are not intended to limit this application.

[0056] According to the background description of this application, path generation methods for assisted driving vehicles refer to the technology of calculating a detailed driving path from the current position to the target position based on environmental perception information, vehicle status, and preset driving targets during the autonomous navigation process. Current vehicle trajectory generation methods include: model predictive control methods, potential field methods, fast stochastic extended tree methods, B-spline curves, and artificial potential field methods. Current path planning methods generally only consider the vehicle's historical trajectory, source domain feature data, and target domain feature data, and generate future trajectories through multilayer perceptrons and temporal fusion Transformers. However, using the vehicle's historical trajectory without considering the motion trajectories of other known or unknown objects can lead to relatively dangerous vehicle planning.

[0057] Therefore, real-time analysis of scenarios and generation of safety trajectories is essential.

[0058] To address the aforementioned problems, this application provides a trajectory generation method that is not only applicable to vehicles but can also be extended to other mobile devices, such as drones, robots, and aircraft, including:

[0059] Execute the first planning strategy in the first scenario;

[0060] Execute the second planning strategy in the second scenario;

[0061] Output the driving trajectory based on either the first or second planning strategy.

[0062] This can be understood as follows: the complexity of a scene can be graded based on factors such as the surrounding environment, including ground morphology, weather conditions, lighting conditions, types and behaviors of moving / static objects, and traffic density, resulting in a first scene and a second scene with different levels of complexity. The planning strategy can generate the final trajectory by combining a series of conditions, such as the surrounding environment information at different times, the status of mobile devices (e.g., vehicles), driving preferences, regulations and standards, with different algorithms (e.g., sampling method, optimization method, machine learning, world model, occupancy grid network method, etc.).

[0063] Specifically, in combination Figure 1 As shown, taking a vehicle as an example, when the vehicle's surrounding environment is determined to be a first scenario (e.g., a scenario with low complexity), a first planning strategy can be executed to output a driving trajectory. When the vehicle's surrounding environment is determined to be a second scenario (e.g., a scenario with extremely high complexity), a second planning strategy can be executed to output a driving trajectory. Real-time analysis of scenario types and execution of different path planning strategies based on the scenario type can effectively improve the safety and rationality of trajectory planning. This enables intelligent driving systems to safely and quickly traverse corresponding scenarios, improving the user experience.

[0064] In some implementations, the first scenario and the second scenario include:

[0065] The first scenario includes general complex scenarios and / or medium complex scenarios;

[0066] The second scenario includes extremely complex scenarios.

[0067] This can be understood as combining Figure 2As shown, scenarios are categorized according to the complexity of the surrounding environment, ranging from low to high complexity into generally complex, moderately complex, and extremely complex scenarios. The first and second scenarios can be understood as grouping scenarios based on their complexity among various classification results. For example, scenarios with relatively low complexity, such as generally complex and moderately complex scenarios, can be classified as the first scenario, while extremely complex scenarios, such as extremely complex scenarios, can be classified as the second scenario. Different algorithms and processing strategies are selected for trajectory planning and output based on the first or second scenario. By classifying scenarios into generally complex, moderately complex, and extremely complex categories, and selecting different processing strategies based on complexity, planning efficiency is optimized.

[0068] In one specific embodiment, general complex scenarios, medium complex scenarios, and extremely complex scenarios can be defined as follows:

[0069] Complex scenarios generally refer to those with few traffic participants, well-defined road conditions, clear rules, and relatively simple driving tasks; these scenarios place lower demands on the decision-making and control capabilities of driver assistance systems. Typical scenarios include highways, urban expressways / elevated roads, and uninterrupted main roads.

[0070] Medium-complexity scenarios refer to situations where the road environment or traffic participants are dynamically changing, requiring the system to handle multi-task coordination, local decision-making, and interaction. These scenarios place high demands on the perception accuracy, decision-making logic, and control response speed of the assisted driving system. Typical scenarios include: urban arterial roads, multi-lane intersections, curves / slopes, and lightly congested road sections.

[0071] Extremely complex scenarios refer to environments with high dynamics, dense traffic participants, ambiguous rules, or sudden disturbances. These scenarios place extremely high demands on the perception, decision-making, and control capabilities of assisted driving systems, requiring them to cope with multi-dimensional uncertainties. Typical scenarios include: heavily congested road sections, multi-lane disordered intersections, construction / accident sections, and urban villages / narrow roads.

[0072] In some implementations, the method further includes:

[0073] Scene analysis is performed based on the collected data to obtain scene classification results. The collected data is information about the surrounding environment, and the scene classification results include the first scene and the second scene.

[0074] Among them, combined Figure 1 , Figure 2 and Figure 3As shown, the collected data can be understood as environmental data acquired by a data acquisition device, such as a camera, LiDAR, or millimeter-wave radar. The collected data can include images and point clouds of the surrounding environment. Scene analysis can be performed by processing the collected data using algorithms such as VLM (Visual Language Model) and cluster analysis to obtain scene classification results.

[0075] Specifically, taking a vehicle as an example, the vehicle periodically collects environmental data about its surroundings during its journey, and labels and stores this data sequentially according to time. Combining the collected environmental data, such as images and / or point clouds, the system analyzes the complexity of the current environment the vehicle is in. Then, based on the corresponding scene classification, an appropriate processing strategy or algorithm is selected to process the collected data to obtain the final driving trajectory. In this way, real-time analysis of scene types, combined with the real-time collected data and the temporal and spatial dimensions, analyzes the relative positions and trajectories of the vehicle and obstacles, improving the safety and rationality of trajectory planning.

[0076] Furthermore, in one specific embodiment, such as Figure 2 As shown, scene analysis can be understood as being accomplished based on the scene complexity detection module. The scene complexity detection module can be trained in a closed loop on the vehicle side based on the collected data and driving trajectory, continuously improving its performance in distinguishing scene complexity; alternatively, the vehicle side can upload the collected data and driving trajectory to a cloud server, where the cloud server will train the scene complexity detector in a closed loop in the background, continuously improving its performance in distinguishing scene complexity. The vehicle side then receives the optimized scene complexity detector and performs scene analysis.

[0077] In some implementations, scene analysis is performed based on the collected data to obtain scene classification results, including:

[0078] The collected data is analyzed based on the first scene classification model to obtain the first classification result;

[0079] The collected data is analyzed based on the second scene classification model to obtain the second classification result;

[0080] Based on the first and second classification results, the scene classification result is obtained.

[0081] Among them, combined Figure 1 , Figure 2 and Figure 3As shown, scene analysis can be understood as analyzing collected data based on different algorithm models to obtain classification results. The first scene classification model and the second scene classification model refer to different algorithm models. For example, the first scene classification model can be cluster analysis, and the second scene classification model can be a visual language model. The first classification result and the second classification result can be one of the scene classifications such as general complex scene, medium complex scene, and extremely complex scene. Alternatively, the first classification result and the second classification result can be understood as candidate results, and one of them needs to be selected as the scene classification result.

[0082] Specifically, the first scene classification model is used to analyze and process the collected data to obtain a first classification result. Then, a second scene classification model is used to analyze and process the collected data to obtain a second classification result. Finally, from the first and second classification results, based on preset screening criteria such as weighted scoring, reliability indicators, and confidence levels, one is selected as the final scene classification result. Combining two or more algorithms can compensate for the shortcomings of a single algorithm, improve the accuracy and robustness of the results, and is suitable for handling complex problems and making better decisions.

[0083] Furthermore, based on different algorithm models, multiple models such as the first scene classification model, the second scene classification model, and the third scene classification model can be obtained. Then, based on these multiple models, multiple results such as the first classification result, the second classification result, and the third classification result can be obtained respectively. Among the multiple results, one is selected as the final scene classification result according to preset filtering conditions.

[0084] In some implementations, a scene classification result is obtained based on the first classification result and the second classification result, including:

[0085] The first classification result and the second classification result are weighted separately;

[0086] The scene classification result is obtained based on the weighted first classification result and the weighted second classification result.

[0087] This can be understood as follows: different values ​​are first set for different classification results, then each classification result is multiplied by a weighted coefficient to obtain a score value, and finally the scores of each classification result are compared to obtain the final scene classification result.

[0088] Specifically, in combination Figure 1 , Figure 2 and Figure 3As shown, in one specific embodiment, the scene complexity detection module for scene analysis takes an image as input. The input image needs to be fed into a Visual Language Model (VLM) trained based on assisted driving data for three comparisons to distinguish which of the three complexity levels the scene belongs to, and outputs the corresponding second classification result k1*y1, where y1 is the classification result. In addition, before classification, an unsupervised clusterer is trained based on existing database images. This clusterer can classify the input image into one of the three categories and will also output a first classification result k2*y2. If k1*y1 ≥ k2*y2, the image will be classified according to the result of y1; otherwise, it will be classified according to the result of y2. Here, k1 and k2 are the weights of the classifier, and y takes values ​​of 1 (generally complex scene), 2 (medium complex scene), and 3 (extremely complex scene).

[0089] In some implementations, when multiple data points are collected, the method further includes:

[0090] Based on the first scene classification model, the second scene classification model, and multiple collected data, several preliminary scene classification results were obtained;

[0091] The preliminary scene classification results are processed based on the preset scene segmentation conditions to obtain the scene classification results.

[0092] This can be understood as the collection of one or more data points at any given time. For example, if the collected data consists of images, the images could be images containing only the front view of the vehicle, or images from all the vehicle's cameras, i.e., six images in total: front view, rear view, two side front views, and two side rear views. The preliminary scene classification result can be one of the following: a generally complex scene, a moderately complex scene, or an extremely complex scene. The final scene classification result needs to be selected from multiple preliminary scene classification results based on certain conditions.

[0093] Specifically, in combination Figure 1 , Figure 2 and Figure 3 As shown, if multiple data points are collected simultaneously, each data point needs to be analyzed and compared using the first scene classification model and the second scene classification model to obtain multiple preliminary scene classification results. Then, from these preliminary scene classification results, the final scene classification result can be obtained by filtering according to preset conditions. In a specific embodiment, assuming that a single input driving scene consists of images from six different cameras, if one image is classified as an extremely complex scene by the pre-trained algorithm model, it is classified as an extremely complex scene; only if all six driving scenes are classified as generally complex scenes is it classified as a generally complex scene; assuming c iThe scene complexity is represented by y, where i ranges from 1 to 6, representing different viewpoints within a scene, with values ​​of 1 (generally complex scene), 2 (mediumly complex scene), and 3 (extremely complex scene). The final scene classification result is represented by y. In one embodiment, the preset conditions are as follows:

[0094] If y = 3, and c exists... i =3 (at least one camera detected an extremely complex scene);

[0095] y = 1, if all c i =1 (All cameras detected a generally complex scene);

[0096] y = 2, other cases (there are moderately complex or generally complex scenarios, but no extremely complex scenarios).

[0097] In some implementations, the collected data includes data collected at the previous moment and data collected at the current moment. Executing a first planning strategy in the first scenario includes:

[0098] The target is identified and tracked based on data collected in the previous moment and data collected in the current moment, so as to generate the target's activity trajectory;

[0099] Output the driving trajectory based on the target's activity trajectory.

[0100] This can be understood as the data being collected periodically by the data acquisition device and then tagged and stored sequentially according to time. The current moment can be the latest point in time or state being processed, while the previous moment can be one or more points in time or states immediately preceding the current moment; the target can be the vehicle itself, or other vehicles, pedestrians, etc. in the surrounding environment.

[0101] Specifically, in combination Figure 2 and Figure 4As shown, after obtaining the scene classification result of the current environment, if the scene classification result is a general complex scene or a medium complex scene in the first scene, then the first planning strategy is used to predict the driving trajectory. First, based on the data collected at the previous time t0 and the data collected at the current time t1, targets such as the vehicle, other vehicles, and pedestrians are identified and tracked. Then, combined with high-precision maps and real-time environmental information, R activity trajectories of the vehicle, other vehicles, and pedestrians at the previous time t0 and the current time t1 are obtained. Then, based on the R activity trajectories, the activity trajectories of the vehicle, other vehicles, and pedestrians at the next time are predicted. The potential conflicts (such as collision risks) between the vehicle's activity trajectory at the next time and the activity trajectories of other vehicles and pedestrians at the next time are analyzed, and the rationality of different activity trajectory combinations is evaluated. Next, based on goals such as safety, efficiency, and comfort, candidate activity trajectories (such as lane keeping, lane changing, etc.) are generated, and the cost of the activity trajectory is optimized using algorithms (such as A*, dynamic programming). Finally, considering the activity trajectory risk, regulatory constraints, and dynamic changes, the optimal activity trajectory (such as the least conflict and the lowest energy consumption) is selected as the output driving trajectory. In this way, by combining data collected at multiple times, such as images and point cloud data, the trajectory is optimized frame by frame to ensure the safety and continuity of the driving trajectory planning.

[0102] In some implementations, target identification and tracking are performed based on data collected at the previous moment and data collected at the current moment to generate the target's activity trajectory, including:

[0103] Based on the data collected in the previous moment, the target is identified and tracked to generate the first activity trajectory;

[0104] Based on the data collected at the current moment, the target is identified and tracked to generate a second activity trajectory;

[0105] Based on the first and second activity trajectories, the target's activity trajectory is generated.

[0106] Specifically, in combination Figure 2 and Figure 4 As shown, based on the data collected at the previous time t0, targets such as the vehicle, other vehicles, and pedestrians are identified and tracked, resulting in R first activity trajectories. Based on the data collected at the current time t1, targets such as the vehicle, other vehicles, and pedestrians are also identified and tracked, resulting in R second activity trajectories. By comparing the R first activity trajectories at time t0 with the R second activity trajectories at time t1, a path similarity method is used, such as calculating the Euclidean distance between each pair of targets, to obtain the activity trajectory with the least change, which is then output as the activity trajectory for each target.

[0107] In some implementations, target identification and tracking are performed based on data collected at the previous moment to generate a first activity trajectory, including:

[0108] Construct a virtual scene based on data collected in the previous moment;

[0109] Target recognition and tracking are performed based on the virtual scene to generate the first activity trajectory.

[0110] This can be understood as follows: virtual scene construction can be based on collected data, such as images and / or point clouds, and generate virtual scenes with depth information through algorithms, such as world models, occupancy grid network methods, etc.

[0111] Specifically, in combination Figure 2 and Figure 4 As shown, the images and / or point clouds acquired at the previous time t0 are used to generate a virtual scene with depth information through algorithms such as world model and occupancy grid network method. Then, potential targets are identified based on the scene with depth information and tracked by the tracking model to generate the first activity trajectory.

[0112] Furthermore, a virtual scene is constructed based on the data collected in the previous moment, including:

[0113] If the data collected at the current moment is considered as one time, a virtual scene is generated based on the data collected at the previous moment.

[0114] If multiple data points are collected at a given moment, a virtual scene is generated by splicing together the data collected at the previous moment.

[0115] This can be understood as the algorithm for constructing the virtual scene differing depending on the amount of data collected at the previous time step t0. For example, if only one piece of data was collected at the previous time step t0, a virtual scene with depth information is generated using algorithms such as the world model; if multiple pieces of data were collected at the previous time step t0, the data is processed and the scene is rasterized using algorithms such as the occupation grid network method.

[0116] Specifically, in combination Figure 2 and Figure 4 As shown, in a specific embodiment, when the classification result is a general complex scene and a medium complex scene in the first scene, the following steps are performed: Figure 4The operation flow is shown below. If the input image is a forward-looking monocular image, then this image generates a corresponding scene through a world model, and the world model provides the scene's depth. In this single forward-looking image, the target recognition model identifies potential targets and marks them with IDs through a tracking model. The tracking algorithm maintains target recognition in the input image at time t1. In the image at time t0, since the world model has already generated a scene with depth information, the algorithm will randomly generate R paths based on the target location at t0 and the drivable space. If the input data at time t0 also contains the point cloud at that time, then the world model, when generating the scene, possesses the true depth information at that time. Based on the input image or point cloud at time t1, after obtaining the target location at time t1, R random paths are output in the drivable area at time t1 through a pre-trained neural network model. Comparing the R paths in the image at time t0 with the R paths at time t1, the path with the least change is obtained through path similarity methods, such as calculating the Euclidean distance between each pair of distances, and is used as the actual motion path for each target. Next, based on the actual motion paths of each target, the behavior of other objects is predicted. Potential conflicts (such as collision risks) between the vehicle's path and the paths of other vehicles, pedestrians, and other targets are analyzed, and the rationality of different path combinations is evaluated. Based on objectives such as safety, efficiency, and comfort, candidate paths (such as lane keeping and lane changing) are generated, and algorithms (such as A* and dynamic programming) are used to optimize path costs. Finally, considering path risks, regulatory constraints, and dynamic changes, the optimal path (such as the one with the fewest conflicts and the lowest energy consumption) is selected as the output driving trajectory.

[0117] When there are 6 images input at time t0, the occupancy grid network method is used to process the images and rasterize the scene. Key static and dynamic objects are identified in each image, and key dynamic objects are marked. In the 6 images at time t1, objects from the previous time step are tracked. Similarly, drivable areas are divided in the raster maps at times t0 and t1. A pre-trained neural network model outputs R random paths, and distances are calculated. Finally, the driving trajectory is output. If there are 6 images input at time t0 and point cloud information is available, the above process is similar and will not be elaborated further.

[0118] In some implementations, executing a second planning strategy in a second scenario includes:

[0119] Multiple candidate trajectories are generated based on the collected data;

[0120] The driving trajectory is output based on the trajectory curvature and / or trajectory acceleration of multiple candidate trajectories.

[0121] In this context, the candidate trajectory can be generated based on a certain algorithm (such as hierarchical planning algorithm, neural network, machine learning, etc.); the trajectory curvature can be a geometric parameter describing the degree of curvature of the trajectory, used to quantify the degree of turning of the trajectory at a certain point; the trajectory acceleration can be a physical quantity describing the change of velocity of an object (such as a vehicle or robot) when it moves along the planned trajectory, and can include lateral acceleration, longitudinal acceleration, jerk, etc.

[0122] Specifically, when the current environment is determined to be the second scenario, i.e., an extremely complex environment, to ensure the safety of oneself and surrounding pedestrians, multiple candidate trajectories will first be generated based on the collected data. Then, based on the trajectory curvature of each candidate trajectory (e.g., constrained curvature ≤ 0.1m),... -1 ) and / or trajectory acceleration (e.g., longitudinal acceleration absolute value ≤ 0.5 m / s²) 2 Jerk (jerk) ≤ 1.5 m / s² 3 The system filters based on conditions such as [list of conditions] to obtain the final driving trajectory. This allows the vehicle to avoid pedestrians, obstacles, etc., or to output a driving path to follow the vehicle in front, thereby controlling the vehicle to travel at a low speed. In a specific embodiment, the second planning strategy is as follows:

[0123] Step 1: Integrate LiDAR, camera, and high-precision map data to construct an environmental model that includes static / dynamic obstacles, lane topology, and speed limits. Predict the movement range of obstacles using a reachable set and set longitudinal safety distance thresholds and lateral lane centering constraints.

[0124] The second step involves using the A* algorithm to generate candidate paths based on cumulative cost (including distance, squared curvature, and acceleration penalty terms) and a heuristic function (Manhattan distance), prioritizing lane-in-lane paths with low curvature and low acceleration. Then, a smooth trajectory is generated using Nth-order Bézier curves or Mth-order polynomial interpolation to determine the constraint curvature, absolute value of longitudinal acceleration, and Jerk (jerk) value, ensuring no lane changes or abrupt changes.

[0125] Step 3: Based on the Collision Cone theory, the intersection of the path and the reachable set of obstacles is detected in real time. If a risk exists, a conservative deceleration is triggered (acceleration ≤ -3m / s²). 2 ).

[0126] The fourth step, based on the above rules, is to obtain the mapping from the sensing sensor to the output safe and smooth trajectory, that is, the optimal path generation and selection.

[0127] In some implementations, when the vehicle is in the second scenario, the following is also included:

[0128] Upload the collected data and / or driving trajectory in the second scenario to the cloud server to optimize the scenario analysis model on the cloud server.

[0129] Specifically, to continuously improve the performance and accuracy of scene analysis, when the vehicle is in the second scenario, the vehicle-side uploads (e.g., periodically) collected data and / or driving trajectories (such as planned trajectories, actual driving trajectories, etc.) to the cloud server. The cloud server receives the collected data and driving trajectories to continuously train and optimize the scene analysis model. The vehicle-side also receives the optimized scene analysis model to perform scene analysis and obtain scene classification results. In this way, the vehicle-side uploads scene data to the cloud to optimize complex scenarios and receives the optimized scene analysis model, improving the overall intelligence level of the system. Meanwhile, the local SoC (System-on-a-Chip) is responsible for real-time path planning, while the cloud is only used for uploading and optimizing scene data, without relying on real-time cloud computing, reducing communication latency and improving system robustness.

[0130] In one specific embodiment, taking a vehicle as an example, for a scene determined to be extremely complex, multiple cameras on the vehicle will capture surround-view and panoramic images and record point cloud data. After the vehicle passes through the scene using a conservative strategy, it will record the current travel path. Then, the domain controller will transmit the path data, images, and point cloud data back to the cloud server via the vehicle's upload link. Here, "conservative strategy" means that the vehicle will not use lane changes, rapid acceleration, or rapid deceleration to pass through the scene, but will instead use a low-speed or following approach. The cloud server will conduct closed-loop training of the scene complexity detector in the background, continuously improving its performance in distinguishing scene complexity.

[0131] To more clearly illustrate the implementation methods of this application, a vehicle is used as an example; please refer to the following documents as well. Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5 Please refer to the following example for further explanation.

[0132] This invention proposes a method for generating vehicle trajectories by combining cloud and local methods, such as... Figure 2 As shown, t0, t1...tN represent different frames within a time series, with corresponding images for each time series. If the vehicle is equipped with LiDAR or millimeter-wave radar, point clouds may also be present in the corresponding time series. Note that the image at time t0 can be a single image containing only the vehicle's front view, or it can be a combination of images from all the vehicle's cameras, i.e., six images in total: front view, rear view, two side front views, and two side rear views. The image at time t1 is the image at the next time step after t0, and the point cloud at time t1 is the point cloud at the next time step after t0, and so on.

[0133] First, image t0 passes through a scene complexity detection module, which then categorizes the scene into three levels of complexity: moderate, medium, and extremely complex. For moderate and medium-complexity scenes, the local SoC (domain controller) calculates and constructs a virtual scene, creating multiple shadows within it. Based on the input endpoint at that moment, it draws the trajectory points within the scene; here, a shadow refers to the trajectory line to be generated. Then, at time t1, the image is input into a multi-frame graph joint detection module, which compares multiple paths and ultimately selects the optimal path for output. For scenes determined to be extremely complex, multiple cameras on the vehicle capture surround-view and panoramic images and record point cloud data. Once the vehicle has passed through the scene using a conservative strategy, it records the current travel path. The domain controller then transmits the path data, image, and point cloud data back to the cloud server via the vehicle's upload link. Here, "conservative strategy" means that the vehicle will not use lane changes, rapid acceleration, or rapid deceleration to pass through the scene, but rather adopt a low-speed or following approach. The cloud server will continuously train the scene complexity detector in the background, constantly improving its performance in distinguishing scene complexity.

[0134] Determining scene complexity, such as Figure 3 As shown, this classifier only requires images as input, and there are two cases: First, if the input is only data from the forward-facing camera, the input image needs to be fed into a Visual Language Model (VLM) trained on intelligent driving data for three comparisons to distinguish which of the three complexity levels the scene belongs to, and output the corresponding result k1*y1, where y1 is the classification result. Second, before classification, an unsupervised clusterer is trained based on existing database images. This clusterer can classify the input image into one of the three categories, and it will also output a result k2*y2. If k1*y1 ≥ k2*y2, the image will be classified according to the result of y1; otherwise, it will be classified according to the result of y2. Here, k1 and k2 are the weights of the classifier.

[0135] Suppose that a single input driving scene consists of images from 6 different cameras. If one image is classified as an extremely complex scene by the pre-trained visual language model, then the scene is classified as an extremely complex scene. Only if all 6 driving scenes are classified as moderately complex scenes will the scene be classified as a moderately complex scene. Let c_i represent the scene complexity, where i ranges from 1 to 6, representing different viewpoints of an image in a scene, with values ​​of 0 (moderately complex scene), 1 (mediumly complex scene), and 2 (extremely complex scene). The final scene classification result is represented by y, as shown in the formula below.

[0136] If y = 3, and c exists... i =3 (at least one camera detected an extremely complex scene);

[0137] y = 1, if all c i =1 (All cameras detected a generally complex scene);

[0138] y = 2, other cases (there are moderately complex or generally complex scenarios, but no extremely complex scenarios).

[0139] In scenarios where the classification results are of general and moderate complexity, perform the following: Figure 4 The operation flow is shown below. If the input image is a forward-looking monocular image, then this image generates a corresponding scene through a world model, and the world model provides the scene's depth. In this single forward-looking image, the target recognition model identifies potential targets and marks them with IDs through a tracking model. The tracking algorithm maintains target recognition in the input image at time t1. In the image at time t0, since the world model has already generated a scene with depth information, the algorithm randomly generates R paths based on the target location at t0 and the drivable space. If the input data at time t0 also contains the point cloud at that time, then the world model generates the scene with the true depth information at that time. Based on the input image or point cloud at time t1, after obtaining the target location at time t1, R random paths are output in the drivable area at time t1 through a pre-trained neural network model. Comparing the R paths in the image at time t0 with the R paths at time t1, the path with the least change is obtained through path similarity methods, such as calculating the Euclidean distance between each pair of distances, and finally the optimal path is output.

[0140] The process for outputting the optimal path is as follows:

[0141] 1. Collect the R paths of the vehicle status and dynamic objects (vehicles, pedestrians, etc.) at times T0 and T1, and combine them with high-precision maps and real-time environmental information.

[0142] 2. Predict the behavior of other objects, analyze potential conflicts between paths (such as collision risks), and evaluate the rationality of different path combinations.

[0143] 3. Vehicle Path Generation: Based on goals such as safety, efficiency, and comfort, candidate paths (such as lane keeping, lane changing, etc.) are generated, and the path cost is optimized using algorithms (such as A*, dynamic programming).

[0144] 4. Decision Selection: Considering path risks, regulatory constraints, and dynamic changes, select the optimal path (e.g., the one with the fewest conflicts and the lowest energy consumption) as the output.

[0145] When there are 6 images input at time t0, the occupancy grid network method is used to process the images and rasterize the scene. Key static and dynamic objects are identified in each image, and key dynamic objects are marked. In the 6 images at time t1, objects from the previous time step are tracked. Similarly, drivable areas are divided in the raster maps at times t0 and t1. A pre-trained neural network model outputs R random paths, distances are calculated, and finally, the optimal path is output. If there are 6 images input at time t0 and point cloud information is available, the above process is similar and will not be elaborated further.

[0146] It is worth noting that the cloud server collects the decision results of all vehicles in different scenarios and at different times, and optimizes the decisions of different vehicles in the same scenario. Before the vehicle arrives at the scenario, the data will be pushed to the vehicle in a timely manner. For example, before a vehicle arrives in a new city, the scenario passage strategy has been optimized in advance for the complex road sections of that city.

[0147] Figure 5 The controller of the present invention is shown, which has a memory for storing program code, a separately running processor, and real-time access to camera sensors, map information, vehicle speed information, etc. via a communication interface.

[0148] This application also provides an electronic device, including:

[0149] Memory, on which computer programs / instructions are stored;

[0150] A processor is configured to execute the computer program / instructions in the memory to implement the steps of the parking space determination method described above.

[0151] This application also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the above-described parking space determination method.

[0152] This application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described parking space determination method.

[0153] This application also provides a vehicle, characterized in that the vehicle includes the electronic equipment described above, or a parking space determination system.

[0154] In this specification, the terms "specifically," "furthermore," "particularly," "can be understood," etc., refer to specific features, structures, materials, or characteristics described in connection with embodiments or examples that are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0155] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0156] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A trajectory generation method, characterized in that, Applied to vehicles, including: Execute the first planning strategy in the first scenario; Execute the second planning strategy in the second scenario; Output the driving trajectory based on the first planning strategy or the second planning strategy.

2. The method according to claim 1, characterized in that, The first scenario and the second scenario include: The first scenario includes general complex scenarios and / or moderately complex scenarios; The second scenario includes extremely complex scenarios.

3. The method according to claim 1 or 2, characterized in that, The method further includes: Scene analysis is performed based on the collected data to obtain scene classification results, wherein the collected data is surrounding environment information, and the scene classification results include the first scene and the second scene.

4. The method according to claim 3, characterized in that, The process of analyzing the collected data to obtain scene classification results includes: The collected data is analyzed based on the first scene classification model to obtain the first classification result; The collected data is analyzed based on the second scene classification model to obtain the second classification result; The scene classification result is obtained based on the first classification result and the second classification result.

5. The method according to claim 4, characterized in that, When the collected data is multiple, the method further includes: Based on the first scene classification model, the second scene classification model, and the collected data, multiple preliminary scene classification results are obtained; The multiple preliminary scene classification results are filtered according to preset scene division conditions to obtain the scene classification result.

6. The method according to claim 4, characterized in that, The step of obtaining the scene classification result based on the first classification result and the second classification result includes: The first classification result and the second classification result are weighted respectively; The scene classification result is obtained based on the weighted first classification result and the weighted second classification result.

7. The method according to any one of claims 1-3, characterized in that, The collected data includes data collected at the previous moment and data collected at the current moment. The execution of the first planning strategy in the first scenario includes: The target is identified and tracked based on the data collected at the previous moment and the data collected at the current moment, so as to generate the target's activity trajectory; The driving trajectory is output based on the target's activity trajectory.

8. The method according to claim 7, characterized in that, The step of identifying and tracking the target based on the data collected at the previous moment and the data collected at the current moment, in order to generate the target's activity trajectory, includes: Based on the data collected in the previous moment, the target is identified and tracked to generate a first activity trajectory; Based on the data collected at the current moment, the target is identified and tracked to generate a second activity trajectory; The target's activity trajectory is generated based on the first activity trajectory and the second activity trajectory.

9. The method according to claim 8, characterized in that, The step of identifying and tracking the target based on the data collected at the previous moment, and generating a first activity trajectory, includes: A virtual scene is constructed based on the data collected in the previous moment; The target is identified and tracked based on the virtual scene to generate the first activity trajectory.

10. The method according to claim 9, characterized in that, The construction of a virtual scene from the data collected in the previous moment includes: When the data collected in the previous moment is a single data point, the virtual scene is generated based on the data collected in the previous moment. When there are multiple data points collected in the previous moment, the virtual scene is generated by splicing the data points collected in the previous moment.

11. The method according to claim 1, characterized in that, The execution of the second planning strategy in the second scenario includes: Multiple candidate trajectories are generated based on the collected data; The driving trajectory is output based on the trajectory curvature and / or trajectory acceleration of the multiple candidate trajectories.

12. The method according to any one of claims 1-11, characterized in that, When the vehicle is in the second scenario, the following is also included: Upload the collected data and / or the driving trajectory in the second scenario to the cloud server to optimize the scenario analysis model on the cloud server.

13. An electronic device, characterized in that, include: Memory, on which computer programs / instructions are stored; A processor for executing the computer program / instructions in the memory to implement the steps of the trajectory generation method according to claims 1-12.

14. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the trajectory generation method as described in claims 1-12.

15. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the trajectory generation method as described in claims 1-12.

16. A vehicle, characterized in that, The vehicle includes the electronic equipment as described in claim 13.

Images