End-to-end automatic driving track determination method and related product

Abstract

The invention discloses an end-to-end automatic driving track determination method and a related product. In the scheme, based on sensor data, map information and a navigation target, an environment context feature vector of the autonomous vehicle is obtained; performing automatic driving track flow direction prediction based on the current track state information, the current time and the environment context feature vector of the automatic driving vehicle, and determining an initial flow direction prediction result; correcting the initial flow direction prediction result by using the constraint correction flow direction to obtain a target flow direction prediction result; the constraint correction flow direction is determined based on the physical constraint and the security rule constraint; the physical constraint is constructed based on a Riemannian metric inverse matrix; the security rule constraint is constructed based on the semantic potential field gradient; and processing the target flow direction prediction result to obtain a trajectory determination result of the autonomous vehicle. Compared with the problems of low planning precision and poor real-time response capability in the prior art, the method has obvious advantages.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to an end-to-end autonomous driving trajectory determination method and related products. Background Technology

[0002] With the development of deep learning, end-to-end autonomous driving has gradually become a research hotspot. End-to-end autonomous driving can directly take environmental perception data, such as raw data collected by cameras or LiDAR, as input and output trajectory information directly through deep neural network models.

[0003] Existing end-to-end autonomous driving trajectory determination methods mostly employ generative planning, but this approach still has certain shortcomings in complex urban conditions: existing models often perform data fitting in a flat Euclidean space, ignoring the kinematic characteristics of the vehicle as a nonholonomic constrained system, such as the inability to sideslip or curvature continuity, making it prone to instantaneous lateral movement or unnatural curvature abrupt changes. This deficiency makes the planning results heavily reliant on cumbersome post-processing corrections, which not only limits the accuracy of the planning but also reduces the system's real-time response capability in extreme obstacle avoidance scenarios.

[0004] Improving the accuracy of autonomous driving trajectory determination results is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] To address the aforementioned issues, this application provides an end-to-end autonomous driving trajectory determination method and related products, with the aim of improving the accuracy of autonomous driving trajectory determination results.

[0006] The embodiments of this application disclose the following technical solutions: The first aspect of this application provides an end-to-end autonomous driving trajectory determination method, the method comprising: Based on sensor data, map information, and navigation targets, the environmental context feature vector of autonomous vehicles is obtained; Based on the current trajectory state information, current time, and environmental context feature vector of the autonomous vehicle, the autonomous driving trajectory flow direction is predicted to determine the initial flow direction prediction result; The initial flow direction prediction result is corrected by using constraint-corrected flow direction to obtain the target flow direction prediction result; the constraint-corrected flow direction is determined based on physical constraints and safety rule constraints; the physical constraints are constructed based on the Riemann metric inverse matrix; the safety rule constraints are constructed based on the semantic potential gradient. The target flow prediction results are processed to obtain the trajectory determination results of the autonomous vehicle.

[0007] Optionally, the step of predicting the autonomous driving trajectory flow direction based on the current trajectory state information of the autonomous vehicle, the current time, and the environmental context feature vector, and determining the initial flow direction prediction result, includes: The flow model is used to predict the flow direction of the autonomous driving trajectory based on the current trajectory state information, current time, and environmental context feature vector of the autonomous vehicle, and the initial flow direction prediction result is determined; the flow model is obtained after training.

[0008] Optionally, the training method of the stream model includes: By sampling real driving datasets, a set of expert real trajectories is obtained; By sampling from a standard Gaussian distribution, a set of noise trajectories is obtained; A pairing mapping between the expert real trajectory set and the noise trajectory set is established based on the Riemann distance matching mechanism to obtain the path mapping function, and an initial flow field is constructed based on the path mapping function; The flow model is subjected to supervised learning based on the initial flow field. The parameters of the flow model are determined by minimizing the conditional flow matching loss function between the predicted flow direction and the target flow direction.

[0009] Optionally, the step of establishing a pairing mapping between the expert real trajectory set and the noise trajectory set based on the Riemann distance matching mechanism to obtain the path mapping function includes: Based on the set of expert real trajectories and the set of noisy trajectories, a Riemann cost matrix is ​​constructed; Based on the Riemann cost matrix, the path mapping function from noisy samples to expert samples is determined by minimizing the total global matching cost.

[0010] Optionally, the step of correcting the initial flow direction prediction result using constraint-corrected flow direction to obtain the target flow direction prediction result includes: The target flow direction prediction result is obtained based on the following formula: ; in, This indicates the predicted direction of the target flow; This indicates the initial flow direction prediction result; Indicates the weighting coefficient; denoted as the Riemann metric inverse matrix, which is constructed based on the real-time heading angle of the autonomous vehicle; Represents the semantic potential gradient; This represents the trajectory state of the autonomous vehicle at time t.

[0011] Optionally, processing the target flow prediction result to obtain the trajectory determination result of the autonomous vehicle includes: The target flow prediction result is subjected to iterative integral evolution of ordinary differential equations in the normalized time domain to obtain the trajectory determination result of the autonomous vehicle.

[0012] A second aspect of this application provides an end-to-end autonomous driving trajectory determination device, the device comprising: The environmental context feature vector determination module is used to obtain the environmental context feature vector of autonomous vehicles based on sensor data, map information and navigation targets; The initial flow direction determination module is used to predict the autonomous driving trajectory flow direction based on the current trajectory state information of the autonomous vehicle, the current time, and the environmental context feature vector, and to determine the initial flow direction prediction result. The flow direction correction module is used to correct the initial flow direction prediction result using constraint-corrected flow direction to obtain the target flow direction prediction result; the constraint-corrected flow direction is determined based on physical constraints and safety rule constraints; the physical constraints are constructed based on the Riemann metric inverse matrix; the safety rule constraints are constructed based on the semantic potential field gradient. The autonomous driving trajectory determination module is used to process the target flow prediction results to obtain the trajectory determination results of the autonomous driving vehicle.

[0013] Optionally, the initial flow direction determination module is used to: The flow model is used to predict the flow direction of the autonomous driving trajectory based on the current trajectory state information, current time, and environmental context feature vector of the autonomous vehicle, and the initial flow direction prediction result is determined; the flow model is obtained after training.

[0014] A third aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the end-to-end autonomous driving trajectory determination method provided in any implementation of the first aspect.

[0015] The fourth aspect of this application provides a processor for running a computer program that, when running, executes an end-to-end autonomous driving trajectory determination method as provided in any implementation of the first aspect.

[0016] Compared with the prior art, this application has the following advantages: The end-to-end autonomous driving trajectory determination method provided in this application embodiment obtains an environmental context feature vector of an autonomous vehicle based on sensor data, map information, and navigation targets; predicts the autonomous driving trajectory flow direction based on the current trajectory state information, current time, and environmental context feature vector of the autonomous vehicle to determine an initial flow direction prediction result; corrects the initial flow direction prediction result using constraint-corrected flow direction to obtain a target flow direction prediction result; the constraint-corrected flow direction is determined based on physical constraints and safety rule constraints; the physical constraints are constructed based on the Riemann metric inverse matrix; the safety rule constraints are constructed based on the semantic potential field gradient; and the target flow direction prediction result is processed to obtain the trajectory determination result of the autonomous vehicle.

[0017] Based on the current trajectory state information, current time, and environmental context feature vectors of the autonomous vehicle, the autonomous driving trajectory flow direction is predicted. The initial flow direction prediction result can fully perceive the road structure, obstacle distribution, navigation target, and real-time vehicle motion state, avoiding planning deviations caused by relying on only a single piece of information, thereby improving the accuracy of trajectory output. Physical constraints are constructed based on the Riemann metric inverse matrix, directly embedding the vehicle's nonholonomic kinematic characteristics into the flow direction correction process. This ensures that the trajectory always meets the vehicle's kinematic constraints, avoiding non-physical deformations such as lateral slippage, curvature abrupt changes, or steering over-limits, significantly improving the physical accuracy and execution precision of the trajectory. Safety rule constraints constructed based on semantic potential field gradients transform obstacles and road boundaries into smooth gradient guiding forces, enabling the trajectory to accurately conform to safety boundaries and smoothly avoid obstacles. This avoids jitter, jumps, and intrusion into dangerous areas caused by hard constraints. This continuous guidance through safety rules significantly improves the accuracy of localization and obstacle avoidance in complex scenarios. By uniformly correcting the initial flow direction prediction results through constraint correction, a deep integration of driving intention, physical feasibility and environmental safety is achieved, so that the autonomous driving trajectory determination results not only conform to the driving behavior rules, but also strictly meet the dual constraints of vehicle and environment, thereby improving the overall accuracy of autonomous driving trajectory determination results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of an end-to-end autonomous driving trajectory determination method provided in this application embodiment; Figure 2This is a schematic diagram of the structure of an end-to-end autonomous driving trajectory determination device provided in an embodiment of this application. Detailed Implementation

[0020] As described earlier, current end-to-end autonomous driving trajectory determination methods mostly employ generative planning. However, this approach still has certain shortcomings in complex urban conditions: existing models often fit data in a flat Euclidean space, ignoring the kinematic characteristics of the vehicle as a nonholonomic constrained system, such as the inability to sideslip or curvature continuity, making it prone to instantaneous lateral movement or unnatural curvature abrupt changes. This deficiency makes the planning results heavily reliant on cumbersome post-processing corrections, which not only limits the accuracy of the planning but also reduces the system's real-time response capability in extreme obstacle avoidance scenarios.

[0021] In view of the above problems, this application proposes an end-to-end autonomous driving trajectory determination method and related products. Based on sensor data, map information, and navigation targets, an environmental context feature vector of the autonomous vehicle is obtained. Based on the current trajectory state information, current time, and environmental context feature vector of the autonomous vehicle, the autonomous driving trajectory flow direction is predicted to determine an initial flow direction prediction result. The initial flow direction prediction result is corrected using constraint-corrected flow direction to obtain a target flow direction prediction result. The constraint-corrected flow direction is determined based on physical constraints and safety rule constraints. The physical constraints are constructed based on the Riemann metric inverse matrix; the safety rule constraints are constructed based on the semantic potential field gradient. The target flow direction prediction result is processed to obtain the trajectory determination result of the autonomous vehicle.

[0022] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0023] See Figure 1 This figure is a flowchart of an end-to-end autonomous driving trajectory determination method provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps: S101. Based on sensor data, map information, and navigation targets, obtain the environmental context feature vector of the autonomous vehicle.

[0024] Sensor data is transformed into a high-dimensional semantic space to provide a basis for flow direction prediction. Based on the sensor data, the vehicle's state vector is obtained. Where x represents the lateral position coordinate of the vehicle in the bird's-eye view coordinate system; y represents the longitudinal position coordinate of the vehicle in the bird's-eye view coordinate system; Indicates the heading angle; Indicates instantaneous velocity; Indicates acceleration; This represents the current curvature. Map information is extracted using an encoding network to obtain the static features of the map. With dynamic target characteristics The feature fusion layer performs cross-attention calculation on the vehicle state vector, map static features, and dynamic target features to obtain the environmental context feature vector. : ; in, This indicates cross attention; This represents a multilayer perceptron.

[0025] S102. Based on the current trajectory state information of the autonomous vehicle, the current time, and the environmental context feature vector, perform autonomous driving trajectory flow prediction to determine the initial flow prediction result.

[0026] In one feasible implementation: The flow model is used to predict the flow direction of the autonomous driving trajectory based on the current trajectory state information, current time, and environmental context feature vector of the autonomous vehicle, and the initial flow direction prediction result is determined; the flow model is obtained after training.

[0027] In one feasible implementation, the training method of the flow model includes: By sampling real driving datasets, a set of expert real trajectories is obtained.

[0028] By sampling the standard Gaussian distribution, a set of noise trajectories is obtained.

[0029] A pairing mapping between the expert real trajectory set and the noise trajectory set is established based on the Riemann distance matching mechanism to obtain the path mapping function, and the initial flow field is constructed based on the path mapping function.

[0030] The flow model is subjected to supervised learning based on the initial flow field. The parameters of the flow model are determined by minimizing the conditional flow matching loss function between the predicted flow direction and the target flow direction.

[0031] Extract the set of expert real trajectories corresponding to the current scenario from the real driving dataset. , for target distribution This serves as the final convergence target for trajectory evolution. Simultaneously, it follows a predefined standard Gaussian distribution. Simultaneous sampling yields an equal number of original noise trajectory sets. , is the prior distribution ,in, This is a hyperparameter, which can be set to 64. It is dynamically adjusted based on the redundancy of real-time computing power to balance the diversity of generated trajectories with real-time computation.

[0032] In one feasible implementation, a pairing mapping between the expert true trajectory set and the noise trajectory set is established based on the Riemann distance matching mechanism to obtain a path mapping function, including: Based on the set of expert real trajectories and the set of noisy trajectories, a Riemann cost matrix is ​​constructed.

[0033] Based on the Riemann cost matrix, the path mapping function from noisy samples to expert samples is determined by minimizing the total global matching cost.

[0034] By traversing the noise set Each random trajectory sample in With experts Each real trajectory sample in ,structure Riemann cost matrix Riemann cost matrix Each element in The cost of the Riemann geodesic distance between sample pairs is represented by the following formula: ; in, Represents the Riemannian metric tensor that reflects vehicle non-integrity constraints; The environmental semantic potential function is represented by parameters. Adjusting the weights ensures that the matching pairs traversing high-potential areas, such as obstacles or curbs, generate extremely high cost-effectiveness.

[0035] In obtaining the Riemann cost matrix Then, by minimizing the total global matching cost, the mapping function from noisy samples to expert samples is determined. The formula is as follows: ; in, Represents all possible A set of permutations of order.

[0036] Define the normalized time domain and through path mapping function Construct from noise points Point to the experts The geometric framework for the transfer is given by the following formula: ; This path specifies the trajectory at any instant during its evolution. status And satisfy the boundary conditions and By initializing the ordinary differential equation solution environment, this geometric benchmark is used as the base for flow field evolution, enabling the predicted flow vectors in subsequent steps to drive the sample to undergo precise state transitions along specific geometric manifold directions.

[0037] For example, construct a parameter The core of the flow model The flow model receives the trajectory state at the current moment. Evolutionary time step (Current moment) and environmental context feature vector As input, the conditional flow matching loss function between the predicted flow direction and the target flow direction is minimized. To optimize network parameters, loss function The formula is as follows: ; in, This represents the expectation operator, used to evaluate the evolution time step t and the current trajectory state. Take the expected value; This represents the ideal evolution rate determined by Riemann matching.

[0038] After training, the flow model Capable of generating initial flow direction prediction results under unconstrained conditions. .

[0039] S103. The initial flow direction prediction result is corrected by using constraint correction flow direction to obtain the target flow direction prediction result.

[0040] The constraint correction flow direction is determined based on physical constraints and security rule constraints; the physical constraints are constructed based on the Riemann metric inverse matrix; and the security rule constraints are constructed based on the semantic potential gradient.

[0041] In one feasible implementation, the target flow direction prediction result is obtained according to the following formula: ; in, This indicates the predicted direction of the target flow; This indicates the initial flow direction prediction result; Indicates the weighting coefficient; denoted as the Riemann metric inverse matrix, which is constructed based on the real-time heading angle of the autonomous vehicle; Represents the semantic potential gradient; This represents the trajectory state of the autonomous vehicle at time t.

[0042] Riemannian metric inverse matrix based on vehicle real-time heading angle The projector operator, acting as a non-holonomic dynamic constraint, is constructed to ensure that the correction is strictly limited within the permissible range of vehicle kinematics by projecting obstacle avoidance forces onto the vehicle's longitudinal degrees of freedom, thus prohibiting non-physical lateral translations. Gradient term. It represents the direction of the repulsive force exerted by obstacles in the environment on the current trajectory point, guiding the trajectory towards a low-potential-energy safe area.

[0043] At each time step of trajectory evolution In this process, the driving intuition flow direction shown by the initial flow direction prediction results is corrected in real time through physical constraints and safety rule constraints, so as to ensure that the trajectory always maintains physical feasibility and semantic safety during the evolution process.

[0044] A semantic potential energy field is constructed within the configuration space, with the center point of the obstacle assigned the highest potential energy value. The energy decreases exponentially or quadratically with increasing distance. The level of energy directly characterizes the degree of danger in a spatial region: higher energy indicates a greater risk of collision; a safe region, such as the center of an open lane, corresponds to the lowest energy. To automatically optimize the flow direction towards the safe region, obstacle avoidance correction forces are generated by calculating the negative gradient of the potential field, which is the semantic potential field gradient. The calculation formula is as follows: ; in, express The momentary momentum at the location.

[0045] By utilizing the metric properties of Riemannian geometry, anisotropic motion resistance is assigned to the configuration space. This is achieved through the metric matrix. Define the cost of the vehicle moving in different directions. Furthermore, since the vehicle cannot translate laterally, the lateral resistance is assumed to approach infinity. This is achieved using the inverse of the metric matrix. As a projection operator, it maps external forces from multiple directions into the vehicle's physically feasible motion space in real time, as shown in the following formula: ; in, For extremely small positive numbers, it is set here. ; Represents the rotation transformation matrix; Indicates the heading angle at the current trajectory point; This represents the inverse metric matrix in the local coordinate system. The vertical component 1 indicates that the resistance in the direction of the vehicle's head is minimal, allowing the flow direction to pass completely.

[0046] S104. Process the target flow prediction result to obtain the trajectory determination result of the autonomous vehicle.

[0047] In one feasible implementation: The target flow prediction result is subjected to iterative integral evolution of ordinary differential equations in the normalized time domain to obtain the trajectory determination result of the autonomous vehicle.

[0048] By normalizing the time domain Iterative integration is performed within the initial state, driving the trajectory from the initial state. Evolving to a convergent state The output sequence is a physical safety trajectory that meets obstacle avoidance requirements and conforms to the vehicle chassis motion characteristics.

[0049] The end-to-end autonomous driving trajectory determination method provided in this application embodiment obtains an environmental context feature vector of an autonomous vehicle based on sensor data, map information, and navigation targets; predicts the autonomous driving trajectory flow direction based on the current trajectory state information, current time, and environmental context feature vector of the autonomous vehicle to determine an initial flow direction prediction result; corrects the initial flow direction prediction result using constraint-corrected flow direction to obtain a target flow direction prediction result; the constraint-corrected flow direction is determined based on physical constraints and safety rule constraints; the physical constraints are constructed based on the Riemann metric inverse matrix; the safety rule constraints are constructed based on the semantic potential field gradient; and the target flow direction prediction result is processed to obtain the trajectory determination result of the autonomous vehicle.

[0050] Based on the current trajectory state information, current time, and environmental context feature vectors of the autonomous vehicle, the autonomous driving trajectory flow direction is predicted. The initial flow direction prediction result can fully perceive the road structure, obstacle distribution, navigation target, and real-time vehicle motion state, avoiding planning deviations caused by relying on only a single piece of information, thereby improving the accuracy of trajectory output. Physical constraints are constructed based on the Riemann metric inverse matrix, directly embedding the vehicle's nonholonomic kinematic characteristics into the flow direction correction process. This ensures that the trajectory always meets the vehicle's kinematic constraints, avoiding non-physical deformations such as lateral slippage, curvature abrupt changes, or steering over-limits, significantly improving the physical accuracy and execution precision of the trajectory. Safety rule constraints constructed based on semantic potential field gradients transform obstacles and road boundaries into smooth gradient guiding forces, enabling the trajectory to accurately conform to safety boundaries and smoothly avoid obstacles. This avoids jitter, jumps, and intrusion into dangerous areas caused by hard constraints. This continuous guidance through safety rules significantly improves the accuracy of localization and obstacle avoidance in complex scenarios. By uniformly correcting the initial flow direction prediction results through constraint correction, a deep integration of driving intention, physical feasibility and environmental safety is achieved, so that the autonomous driving trajectory determination results not only conform to the driving behavior rules, but also strictly meet the dual constraints of vehicle and environment, thereby improving the overall accuracy of autonomous driving trajectory determination results.

[0051] Based on the end-to-end autonomous driving trajectory determination method described in the preceding embodiments, this application also provides an end-to-end autonomous driving trajectory determination device. Figure 2 This is a schematic diagram of the device. Figure 2 As shown, the end-to-end autonomous driving trajectory determination device includes: The environmental context feature vector determination module 201 is used to obtain the environmental context feature vector of the autonomous vehicle based on sensor data, map information and navigation target.

[0052] The initial flow direction determination module 202 is used to predict the flow direction of the autonomous driving trajectory based on the current trajectory state information of the autonomous vehicle, the current time, and the environmental context feature vector, and to determine the initial flow direction prediction result.

[0053] The flow direction correction module 203 is used to correct the initial flow direction prediction result by using constraint correction flow direction to obtain the target flow direction prediction result; the constraint correction flow direction is determined based on physical constraints and safety rule constraints; the physical constraints are constructed based on the Riemann metric inverse matrix; the safety rule constraints are constructed based on the semantic potential field gradient.

[0054] The autonomous driving trajectory determination module 204 is used to process the target flow prediction result to obtain the trajectory determination result of the autonomous driving vehicle.

[0055] Optionally, the initial flow direction determination module 202 is used to: The flow model is used to predict the flow direction of the autonomous driving trajectory based on the current trajectory state information, current time, and environmental context feature vector of the autonomous vehicle, and the initial flow direction prediction result is determined; the flow model is obtained after training.

[0056] Optionally, the training methods for the streaming model include: By sampling real driving datasets, a set of expert real trajectories is obtained.

[0057] By sampling the standard Gaussian distribution, a set of noise trajectories is obtained.

[0058] A pairing mapping between the expert real trajectory set and the noise trajectory set is established based on the Riemann distance matching mechanism to obtain the path mapping function, and the initial flow field is constructed based on the path mapping function.

[0059] The flow model is subjected to supervised learning based on the initial flow field. The parameters of the flow model are determined by minimizing the conditional flow matching loss function between the predicted flow direction and the target flow direction.

[0060] Optionally, a pairing mapping between the expert's true trajectory set and the noisy trajectory set is established based on the Riemann distance matching mechanism to obtain a path mapping function, including: Based on the set of expert real trajectories and the set of noisy trajectories, a Riemann cost matrix is ​​constructed.

[0061] Based on the Riemann cost matrix, the path mapping function from noisy samples to expert samples is determined by minimizing the total global matching cost.

[0062] Optionally, the flow direction correction module 203 is used for: The target flow direction prediction result is obtained based on the following formula: ; in, This indicates the predicted direction of the target flow; This indicates the initial flow direction prediction result; Indicates the weighting coefficient; denoted as the Riemann metric inverse matrix, which is constructed based on the real-time heading angle of the autonomous vehicle; Represents the semantic potential gradient; This represents the trajectory state of the autonomous vehicle at time t.

[0063] Optionally, the autonomous driving trajectory determination module 204 is used for: The target flow prediction result is subjected to iterative integral evolution of ordinary differential equations in the normalized time domain to obtain the trajectory determination result of the autonomous vehicle.

[0064] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the end-to-end autonomous driving trajectory determination method as described in any of the method embodiments.

[0065] Furthermore, this application embodiment also provides a processor for running a computer program, which executes the end-to-end autonomous driving trajectory determination method as described in any of the foregoing method embodiments.

[0066] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separate. The components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment solution according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0067] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for determining the trajectory of end-to-end automated driving, characterized in that, include: Based on sensor data, map information, and navigation targets, the environmental context feature vector of autonomous vehicles is obtained; Based on the current trajectory state information, current time, and environmental context feature vector of the autonomous vehicle, the autonomous driving trajectory flow direction is predicted to determine the initial flow direction prediction result; The initial flow direction prediction result is corrected by using constraint-corrected flow direction to obtain the target flow direction prediction result; the constraint-corrected flow direction is determined based on physical constraints and safety rule constraints; the physical constraints are constructed based on the Riemann metric inverse matrix; the safety rule constraints are constructed based on the semantic potential gradient. The target flow prediction results are processed to obtain the trajectory determination results of the autonomous vehicle.

2. The method according to claim 1, characterized in that, The process of predicting the autonomous driving trajectory flow direction based on the current trajectory state information of the autonomous vehicle, the current time, and the environmental context feature vector, and determining the initial flow direction prediction result, includes: The flow model is used to predict the flow direction of the autonomous driving trajectory based on the current trajectory state information, current time, and environmental context feature vector of the autonomous vehicle, and the initial flow direction prediction result is determined; the flow model is obtained after training.

3. The method according to claim 2, characterized in that, The training methods for the stream model include: By sampling real driving datasets, a set of expert real trajectories is obtained; By sampling from a standard Gaussian distribution, a set of noise trajectories is obtained; A pairing mapping between the expert real trajectory set and the noise trajectory set is established based on the Riemann distance matching mechanism to obtain the path mapping function, and an initial flow field is constructed based on the path mapping function; The flow model is subjected to supervised learning based on the initial flow field. The parameters of the flow model are determined by minimizing the conditional flow matching loss function between the predicted flow direction and the target flow direction.

4. The method according to claim 3, characterized in that, The method for establishing a pairing mapping between the expert's true trajectory set and the noise trajectory set based on the Riemann distance matching mechanism, resulting in a path mapping function, includes: Based on the set of expert real trajectories and the set of noisy trajectories, a Riemann cost matrix is ​​constructed; Based on the Riemann cost matrix, the path mapping function from noisy samples to expert samples is determined by minimizing the total global matching cost.

5. The method according to claim 1, characterized in that, The step of correcting the initial flow direction prediction result using constraint-corrected flow direction to obtain the target flow direction prediction result includes: The target flow direction prediction result is obtained based on the following formula: ； in, This indicates the predicted direction of the target flow; This indicates the initial flow direction prediction result; Indicates the weighting coefficient; denoted as the Riemann metric inverse matrix, which is constructed based on the real-time heading angle of the autonomous vehicle; Represents the semantic potential gradient; This represents the trajectory state of the autonomous vehicle at time t.

6. The method according to claim 1, characterized in that, The process of processing the target flow prediction result to obtain the trajectory determination result of the autonomous vehicle includes: The target flow prediction result is subjected to iterative integral evolution of ordinary differential equations in the normalized time domain to obtain the trajectory determination result of the autonomous vehicle.

7. An end-to-end autonomous driving trajectory determination device, characterized in that, include: The environmental context feature vector determination module is used to obtain the environmental context feature vector of autonomous vehicles based on sensor data, map information and navigation targets; The initial flow direction determination module is used to predict the autonomous driving trajectory flow direction based on the current trajectory state information of the autonomous vehicle, the current time, and the environmental context feature vector, and to determine the initial flow direction prediction result. The flow direction correction module is used to correct the initial flow direction prediction result using constraint-corrected flow direction to obtain the target flow direction prediction result; the constraint-corrected flow direction is determined based on physical constraints and safety rule constraints; the physical constraints are constructed based on the Riemann metric inverse matrix; the safety rule constraints are constructed based on the semantic potential field gradient. An autonomous driving trajectory determination module is used to process the target flow prediction results to obtain the trajectory determination results of the autonomous driving vehicle.

8. The apparatus according to claim 7, characterized in that, The initial flow direction determination module is used for: The flow model is used to predict the flow direction of the autonomous driving trajectory based on the current trajectory state information, current time, and environmental context feature vector of the autonomous vehicle, and the initial flow direction prediction result is determined; the flow model is obtained after training.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the end-to-end autonomous driving trajectory determination method as described in any one of claims 1-6.

10. A processor, characterized in that, Used to run a computer program, which, when running, performs the end-to-end autonomous driving trajectory determination method as described in any one of claims 1-6.

Images