Driving trajectory planning method and device, equipment and medium

Abstract

The application relates to the technical field of trajectory planning, and discloses a driving trajectory planning method, device, equipment and medium. The method comprises the following steps: acquiring a first environment feature of environment information containing a current driving environment; inserting a plurality of anchor point features into the first environment feature to obtain a second environment feature, wherein the anchor point features are used for representing reference path points corresponding to sample driving trajectories; performing forward diffusion and reverse denoising on the second environment feature to reconstruct corresponding driving trajectories based on the anchor point features, so as to obtain corresponding candidate trajectory features; performing confidence evaluation on the candidate trajectory features, and selecting a target trajectory feature from the candidate trajectory features according to a confidence evaluation result; and performing path planning control on a vehicle based on the target trajectory feature. The application can realize multi-modal trajectory generation in a complex traffic scene, and meet the requirements of multi-modal behavior expression and real-time performance.

Description

Technical Field

[0001] This application relates to the field of trajectory planning technology, and in particular to a driving trajectory planning method, device, equipment and medium. Background Technology

[0002] Currently, end-to-end autonomous driving systems are increasingly employing deep learning methods, integrating perception, prediction, and decision-making / planning into a unified network to improve overall performance. However, existing end-to-end driving trajectory planning methods often adopt a modular design (perception → prediction → planning), which, while mature, suffers from poor generalization ability in complex traffic scenarios. Furthermore, while existing end-to-end driving trajectory planning methods attempt to learn trajectory planning strategies directly from sensor inputs, they often use single-modal regression models or fixed trajectory vocabulary sampling, making it difficult to simultaneously meet the requirements of multimodal behavior representation and real-time performance. Summary of the Invention

[0003] The purpose of this application is to provide a driving trajectory planning method, device, equipment, and medium that can generate multimodal trajectories in complex traffic scenarios, meeting the requirements for multimodal behavior expression and real-time performance.

[0004] This application provides a driving trajectory planning method, including: Obtain the first environmental feature containing environmental information about the current driving environment; Multiple anchor point features are inserted into the first environmental feature to obtain the second environmental feature; the anchor point features are used to characterize the reference path points corresponding to the corresponding sample driving trajectory. The second environmental feature is subjected to forward diffusion and reverse denoising to reconstruct the corresponding driving trajectory based on the anchor point feature, thereby obtaining the corresponding candidate trajectory feature; The confidence level of the candidate trajectory features is evaluated, and the target trajectory feature is selected from the candidate trajectory features based on the confidence level evaluation results. Based on the target trajectory characteristics, the vehicle is subjected to path planning and control.

[0005] In some embodiments, prior to obtaining the first environmental feature containing environmental information about the current driving environment, the method further includes: Acquire environmental sensing features that collect environmental information about the driving environment; The environmental sensing features are encoded to obtain the first environmental feature.

[0006] In some embodiments, the method for generating the anchor point feature includes: Multiple sample path point features are obtained; the sample path point features are composed of multiple path point features and are used to characterize the trajectory information of the sample driving trajectory. Clustering is performed on the path point features in the sample path point features to obtain several feature clusters; Based on the number of path point features in the feature cluster, several path point features are selected as cluster centers to obtain the anchor point features.

[0007] In some embodiments, the forward diffusion and reverse denoising of the second environmental feature includes: Gaussian noise is added to the anchor point features to obtain the initial noisy features; The initial noisy features are forward diffused to obtain noisy state features; The noisy state features are denoised in reverse to obtain the candidate trajectory features.

[0008] In some embodiments, the noisy state feature is a noisy feature prior to convergence to an isotropic standard Gaussian noise feature.

[0009] In some embodiments, prior to the confidence evaluation of the candidate trajectory features, the method further includes: Acquire vehicle status features and map information features; The first environmental feature is used to extract road condition features; Based on the road condition features, vehicle status features, and map information features, the candidate trajectory features are cascaded for denoising to obtain denoised candidate trajectory features.

[0010] In some embodiments, the confidence evaluation of the candidate trajectory features includes: Based on the smoothness index, physical feasibility index, and trajectory morphology stability index of the candidate trajectory features, the consistency confidence score of the candidate trajectory features is calculated. Based on the first environmental characteristic, calculate the external constraint compliance score; Based on the consistency confidence score and the external constraint compliance score, a comprehensive confidence evaluation score for the candidate trajectory features is generated.

[0011] This application embodiment also includes a driving trajectory planning device, comprising: The first module is used to acquire a first environmental feature containing environmental information about the current driving environment; The second module is used to insert multiple anchor point features into the first environmental features to obtain the second environmental features; the anchor point features are used to characterize the reference path points corresponding to the corresponding sample driving trajectory. The third module is used to perform forward diffusion and reverse denoising on the second environmental features in order to reconstruct the corresponding driving trajectory based on the anchor point features and obtain the corresponding candidate trajectory features. The fourth module is used to evaluate the confidence level of the candidate trajectory features and select the target trajectory feature from the candidate trajectory features based on the confidence level evaluation results. The fifth module is used to perform path planning and control of the vehicle based on the target trajectory characteristics.

[0012] This application also includes an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the driving trajectory planning method described above.

[0013] This application embodiment also includes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described driving trajectory planning method.

[0014] The beneficial effects of this application are as follows: By inserting anchor point features into the current driving environment features, these anchor point features are used to characterize the reference path points corresponding to the sample driving trajectory. Combined with a diffusion denoising process, diverse driving trajectories are reconstructed, and then the target trajectory is selected through confidence evaluation, thereby enabling vehicle path planning and control. Thus, by introducing anchor point features and combining forward diffusion and backward denoising mechanisms to reconstruct the driving trajectory, the anchor point features, as abstractions of key points in the sample driving trajectory, provide powerful guidance for trajectory reconstruction. Based on limited anchor point information, diverse candidate trajectories that conform to scene constraints can be generated, thereby avoiding the limitations of single regression models or fixed trajectory vocabulary sampling. In the process of gradually recovering the trajectory from noise through a diffusion model, the multimodal characteristics of driving behavior can be better captured, generating more natural trajectories that are more in line with actual driving habits. At the same time, since forward diffusion and backward denoising operations are performed starting from anchor point features, the generation speed of candidate trajectory features is greatly improved, enhancing the real-time performance of driving trajectory planning. Attached Figure Description

[0015] Figure 1 This is an application environment diagram of the driving trajectory planning method provided in the embodiments of this application.

[0016] Figure 2 This is a flowchart of the driving trajectory planning method provided in the embodiments of this application.

[0017] Figure 3 This is a flowchart of the anchor point feature generation method provided in the embodiments of this application.

[0018] Figure 4 This is a flowchart of a method for forward diffusion and reverse denoising of a second environmental feature provided in an embodiment of this application.

[0019] Figure 5This is a flowchart of a method for evaluating the confidence level of candidate trajectory features provided in an embodiment of this application.

[0020] Figure 6 This is a schematic diagram of the driving trajectory planning device provided in the embodiments of this application.

[0021] Figure 7 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and drawings are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application. Furthermore, the information, data, and signals involved in the embodiments of this application are all authorized by relevant parties or have been fully authorized by all parties, and the collection, use, and processing of related data comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0025] In traditional end-to-end driving trajectory planning methods, the modular design architecture leads to a disconnect in information transmission between the perception, prediction, and decision-making planning modules, resulting in a significant reduction in the system's generalization ability in complex traffic scenarios. Specifically, the mapping relationship between environmental features and trajectory planning strategies is not adequately modeled, causing insufficient adaptability of the generated trajectory to dynamic traffic participants. Furthermore, methods that directly learn trajectory planning strategies from sensor inputs are constrained by single-modality regression model outputs or fixed trajectory vocabulary sampling, making it difficult to effectively represent the multimodal characteristics of driving behavior and meet real-time performance requirements. Consequently, under unreasonable computational resource allocation mechanisms, the trajectory planning process generates redundant computational burdens, leading to increased system response latency and decreased trajectory continuity. For example, in multi-directional intersections in cities, when pedestrians cross, non-motorized vehicles cut in, and multi-vehicle interactions occur, the modular approach suffers from a lack of coordination between the static target detection results output by the perception module and the dynamic trajectory deduction by the prediction module. This leads to frequent discontinuities or conflicts in the trajectories generated by the planning module. Specifically, when a vehicle approaches an intersection, the failure to integrate pedestrian intention prediction information with environmental features results in a trajectory planning state of sudden braking or unexpected deceleration. The single-mode regression model only outputs a single deterministic trajectory, which cannot cover multiple reasonable driving behavior patterns such as yielding and accelerating through. The fixed trajectory vocabulary sampling strategy, due to the limitations of the predefined trajectory set, requires a large amount of redundant calculation when generating candidate trajectories, causing the planning cycle to exceed the real-time control requirements, thereby affecting traffic flow and increasing potential safety risks.

[0026] If the aforementioned technical issues are not resolved, the reliability of the driving trajectory planning system in complex scenarios cannot be guaranteed. Uncertainty in the trajectory generation process may lead to mismatch in control commands, which in turn may cause the vehicle to lose control. In addition, insufficient real-time performance will make it difficult for the system to adapt to highly dynamic traffic environments, resulting in accumulated delays in the planning-control link, which will reduce the overall safety level of the autonomous driving system and user acceptance.

[0027] Based on this, embodiments of this application provide a driving trajectory planning method, apparatus, device, and medium. By inserting anchor point features into the current driving environment features to characterize sample trajectory points, and combining a diffusion denoising process to reconstruct diverse driving trajectories, and then optimizing and selecting target trajectories through confidence evaluation, the method solves the problems of information fragmentation caused by modular architecture and the inability of single-mode output to cover multimodal behavior. It can realize multimodal trajectory generation in complex traffic scenarios and meet the requirements of multimodal behavior expression and real-time performance.

[0028] Figure 1 This diagram illustrates the application environment of the driving trajectory planning method provided in the embodiments of this application. (See attached diagram.) Figure 1This method is applied to a driving trajectory planning system. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be at least one of a mobile phone, tablet, laptop, or in-vehicle terminal. The server 120 can be a standalone server or a server cluster consisting of several servers. The terminal 110 sends a first environmental feature containing environmental information of the current driving environment to the server 120. The server 120 acquires the first environmental feature containing the environmental information of the current driving environment, inserts multiple anchor features into the first environmental feature to obtain a second environmental feature, performs forward diffusion and backward denoising on the second environmental feature to reconstruct the corresponding driving trajectory based on the anchor features, obtains corresponding candidate trajectory features, evaluates the confidence of the candidate trajectory features, selects a target trajectory feature from the candidate trajectory features based on the confidence evaluation results, and performs path planning control of the vehicle based on the target trajectory feature. The anchor features are used to characterize the reference path points corresponding to the sample driving trajectory.

[0029] It should be understood that Figure 1 The application scenarios shown are merely examples. In practical applications, the driving trajectory planning method provided in this application embodiment can also be applied to other scenarios. For example, the above-described driving trajectory planning method can be directly applied to terminal 110. Terminal 110 is used to obtain a first environmental feature containing environmental information of the current driving environment, insert multiple anchor point features into the first environmental feature to obtain a second environmental feature, perform forward diffusion and reverse denoising on the second environmental feature to reconstruct the corresponding driving trajectory based on the anchor point features, obtain corresponding candidate trajectory features, evaluate the confidence of the candidate trajectory features, select a target trajectory feature from each candidate trajectory feature according to the confidence evaluation result, and perform path planning control of the vehicle based on the target trajectory feature.

[0030] See Figure 2 In one embodiment, a driving trajectory planning method is provided. The execution subject of the method can be a terminal or a server, including but not limited to steps S201 to S205.

[0031] Step S201: Obtain the first environmental feature containing environmental information of the current driving environment.

[0032] Current driving environment information refers to various data about the vehicle's current driving environment, such as road topology, traffic signs, the positions of other vehicles and pedestrians, and weather conditions. This information forms the basis for the vehicle's decision-making and planning.

[0033] The first environmental feature refers to the structured data used to describe the current driving environment, obtained after preliminary processing and abstraction of the raw environmental information. This feature is usually represented in the form of a vector or tensor and can be directly processed by subsequent algorithm modules.

[0034] There are several ways to acquire the first environmental feature containing environmental information about the current driving environment. For example, pre-stored road map data and traffic rule information can be manually input into the system to form a description of the current driving environment. Alternatively, raw environmental data can be directly collected by the vehicle's sensor system, such as cameras, radar, or lidar, and then a basic data processing module can perform simple format conversion and preliminary integration to generate a structured first environmental feature.

[0035] Step S202: Insert multiple anchor point features into the first environmental feature to obtain the second environmental feature.

[0036] Anchor point features are used to characterize the reference path points corresponding to the given sample driving trajectory. In essence, anchor point features are key reference path points used to characterize a specific sample driving trajectory during the driving trajectory planning process. These anchor point features serve as constraints or guides for trajectory reconstruction, capturing the core shape and key turning points of the sample trajectory. Reference path points are discrete points on the ideal or typical path that the vehicle should follow in a specific driving scenario. Anchor point features are abstract representations of these reference path points.

[0037] Sample driving trajectories refer to a series of representative driving behaviors or paths that are collected or generated in advance, such as the historical driving trajectory of the current vehicle. Each sample driving trajectory corresponds to a specific driving behavior, such as going straight, turning left, turning right, or changing lanes. These sample trajectories can cover a variety of driving scenarios and behavior patterns, providing a learning foundation for trajectory generation.

[0038] The second environmental feature refers to an enhanced environmental representation formed by inserting anchor point features on top of the first environmental feature. This feature integrates current environmental information with the guidance information required for trajectory reconstruction.

[0039] Inserting multiple anchor features into the first environment feature can be achieved by predefining a set of typical driving scenarios and specifying several key reference path points of sample driving trajectories as anchor points for each scenario. For example, scenario A might specify reference path points representing three sample driving trajectories: going straight, turning left, or turning right. When a vehicle enters a scenario, the corresponding anchor feature is directly selected from a predefined library and embedded into the first environment feature. Alternatively, historical driving data can be analyzed to identify representative sample driving trajectories, and representative reference path points from these trajectories can be selected as anchor features. These anchor features are then directly added to the first environment feature.

[0040] Step S203: Perform forward diffusion and reverse denoising on the second environmental features to reconstruct the corresponding driving trajectory based on the anchor point features, and obtain the corresponding candidate trajectory features.

[0041] Forward diffusion refers to the process in a diffusion model where noise is gradually added to data (such as anchor features), causing it to gradually evolve into a random noise distribution. This process simulates the evolution of information from clear to fuzzy.

[0042] Reverse denoising refers to the process opposite to forward diffusion, which involves progressively removing noise from noisy data to recover the original data or generate new data. In trajectory planning, this process is used to reconstruct the driving trajectory from noisy anchor point features.

[0043] Forward diffusion and backward denoising of the second environmental features can be performed by using the anchor features as initial conditions and gradually introducing Gaussian noise. Then, through a simple iterative process, the features after introducing Gaussian noise are subjected to small random perturbations in each iteration, and the trajectory shape is adjusted according to the distance from the anchor features until a series of trajectories that roughly match the anchor features are generated, thus obtaining candidate trajectory features. Alternatively, a basic generative model can be used to directly generate trajectories from random input and anchor features, and candidate trajectory features can be generated through repeated training.

[0044] Step S204: Optimize the confidence of the candidate trajectory features, and select the target trajectory features from each candidate trajectory feature based on the confidence evaluation results.

[0045] Confidence evaluation refers to a multi-dimensional assessment of the characteristics of generated candidate trajectories to measure their feasibility, safety, comfort, and compliance with environmental constraints. The evaluation results are used to select the optimal trajectory.

[0046] The target trajectory feature refers to the single or set of optimal driving trajectories selected from all candidate trajectory features after confidence evaluation, which best meets the current driving needs and environmental constraints. This feature will be directly used for vehicle path planning and control.

[0047] Confidence evaluation of candidate trajectory features can be achieved by setting a simple threshold rule and calculating a comprehensive score for each candidate trajectory feature. This score can be based on one or a few general indicators, such as trajectory length and average distance from the lane center. Trajectories with scores above a preset threshold are then considered qualified, and the trajectory with the highest score among these qualified trajectories is selected as the target trajectory feature. Alternatively, human experts can visually inspect the generated candidate trajectories and select the optimal target trajectory feature based on experience.

[0048] Step S205: Based on the target trajectory characteristics, perform path planning and control for the vehicle.

[0049] Path planning control refers to generating lateral and longitudinal control commands for a vehicle based on the selected target trajectory characteristics, in order to guide the vehicle to travel safely and smoothly along the target trajectory.

[0050] Based on the target trajectory characteristics, vehicle path planning and control can be performed. This can be achieved by directly converting the selected target trajectory characteristics into a series of discrete waypoints, which are then sequentially tracked by the vehicle's underlying controller (such as a PID controller). Alternatively, the target trajectory characteristics can be input into a simple trajectory tracking algorithm. This algorithm calculates the required steering angle and speed commands based on the deviation between the vehicle's current position and the target trajectory, and sends these commands directly to the vehicle's actuators.

[0051] The following example will provide a more detailed explanation of the above technical solution: Suppose a vehicle is traveling at location A and needs to perform a complex lane change maneuver. First, the executing entity acquires a first environmental feature containing environmental information about the current driving environment. Specifically, the vehicle's perception system collects environmental data such as the road geometry, lane markings, the relative positions and speeds of surrounding vehicles, and traffic light status. This raw data undergoes preliminary processing, such as format conversion and initial integration, to form a structured first environmental feature, which clearly describes the traffic scenario the vehicle is currently in.

[0052] Subsequently, multiple anchor features are inserted into the first environmental feature to obtain the second environmental feature. These anchor features are generated based on key reference path points corresponding to sample driving trajectories of similar scenarios in historical driving data. For example, in a lane-changing scenario, anchor features may include the lane-changing start point, the lane-changing midpoint, and the lane-changing end point. These anchor features are embedded into the first environmental feature, providing crucial guidance and constraints for subsequent trajectory generation, ensuring that the generated trajectory conforms to the expected behavioral pattern.

[0053] Based on this, forward diffusion and reverse denoising are performed on the second environmental feature to reconstruct the corresponding driving trajectory based on the anchor point feature, obtaining the corresponding candidate trajectory features. Specifically, the executing agent first treats the anchor point feature as an initial state and gradually adds random noise to it, simulating the evolution of the trajectory from clear to blurry (forward diffusion). Then, through a learned denoising model, noise is gradually removed from these noisy states, and combined with the guidance of the anchor point feature, multiple driving trajectories that conform to the corresponding behavioral intentions (left turn, right turn, lane change, etc.) are reconstructed. These reconstructed driving trajectories constitute the corresponding candidate trajectory features, and each candidate trajectory feature represents a behavioral intention.

[0054] Next, the confidence level of each candidate trajectory feature is evaluated, and the target trajectory feature is selected from these candidates based on the confidence level evaluation results. For example, the executing entity will assess the length of the trajectory corresponding to each candidate trajectory feature, the average distance from the lane center, whether there is a collision with surrounding vehicles, whether it exceeds the lane boundary, and whether it complies with traffic rules. By calculating and comprehensively evaluating these indicators, a confidence score is generated for each candidate trajectory. A trajectory with a higher score indicates that it performs better in terms of safety, comfort, and feasibility. Finally, the executing entity will select the trajectory with the highest confidence score as the target trajectory feature, which is considered the optimal driving path in the current scenario.

[0055] Finally, based on the target trajectory characteristics, path planning and control are performed on the vehicle. The selected target trajectory characteristics are sent to the vehicle's motion controller. The motion controller calculates and outputs the vehicle's steering angle, throttle, and braking commands in real time based on the target trajectory, thereby guiding the vehicle to precisely complete the lane change operation along the target trajectory. Through this series of steps, the vehicle can achieve safe and efficient driving trajectory planning in a complex lane change scenario, from environmental perception to final motion control.

[0056] Based on the above examples, the driving trajectory planning method provided in this embodiment demonstrates significant technical contributions. In the prior art, end-to-end autonomous driving systems often adopt a modular design, such as perception, prediction, and planning being independent of each other. This often leads to insufficient generalization ability in complex traffic scenarios. For example, when a vehicle changes lanes at location A, traditional modular methods may result in a less robust or unsuitable planned trajectory due to the accumulation of errors in the perception module or the limited ability of the prediction module to express multimodal behavior.

[0057] In contrast, this embodiment introduces anchor point features and combines forward diffusion and reverse denoising mechanisms to reconstruct the driving trajectory, effectively solving the problems of poor generalization ability and poor real-time performance of existing methods in complex scenarios. In the lane-changing example above, anchor point features, as abstractions of key points of the sample driving trajectory, provide strong guidance for trajectory reconstruction. This enables the executing agent to generate diverse candidate trajectories that conform to scenario constraints based on limited anchor point information, thereby avoiding the limitations of single regression models or fixed trajectory vocabulary sampling in traditional methods. The mechanism of gradually recovering the trajectory from noise through the diffusion model can better capture the multimodal characteristics of driving behavior and generate more natural trajectories that conform to actual driving habits. At the same time, since forward diffusion and reverse denoising operations are performed starting from anchor point features, the generation speed of candidate trajectory features is greatly improved, enhancing the real-time performance of driving trajectory planning.

[0058] In some embodiments, before obtaining the first environmental feature containing environmental information of the current driving environment, the method further includes: obtaining environmental sensing features that collect environmental information of the driving environment; and performing feature encoding on the environmental sensing features to obtain the first environmental feature.

[0059] This environmental sensing feature can be acquired through various vehicle-mounted sensors. For example, LiDAR can be used to acquire point cloud data for sensing obstacles, road boundaries, and terrain information; millimeter-wave radar can be used to acquire target speed, distance, and angle information for detecting dynamic targets such as vehicles and pedestrians; and cameras can be used to acquire image or video data for identifying lane lines, traffic signs, traffic lights, and other visual features. Furthermore, ultrasonic sensors can be used to acquire information about nearby obstacles, or GPS and inertial measurement units can be used to acquire the vehicle's own position, speed, and attitude information; all of this data can be used as part of the environmental sensing feature.

[0060] Feature encoding of environmental sensing characteristics refers to transforming raw, heterogeneous environmental sensing features into a unified, structured, and algorithm-friendly representation—the first environmental feature. Its purpose is to extract the most critical and effective information for driving trajectory planning, remove redundancy or noise, and provide high-quality input for subsequent anchor insertion and diffusion denoising processes. This feature encoding can be performed using deep learning models. For example, convolutional neural networks can be used to process image data to extract visual features, recurrent neural networks or Transformer models can be used to process sequence data to capture temporal information, or graph neural networks can be used to process the environmental topology after sensor fusion. Alternatively, traditional machine learning methods can be used for feature encoding. For example, feature engineering can be used to manually extract road geometry features, obstacle bounding boxes, lane line parameters, etc., and encode them into vector or matrix forms; or dimensionality reduction techniques such as principal component analysis can be used to compress and extract features from high-dimensional sensing data.

[0061] The proposed solution first acquires environmental sensor features by collecting environmental information about the driving environment. Then, it encodes these raw, multimodal environmental sensor features to generate the first environmental feature for subsequent trajectory planning. Specifically, various sensors on the vehicle, such as LiDAR, millimeter-wave radar, and cameras, continuously collect raw data about the surrounding environment, which constitutes the environmental sensor features. Subsequently, a dedicated feature encoding module processes this heterogeneous sensor data. For example, by fusing data from different sensors, it extracts key information such as road structure, obstacle locations, and traffic participant behavior, transforming it into a unified, structured representation. This encoding process not only effectively filters out noise and redundant information from the raw data but also abstracts the complexity of the environment into feature vectors or feature maps that are more compatible with trajectory planning algorithms. This approach ensures that the first environmental feature input to the subsequent trajectory planning module is of high quality and high accuracy, providing a solid foundation for subsequent steps such as anchor insertion, forward diffusion, and reverse denoising, significantly improving the accuracy and robustness of the final driving trajectory planning.

[0062] In one specific implementation, a vehicle can be equipped with a sensor suite, including a lidar mounted on the roof, multiple millimeter-wave radars mounted around the vehicle, and a high-resolution camera mounted behind the windshield. During vehicle operation, the lidar continuously scans the surrounding environment, generating 3D point cloud data to construct a geometric model of the local environment; the millimeter-wave radars monitor the speed and distance of dynamic targets such as vehicles and pedestrians in front and to the sides in real time; the camera captures road images to identify lane lines, traffic signs, and traffic lights. These raw lidar point clouds, millimeter-wave radar target lists, and camera image data constitute the environmental sensing features. These environmental sensing features are then input into a fusion perception module. This module can employ deep learning-based sensor fusion algorithms, for example, projecting the lidar point cloud onto a bird's-eye view feature map, extracting visual features from the camera images using a convolutional neural network, and embedding the millimeter-wave radar target information into these feature maps. Finally, the fusion perception module outputs a multi-channel feature map containing road topology, obstacle occupancy grids, drivable areas, and semantic information about traffic participants, serving as the first environmental feature.

[0063] Through the above technical solution, this application can effectively extract high-quality, high-precision first environmental features from raw, heterogeneous driving environment information. This explicit feature acquisition and encoding process avoids noise interference and information redundancy problems that may arise from directly using raw data, ensuring that the environmental information input into the subsequent trajectory planning process consists of optimized and abstracted key features. This allows subsequent steps such as anchor feature insertion, forward diffusion, and reverse denoising to be performed on a more reliable and accurate environmental perception basis, thereby significantly improving the accuracy of reconstructed driving trajectories and the effectiveness of confidence evaluation, ultimately enhancing the overall performance and safety of driving trajectory planning.

[0064] See Figure 3 In one embodiment, the method for generating anchor point features includes, but is not limited to, steps S301 to S303.

[0065] Step S301: Obtain features of multiple sample path points.

[0066] The sample path point features consist of multiple path point features used to characterize the trajectory information of the sample driving trajectory. In essence, sample path point features refer to a set of data extracted from historical driving data, simulation data, or expert driving demonstrations to describe the vehicle's state at a specific point in time or spatial location. Each path point feature may contain information such as the vehicle's position (e.g., latitude and longitude coordinates or X and Y coordinates in a local coordinate system), speed, acceleration, heading angle, and curvature. The collection of these sample path point features constitutes the trajectory information of the sample driving trajectory, collectively depicting the vehicle's actual or expected driving path and dynamic behavior in different scenarios.

[0067] These features can be obtained through data fusion processing of onboard sensors (such as GPS, IMU, radar, and cameras), or through simulation generation using high-precision map data combined with vehicle kinematic models.

[0068] Step S302: Cluster the path point features in the sample path point features to obtain several feature clusters.

[0069] Clustering pathpoint features within a sample is the process of grouping these features together, aiming to classify similar features into a single feature cluster. This clustering operation can be based on the spatial similarity of pathpoint features, kinematic parameter similarity (such as velocity, heading, and curvature), or a combination of both. For example, K-means, DBSCAN, or hierarchical clustering algorithms can be used. Through clustering, a large number of discrete sample pathpoint features can be organized into a set with inherent structure and patterns, thereby revealing typical path patterns under different driving scenarios or driving behaviors.

[0070] Step S303: Based on the number of path point features in the feature cluster, select several path point features as cluster centers to obtain anchor point features.

[0071] After obtaining the feature clusters, the selection of anchor features is based on the number of path point features contained in each cluster. Clusters with a larger number of features typically represent driving patterns or path regions that occur more frequently or are more representative in the sample data. Therefore, the cluster centers of these clusters can be selected as anchor features. The cluster center can be the average value of all path point features in the cluster (centroid), or it can be the actual path point in the cluster closest to the centroid (center point). This selection method ensures that the generated anchor features effectively cover the main patterns in the sample driving trajectory and provide statistically significant reference points for subsequent trajectory reconstruction.

[0072] This application's solution systematically extracts and refines anchor point features from a large number of sample driving trajectories to optimize the initial conditions for driving trajectory planning. First, by acquiring multiple sample path point features containing information such as vehicle position, speed, and heading, the dynamic and geometric characteristics of the sample driving trajectories are comprehensively captured. These sample path point features form the basis for subsequent analysis, collectively constituting a rich description of historical or expected driving behavior. Second, these sample path point features are clustered, grouping path points with similar characteristics into different feature clusters. This step effectively reduces the dimensionality and performs pattern recognition on the complex sample data, revealing potential typical driving patterns. Through clustering, multiple driving paths and behavioral patterns that vehicles may adopt in different scenarios can be identified. Finally, based on the number of path point features in each feature cluster, several path point features are selected as cluster centers, thus obtaining the final anchor point features. This quantity-based selection strategy ensures that the selected anchor point features are not only representative of each cluster but also prioritize path patterns that occur frequently and are highly representative in the sample data. These optimized anchor features are then inserted into the first environmental features to form the second environmental features, providing high-quality reference points for subsequent forward diffusion and backward denoising processes. In this way, the proposed solution can generate more representative anchor features that better reflect real-world driving scenarios, thereby making the candidate trajectory features reconstructed based on these anchor features more accurate and diverse, significantly improving the effectiveness and robustness of the overall driving trajectory planning.

[0073] As a specific implementation method, when acquiring multiple sample pathpoint features, GPS positioning data, inertial measurement unit (IMU) data, and vehicle bus data collected by autonomous driving test vehicles under different road types (such as highways, urban roads, and rural roads) and traffic conditions can be utilized. After time synchronization and fusion processing, the vehicle position (e.g., latitude and longitude in the WGS84 coordinate system, or converted to X and Y coordinates in a local Cartesian coordinate system), speed, acceleration, and heading angle corresponding to each timestamp can be extracted. This information is then combined into a pathpoint feature vector. For example, a pathpoint feature can be represented as [X, Y, speed, heading angle]. When clustering the pathpoint features in the sample pathpoint features, the K-means clustering algorithm can be used. First, a suitable number of clusters K is determined empirically or through the silhouette coefficient method. Then, all sample pathpoint features are used as input, and the K-means algorithm is run to divide them into K feature clusters. During clustering, the spatial location (X, Y coordinates) of pathpoint features can be primarily considered as the clustering dimension to identify spatially similar sets of pathpoints. When selecting several pathpoint features as cluster centers based on the number of pathpoint features in the feature clusters, the number of pathpoint features contained in each feature cluster can be counted. A threshold is set; for example, if the number of pathpoint features in a cluster exceeds 1% of the total number of pathpoint features in the sample, the cluster is considered sufficiently representative. For clusters that meet this condition, the average value of all their pathpoint features is calculated to obtain the centroid of the cluster. These centroids are used as the final anchor features. For example, if the centroid of a cluster is [X_c, Y_c, V_c, H_c], then this centroid is an anchor feature used to characterize the typical driving pathpoints represented by that cluster.

[0074] Through the above technical solution, this application can systematically and efficiently extract representative anchor point features from massive sample driving trajectory data. This anchor point generation method based on clustering and quantity screening avoids the inaccuracies and incompleteness that may result from random or empirical anchor point selection. The generated anchor point features can more accurately reflect typical path patterns and driving behaviors in actual driving scenarios, thus providing a more solid and reliable foundation for the subsequent trajectory reconstruction process. When these high-quality anchor point features are inserted into the first environment features, they can guide the diffusion model to generate candidate trajectory features that are more in line with actual road conditions and driving habits, significantly improving the accuracy, diversity, and physical feasibility of trajectory planning, thereby enhancing the intelligence level and safety of vehicle path planning and control.

[0075] See Figure 4In one embodiment, the method for forward diffusion and reverse denoising of the second environmental feature includes, but is not limited to, steps S401 to S403.

[0076] Step S401: Add Gaussian noise to the anchor point features to obtain the initial noisy features.

[0077] Step S402: Perform forward diffusion on the initial noisy features to obtain noisy state features.

[0078] Step S403: Perform reverse denoising on the noisy state features to obtain candidate trajectory features.

[0079] Adding Gaussian noise to anchor features aims to provide a starting point with initial random perturbation for the subsequent forward diffusion process. Introducing this noise into the anchor features increases the diversity of the data samples and provides rich training data for the diffusion model to learn the denoising process. This step can be implemented in two ways: one is to directly superimpose Gaussian random numbers with a specific mean (e.g., zero) and variance (e.g., a preset fixed value or a value dynamically adjusted based on the anchor features) onto the numerical representation of the anchor features; the other is to utilize a dedicated noise generation module that takes the anchor features as input and outputs Gaussian noise matching its dimensions, which is then superimposed onto the anchor features.

[0080] Forward diffusion of initial noisy features aims to gradually transform them into a more highly randomized state. The forward diffusion process is typically a Markov chain, iteratively adding noise to the features at multiple time steps, causing the features to gradually lose their original structural information and eventually approach a pure noise distribution. This process provides the inverse denoising model with data samples at different noise levels, enabling it to learn the ability to recover the original information under various noise intensities. This step can be implemented in two ways: one is by using a pre-defined noise scheduling function, adding Gaussian noise to the features of the current state at each time step t according to the scheduling function, and repeating this process until a preset maximum time step or noise level is reached; the other is by designing a parameterized diffusion function that can directly calculate the corresponding noisy state features based on given time step or noise level parameters, thus avoiding the computational overhead of iterative iteration.

[0081] Reverse denoising of noisy state features aims to progressively recover clear and meaningful trajectory information from highly noisy states. The reverse denoising process is typically performed by a trained deep learning model (e.g., a neural network-based denoiser) that learns to predict and remove noise added during forward diffusion. By iteratively subtracting the predicted noise from the noisy state features, the model can progressively transform the features from a purely noisy state back to a trajectory shape close to that represented by the original anchor features. This step can be implemented in two ways: one is by using a U-Net architecture neural network model that takes the noisy state features and the current time step as input, outputs the predicted noise, subtracts the predicted noise from the noisy state features to obtain the denoised features, and repeats this process until the initial time step is reached; another approach is to use a score matching-based method, training a model to estimate the gradient of the data distribution (i.e., the fractional function), and then using sampling methods such as Langevin dynamics to iteratively update the noisy state features along the direction of the fractional function to achieve denoising and trajectory reconstruction.

[0082] This application's scheme uses anchor point features as the starting point for trajectory reconstruction and introduces Gaussian noise to lay the foundation for the subsequent diffusion process. The forward diffusion process systematically transforms anchor point features into a series of noisy state features with different noise levels. This not only simulates the various uncertainties that may exist in trajectories in the real world but also provides rich learning samples for the reverse denoising model. Subsequently, the reverse denoising process utilizes the learned denoising capabilities to gradually recover clear candidate trajectory features from these noisy state features. This progressive reconstruction mechanism from noise to clarity enables the system to effectively generate diverse and high-quality trajectories from sparse or incomplete anchor point information, while ensuring the smoothness and physical feasibility of the reconstructed trajectory. This solves the problem of how to construct an accurate and robust diffusion-denoising mechanism to generate high-quality trajectories.

[0083] The following is a concrete example. Assume the anchor features are represented as a series of two-dimensional coordinate points; for example, each anchor feature can be a (x, y) coordinate pair. When adding Gaussian noise to the anchor features, for each anchor feature's (x, y) coordinate, a Gaussian random number with a mean of 0 and a standard deviation of σ can be superimposed on both the x and y components, resulting in an initial noisy feature (x', y'). Here, σ can be a preset small value, such as 0.1. Subsequently, the initial noisy feature is forward diffused. This process can be a multi-step iterative process. For example, at each time step t (from 1 to T), according to a preset noise scheduling function β_t, the current feature is weighted and mixed with random noise following a standard Gaussian distribution, gradually increasing the noise level of the feature until, at time step T, the feature is completely transformed into pure Gaussian noise. For example, the diffusion process in DDPM (Denoising Diffusion Probabilistic Models) can be used. Finally, the noisy state feature is reverse-denoised. This can be accomplished by a deep neural network model (e.g., a denoising network based on the U-Net architecture). This network receives the noisy state features at a certain time step t and time step t as input, and is trained to predict the noise to be added to those features. By subtracting the predicted noise from the noisy features, a slightly "cleaner" feature is obtained. This process starts at time step T and iterates progressively up to time step 1, ultimately generating a denoised candidate trajectory feature, which is a series of smooth path points that conform to anchor constraints.

[0084] Through the above technical solution, this application provides an efficient and robust trajectory reconstruction mechanism. By adding Gaussian distributed noise to the anchor point features and performing forward diffusion, necessary randomness and diversity can be introduced into trajectory generation, avoiding the uniformity of generated trajectories. Subsequently, through a reverse denoising process, high-quality candidate trajectory features can be accurately recovered from these noisy state features, ensuring the smoothness, physical feasibility, and effective utilization of anchor point information in the reconstructed trajectory. This significantly improves the accuracy and adaptability of driving trajectory planning, especially in the face of complex and ever-changing driving environments, enabling the generation of driving trajectories that better meet actual needs.

[0085] In some embodiments, the noisy state features are the noisy features prior to convergence to isotropic standard Gaussian noise features.

[0086] Converging to isotropic standard Gaussian noise features refers to the final state of the diffusion process, where noise is sufficiently added to completely mask the original information in the feature, and the noise exhibits the same statistical characteristics across all dimensions, conforming to a standard Gaussian distribution. Noisy features preceding convergence to isotropic standard Gaussian noise features refer to the process of preventing the diffusion from reaching a state of pure random noise during forward diffusion; instead, it stops before reaching that state. For example, it can stop when the number of diffusion steps reaches a preset threshold, or when the signal-to-noise ratio of the noisy feature falls below a certain preset value.

[0087] This application reconstructs the driving trajectory by performing forward diffusion and reverse denoising on the second environmental features. The forward diffusion process aims to gradually add noise to the anchor point features to simulate the evolution of the trajectory from clear to blurry, generating a series of noisy state features. To improve the generation speed of candidate trajectory features and thus enhance the real-time performance of driving trajectory planning, this application limits the noisy state features to those that converge to isotropic standard Gaussian noise features. This means that during the forward diffusion process, the number of diffusion steps will not reach a completely randomized level, but will stop at an intermediate state where the trajectory information is not completely lost and some structural features are still retained. In this way, it is ensured that the noisy state features still contain enough original trajectory information. While shortening the forward diffusion time, only a very small number of reverse denoising steps (e.g., 2 steps or a few steps) are needed to obtain the required candidate trajectory features, significantly reducing the number of iterations for both forward diffusion and reverse denoising.

[0088] As a specific implementation method, a Markov chain-based diffusion process can be used when forward diffusing the initial noisy features to obtain noisy state features. In this process, Gaussian noise is gradually added to the features through a series of discrete time steps. To ensure that the noisy state features are the noisy features before converging to isotropic standard Gaussian noise features, a maximum diffusion step threshold can be set. For example, when the diffusion steps reach 80% of the total diffusion steps, the forward diffusion process stops, and the noisy features at this point are taken as the noisy state features. Alternatively, the signal-to-noise ratio (SNR) of the noisy features can be monitored in real time. When the SNR drops to a preset critical value, diffusion stops, and the features at that moment are taken as the noisy state features. For example, a pre-trained noise scheduler can be used to control the amount of noise added at each step and truncate the diffusion process according to a preset stopping condition, ensuring that the generated noisy state features have sufficient noise diversity without completely losing the structural information of the original trajectory.

[0089] By using the above technical solution, the noisy state features in the forward diffusion process are limited to the state before they converge to isotropic standard Gaussian noise features. This ensures that the noisy state features still contain enough original trajectory information. On the basis of shortening the forward diffusion time, only a very small number of reverse denoising steps are needed to obtain the candidate trajectory features that meet the requirements. This greatly reduces the number of iterations of both forward diffusion and reverse denoising, improves the generation speed of candidate trajectory features, and thus improves the real-time performance of driving trajectory planning.

[0090] In some embodiments, before evaluating the confidence level of the candidate trajectory features, the method further includes: acquiring vehicle state features and map information features; extracting features from the first environmental features to obtain road condition features; and performing cascaded denoising on the candidate trajectory features based on the road condition features, vehicle state features, and map information features to obtain denoised candidate trajectory features.

[0091] Vehicle status characteristics refer to the feature information describing the current motion state and internal parameters of a vehicle. The concept encompasses the vehicle's dynamic and static attributes at a specific moment, such as speed, acceleration, angular velocity, steering angle, gear position, and braking status. These characteristics can be acquired in real time through the vehicle's internal sensor network (such as CAN bus data, IMU sensors, wheel speed sensors, etc.) or read through the on-board diagnostic (OBD) interface.

[0092] Map information features refer to the characteristic information describing the geospatial and road topology of the driving environment. These concepts include, but are not limited to, road geometry (such as lane lines, curbs, and curvature), traffic rules (such as speed limits and no-entry zones), traffic facilities (such as traffic lights and signs), and semantic information contained in high-precision maps (such as lane type, gradient, and road surface material). These features can be retrieved from pre-stored high-precision map databases or through in-vehicle navigation systems or real-time map service interfaces.

[0093] Road condition features refer to the characteristic information extracted from the primary environmental features that reflects the current road traffic conditions and environmental obstacles. These features include, but are not limited to, the location, type, and speed of obstacles ahead, traffic flow density, lane occupancy, road surface slipperiness, and visibility. The extraction of road condition features can utilize deep learning models to analyze and identify the primary environmental features (e.g., raw or pre-processed data from sensors such as cameras, radar, and lidar). For example, target detection algorithms can be used to identify vehicles and pedestrians, while semantic segmentation algorithms can be used to identify drivable areas and obstacles.

[0094] Cascaded denoising refers to a multi-stage, multi-information fusion process for noise suppression and feature optimization. Its function is to iteratively or hierarchically refine the initially generated candidate trajectory features using multi-source information (road condition features, vehicle state features, and map information features) to eliminate residual noise, correct unreasonable parts, and improve the smoothness, physical feasibility, and environmental adaptability of the trajectory. Cascaded denoising can be implemented using various algorithms. For example, methods based on Kalman filtering or particle filtering can be used, taking the candidate trajectory features as observations and combining them with vehicle dynamics models and environmental constraints for state estimation and prediction, thereby smoothing the trajectory. Alternatively, a trajectory optimization model based on deep neural networks can be used, taking the candidate trajectory features along with road conditions, vehicle state, and map information as input, and outputting smoother and more constraint-compliant denoised trajectory features.

[0095] This application's solution introduces vehicle state features, map information features, and road condition features extracted from the first environmental features. Before evaluating the confidence of candidate trajectory features, these features undergo cascaded denoising. Specifically, after initially generating candidate trajectory features, the system acquires the vehicle's current precise state information, such as speed, acceleration, and steering angle, while simultaneously querying high-precision map information for the current area, including lane lines, speed limits, and road curvature. Furthermore, through in-depth analysis of the first environmental features, real-time road condition information is extracted, such as obstacles ahead, traffic congestion, or special road surface conditions. Subsequently, this multi-source information (road condition features, vehicle state features, and map information features) is used as input to perform cascaded denoising on the initially generated candidate trajectory features. This process utilizes the vehicle's own kinematic and dynamic constraints, the road's geometric and semantic constraints, and real-time environmental obstacle information to refine and optimize the candidate trajectories. For example, if a candidate trajectory initially exhibits uneven turns or slight deviations from lane lines, the cascaded denoising process adjusts it by incorporating the vehicle's steering capabilities and lane information from the map, making it smoother and more strictly aligned with the lanes. In this way, potential noise or inconsistencies in the original candidate trajectory features are effectively suppressed and corrected, resulting in higher-quality, more realistically suited denoised candidate trajectory features. This series of operations allows subsequent confidence evaluation to be performed on a more reliable and accurate set of trajectories, significantly improving the accuracy and safety of the final trajectory planning.

[0096] As a specific implementation method, after acquiring candidate trajectory features, vehicle status features, such as current speed, steering wheel angle, throttle opening, and brake pedal travel, can first be obtained via the vehicle's CAN bus. Simultaneously, map information features, including the current lane centerline, lane width, upcoming intersection information, and speed limit signs, are obtained from the vehicle's high-precision map module. Next, using the first environmental features collected by the vehicle's cameras and radar sensors, a pre-trained convolutional neural network model is used to extract road condition features, such as identifying the position and speed of vehicles ahead, pedestrians, traffic cones, and other obstacles, as well as determining whether there is standing water or potholes on the road surface. Subsequently, these vehicle status features, map information features, and road condition features, along with the initially generated candidate trajectory features, are input into a deep learning-based trajectory optimization network. This network can be a sequence-to-sequence model, such as a Transformer or LSTM network, containing multiple denoising layers, each using different types of contextual information to correct the trajectory. For example, the first layer might focus on smoothing the trajectory using vehicle dynamics models, the second layer might combine map information to ensure the trajectory doesn't deviate from the lane, and the third layer might use road condition features to adjust the trajectory to avoid obstacles. Through multi-level iterative optimization, a series of smooth, safe, and environmentally compliant denoised candidate trajectory features are finally output.

[0097] By employing the aforementioned technical solution, a cascaded denoising process based on multi-source information is introduced before evaluating the confidence level of candidate trajectory features, effectively improving the quality of these features. This allows subsequent confidence evaluations to be based on more accurate and reliable trajectory data, thereby significantly enhancing the safety, comfort, and physical feasibility of the final selected driving trajectory. This solution effectively addresses the noise and uncertainty issues that may exist in the initially generated candidate trajectories, providing a solid foundation for accurate and robust path planning in complex and ever-changing driving environments.

[0098] See Figure 5 In one embodiment, the method for evaluating the confidence of candidate trajectory features includes, but is not limited to, steps S501 to S503.

[0099] Step S501: Calculate the consistency confidence score of the candidate trajectory features based on the smoothness index, physical feasibility index, and trajectory morphology stability index of the candidate trajectory features. Step S502: Calculate the external constraint compliance score based on the first environmental characteristics; Step S503: Based on the consistency confidence score and the external constraint compliance score, generate a comprehensive confidence evaluation score for the candidate trajectory features.

[0100] The smoothness index is used to measure the continuity and stability of a trajectory to avoid abrupt steering or acceleration / deceleration during vehicle operation. This index can be quantified by calculating the rate of change of trajectory curvature, the norm of acceleration or jerk, or indirectly reflected by assessing changes in distance and angle between trajectory points.

[0101] Physical feasibility metrics are used to assess whether a trajectory conforms to the vehicle's own dynamic constraints and physical limitations. These metrics may include checking whether the speed, acceleration, and steering angular velocity on the trajectory are within the vehicle's maximum permissible range, and may also consider factors such as tire grip and weight transfer to ensure the trajectory is physically feasible.

[0102] The trajectory morphology stability index measures whether a trajectory maintains its shape and is not prone to drastic changes when faced with minor disturbances or environmental changes. This index can be quantified by assessing the trajectory's sensitivity to changes in initial conditions or environmental parameters, such as the degree of shape change under small disturbances; it can also be determined by analyzing the trajectory's geometric characteristics, such as the continuity of the radius of curvature and the presence of self-intersections.

[0103] The consistency confidence score is a comprehensive assessment of the intrinsic quality of the trajectory itself. This score can be a weighted average of the smoothness, physical feasibility, and trajectory morphological stability indicators mentioned above, or it can be obtained through the fusion of these indicators using a machine learning model. Alternatively, it can be based on fuzzy logic or an expert system, making a comprehensive judgment based on the degree of conformity of each indicator.

[0104] The external constraint compliance score is used to assess the compatibility of a trajectory with its external environment (such as road boundaries, obstacles, and traffic rules). This score can be quantified by calculating the distance between the trajectory and road boundaries, the safe distance from obstacles, and whether traffic rules (such as lane markings and speed limits) are violated. Alternatively, it can be determined by analyzing the trajectory's position within the first environmental feature to assess whether it is within a drivable area and whether there is a risk of collision with dynamic obstacles.

[0105] The overall confidence score combines the intrinsic quality of the trajectory with its compatibility with the external environment to provide a final comprehensive evaluation of the trajectory. This score can be a weighted sum, product, or other fusion function of the consistency confidence score and the external constraint compliance score, or it can be achieved through a multi-objective optimization algorithm that maximizes the overall confidence score while satisfying all constraints.

[0106] The proposed solution, when evaluating the confidence level of candidate trajectory features, first starts from the intrinsic attributes of the trajectory itself. It quantifies the intrinsic quality of the trajectory by calculating its smoothness index, physical feasibility index, and trajectory morphological stability index, and generates a consistency confidence score accordingly. This ensures that the trajectory remains comfortable, stable, and conforms to vehicle dynamics constraints during vehicle execution. Simultaneously, to ensure the trajectory's compatibility with the actual driving environment, the solution further calculates the trajectory's external constraint compliance score based on a first environmental feature, assessing whether the trajectory safely avoids obstacles, complies with traffic rules, and adapts to road structures. Finally, a comprehensive confidence evaluation score is generated by fusing the consistency confidence score and the external constraint compliance score. This multi-dimensional, comprehensive evaluation mechanism ensures that the target trajectory features selected from the candidate trajectory features not only exhibit excellent intrinsic quality but also highly match complex driving environments, thereby significantly improving the reliability and safety of path planning.

[0107] As a specific implementation, when calculating the smoothness index of the candidate trajectory features, the rate of curvature change at each point on the trajectory can be evaluated. For example, the smoothness can be measured by calculating the curvature of three consecutive points on the trajectory and obtaining their second derivative. For the physical feasibility index, it can be checked whether the velocity, acceleration, and lateral acceleration at any point on the trajectory exceed the vehicle's maximum design limits. For example, the maximum longitudinal acceleration can be set to 3 m / s², and the maximum lateral acceleration can be set to 5 m / s². The trajectory morphological stability index can be evaluated based on the degree of morphological change after a small perturbation to the trajectory. For example, a small velocity or angle perturbation can be applied at the starting point of the trajectory, and then the offset at the end of the trajectory can be observed. The consistency confidence score can be obtained by comprehensively evaluating these indices using a pre-trained neural network model. The external constraint compliance score can be calculated based on the distance and interaction between the trajectory and elements such as lane lines, obstacles, and traffic signs identified in the first environmental features. For example, the minimum distance between the trajectory and the nearest obstacle should be greater than a safety threshold, and the trajectory should always remain within the drivable area. Ultimately, the overall confidence score can be obtained by weighting and summing the consistency confidence score and the external constraint compliance score. The weights can be adjusted according to the needs of the actual application scenario. For example, in the highway scenario, the weight of the smoothness index can be appropriately increased, while in the urban congestion scenario, the weight of the external constraint compliance index can be higher.

[0108] Through the aforementioned technical solution, the confidence evaluation of candidate trajectory features is no longer limited to a single dimension, but comprehensively considers the intrinsic quality of the trajectory itself, such as its smoothness, physical feasibility, and morphological stability, as well as its degree of conformity with the external environment reflected by the first environmental features. This comprehensive evaluation mechanism can effectively avoid selecting trajectories that perform well in some aspects but have defects in other key aspects, thereby significantly improving the reliability, safety, and comfort of the selected target trajectory features. Vehicles can perform path planning and control based on more optimized trajectories, improving the decision-making quality of the autonomous driving system and the user experience.

[0109] See Figure 6 This application also provides a driving trajectory planning device that can implement the above-described driving trajectory planning method. The device includes: The first module 601 is used to acquire a first environmental feature containing environmental information of the current driving environment; The second module 602 is used to insert multiple anchor point features into the first environmental features to obtain the second environmental features; the anchor point features are used to characterize the reference path points corresponding to the sample driving trajectory. The third module 603 is used to perform forward diffusion and reverse denoising on the second environmental features in order to reconstruct the corresponding driving trajectory based on the anchor point features and obtain the corresponding candidate trajectory features. The fourth module 604 is used to evaluate the confidence of candidate trajectory features and select target trajectory features from each candidate trajectory feature based on the confidence evaluation results. The fifth module, 605, is used for path planning and control of vehicles based on target trajectory characteristics.

[0110] The specific implementation of this driving trajectory planning device is basically the same as the specific implementation of the driving trajectory planning method described above, and will not be repeated here.

[0111] Figure 7 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application.

[0112] The following reference Figure 7 To describe an electronic device 700 according to such an embodiment of the present disclosure. Figure 7 The electronic device 700 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0113] like Figure 7 As shown, the electronic device 700 is presented in the form of a general-purpose computing device. The components of the electronic device 700 may include, but are not limited to: at least one processing unit 710, at least one storage unit 720, a bus 730 connecting different system components (including storage unit 720 and processing unit 710), a display unit 740, etc.

[0114] The storage unit stores program code, which can be executed by the processing unit 710, causing the processing unit 710 to perform the steps described in the above-described driving trajectory planning method section of this specification according to various exemplary embodiments of this disclosure.

[0115] Storage unit 720 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 7201 and / or cache memory 7202, and may further include a read-only memory (ROM) 7203.

[0116] The storage unit 720 may also include a program / utility 7204 having a set (at least one) program module 7205, such program module 7205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0117] Bus 730 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0118] Electronic device 700 can also communicate with one or more external devices 700' (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 700, and / or with any device that enables electronic device 700 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 750. Furthermore, electronic device 700 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 760. Network adapter 760 can communicate with other modules of electronic device 700 via bus 730. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 700, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0119] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0120] The driving trajectory planning method, apparatus, device, and medium provided in this application insert anchor point features into the current driving environment features. These anchor point features characterize the reference path points corresponding to the sample driving trajectory. Combined with a diffusion denoising process, diverse driving trajectories are reconstructed, and then the target trajectory is selected through confidence evaluation, thereby enabling path planning and control of the vehicle. Thus, by introducing anchor point features and combining forward diffusion and backward denoising mechanisms to reconstruct the driving trajectory, the anchor point features, as abstractions of key points of the sample driving trajectory, provide powerful guidance for trajectory reconstruction. Based on limited anchor point information, diverse candidate trajectories that conform to scene constraints can be generated, thereby avoiding the limitations of single regression models or fixed trajectory vocabulary sampling. In the process of gradually recovering the trajectory from noise through a diffusion model, the multimodal characteristics of driving behavior can be better captured, generating more natural trajectories that are more in line with actual driving habits. Simultaneously, since forward diffusion and backward denoising operations are performed starting from anchor point features, the generation speed of candidate trajectory features is greatly improved, enhancing the real-time performance of driving trajectory planning.

[0121] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to 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, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the methods described above according to the embodiments of this disclosure.

[0122] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0123] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0124] Those skilled in the art will understand that the above modules can be distributed in the device as described in the embodiments, or they can be modified accordingly and placed in one or more devices that are unique to this embodiment. The modules in the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0125] Exemplary embodiments of this disclosure have been specifically shown and described above. It should be understood that this disclosure is not limited to the detailed structures, arrangements, or implementations described herein; rather, this disclosure is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended claims.

Claims

1. A driving trajectory planning method, characterized in that, include: Obtain the first environmental feature containing environmental information about the current driving environment; Multiple anchor point features are inserted into the first environmental feature to obtain the second environmental feature; the anchor point features are used to characterize the reference path points corresponding to the corresponding sample driving trajectory. The second environmental feature is subjected to forward diffusion and reverse denoising to reconstruct the corresponding driving trajectory based on the anchor point feature, thereby obtaining the corresponding candidate trajectory feature; The confidence level of the candidate trajectory features is evaluated, and the target trajectory feature is selected from the candidate trajectory features based on the confidence level evaluation results. Based on the target trajectory characteristics, the vehicle is subjected to path planning and control.

2. The driving trajectory planning method according to claim 1, characterized in that, Before acquiring the first environmental feature containing environmental information about the current driving environment, the method further includes: Acquire environmental sensing features that collect environmental information about the driving environment; The environmental sensing features are encoded to obtain the first environmental feature.

3. The driving trajectory planning method according to claim 1, characterized in that, The method for generating the anchor point features includes: Multiple sample path point features are obtained; the sample path point features are composed of multiple path point features and are used to characterize the trajectory information of the sample driving trajectory. Clustering is performed on the path point features in the sample path point features to obtain several feature clusters; Based on the number of path point features in the feature cluster, several path point features are selected as cluster centers to obtain the anchor point features.

4. The driving trajectory planning method according to claim 1, characterized in that, The forward diffusion and reverse denoising of the second environmental feature includes: Gaussian noise is added to the anchor point features to obtain the initial noisy features; The initial noisy features are forward diffused to obtain noisy state features; The noisy state features are denoised in reverse to obtain the candidate trajectory features.

5. The driving trajectory planning method according to claim 4, characterized in that, The noisy state features are the noisy features before they converge to isotropic standard Gaussian noise features.

6. The driving trajectory planning method according to claim 1, characterized in that, Before evaluating the confidence level of the candidate trajectory features, the method further includes: Acquire vehicle status features and map information features; The first environmental feature is used to extract road condition features; Based on the road condition features, vehicle status features, and map information features, the candidate trajectory features are cascaded for denoising to obtain denoised candidate trajectory features.

7. The driving trajectory planning method according to claim 1, characterized in that, The confidence evaluation of the candidate trajectory features includes: Based on the smoothness index, physical feasibility index, and trajectory morphology stability index of the candidate trajectory features, the consistency confidence score of the candidate trajectory features is calculated. Based on the first environmental characteristic, calculate the external constraint compliance score; Based on the consistency confidence score and the external constraint compliance score, a comprehensive confidence evaluation score for the candidate trajectory features is generated.

8. A driving trajectory planning device, characterized in that, include: The first module is used to acquire a first environmental feature containing environmental information about the current driving environment; The second module is used to insert multiple anchor point features into the first environmental features to obtain the second environmental features; the anchor point features are used to characterize the reference path points corresponding to the corresponding sample driving trajectory. The third module is used to perform forward diffusion and reverse denoising on the second environmental features in order to reconstruct the corresponding driving trajectory based on the anchor point features and obtain the corresponding candidate trajectory features. The fourth module is used to evaluate the confidence level of the candidate trajectory features and select the target trajectory feature from the candidate trajectory features based on the confidence level evaluation results. The fifth module is used to perform path planning and control of the vehicle based on the target trajectory characteristics.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the driving trajectory planning method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the driving trajectory planning method according to any one of claims 1 to 7.

Images