Automatic driving double model hierarchical fusion safety decision method, system, device and medium
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEFEI INSTITUTE OF PHYSICAL SCIENCE CHINESE ACADEMY OF SCIENCES
- Filing Date
- 2026-02-06
- Publication Date
- 2026-06-02
Smart Images
Figure CN122126307A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous driving technology, and in particular to a method, system, device and medium for autonomous driving dual-model hierarchical fusion safety decision-making. Background Technology
[0002] The decision-making framework of autonomous driving systems mainly follows two technical approaches: modular methods based on explicit rules and data-driven end-to-end learning methods. Rule-based methods rely on predefined traffic rules, geometric constraints, and optimization models, offering advantages such as transparent decision-making processes, logical interpretability, and ease of safety verification. However, their performance heavily depends on the completeness of the rule base, and their generalization ability is limited in complex, dynamic, or long-tailed scenarios, making it difficult to cover all driving situations. End-to-end learning methods directly map raw sensor data to control commands through deep neural networks, enabling them to learn complex driving patterns and scenario features from large-scale data. They exhibit strong performance potential and generalization ability within known data distributions. However, their decision-making process is "black box" in nature, lacking interpretability and formal safety guarantees. When sensors fail, data distribution shifts, or model confidence is insufficient, unpredictable and unsafe behaviors may occur.
[0003] To balance performance and security, existing technologies attempt to introduce redundancy or master-slave switching architectures, such as running multiple models in parallel or switching to a backup model when the master model is detected to be faulty. However, such solutions mostly remain at the level of simple comparison or switching at the system output level, lacking effective mechanisms for fine-grained, structured alignment, evaluation, and fusion of the outputs of data-driven and rule-driven models at multiple key decision levels such as "perception, prediction, and planning." Therefore, existing solutions struggle to achieve complementary advantages between the two technical approaches, cannot dynamically select the most reliable or optimal intermediate result in complex scenarios, and lack a system safety fallback mechanism to ensure that the system can still perform the lowest-risk operation when both models face uncertainty or failure.
[0004] Therefore, designing a fusion framework in autonomous driving systems that can dynamically integrate data-driven capabilities and rule-driven verifiability at multiple decision levels, and has a fail-safe fallback mechanism, has become a key challenge in improving the overall safety, reliability, and scenario adaptability of autonomous driving systems. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the existing technology. To achieve the above objectives, an autonomous driving dual-model hierarchical fusion safety decision-making method, system, device, and medium are adopted to solve the problems mentioned in the background technology.
[0006] A dual-model hierarchical fusion safety decision-making method for autonomous driving includes the following steps:
[0007] Simultaneously acquire and process camera images, LiDAR point cloud data, and vehicle status interface input information to form multimodal sensor data; The multimodal sensor data is input in parallel into a dual model consisting of a data-driven end-to-end learning model and a rule-driven model based on explicit rules and geometric constraints. The two models perform inference independently and output their respective perception results, prediction results, and planning results. In the perception layer, prediction layer and planning layer of the dual model, the outputs of the end-to-end learning model and the rule-driven model are dynamically compared, evaluated and fused to generate the final execution trajectory. The fusion is based on verifiable geometric rationality, multidimensional safety score or trajectory safety assessment. A safety assessment mechanism independent of the dual models is introduced to perform real-time safety judgment on the final execution trajectory. When it is determined that neither of the dual model outputs meets the preset safety conditions, emergency braking or the minimum risk strategy is triggered and a safe trajectory is output; otherwise, the vehicle control command corresponding to the final execution trajectory is output.
[0008] As a further aspect of the present invention: the vehicle status interface input information in step S1 includes one or more of vehicle positioning information, inertial measurement information and / or high-precision map prior information.
[0009] As a further aspect of the present invention: the end-to-end learning model in step S2 adopts a sparse bird's-eye view feature construction method based on an attention mechanism, performing feature transformation only on key regions to reduce computational load, as expressed in the following expression:
[0010] in, For view transformation functions, Represents the characteristic coordinates, Representing the characteristics of integration, It's an attention score. For the threshold, This represents the number of candidate features. The perception, prediction, and planning results are output synchronously through a multi-task head, expressed as follows:
[0011] in, This indicates a multi-task header used for multi-task decoupling processing of the sparse bird's-eye view features and generating corresponding task outputs, wherein... This represents the perceptual result output by the end-to-end learning model. This represents the prediction result output by the end-to-end learning model. This represents the planning result output by the end-to-end learning model.
[0012] As a further aspect of the present invention: in step S2, the rule-driven model sequentially executes sensing based on sensor data, prediction based on a physical motion model, and planning based on an optimization algorithm. The prediction task uses a constant velocity model or a constant acceleration model to extrapolate the dynamic target trajectory, expressed as:
[0013] in, For the first The current location of each target. For velocity vectors, For predicting time.
[0014] As a further aspect of the present invention: the fusion of the sensing layer in step S3 specifically includes: First, establish the correspondence between targets in the perception results of the two models. Then, perform geometric verification on the target detection results output by the end-to-end learning model, including size, position and motion state. If the verification is successful and the confidence level is higher than the first threshold, the end-to-end result is adopted. Otherwise, the result of the rule-driven model is adopted. The Hungarian algorithm or the nearest neighbor matching algorithm is used to establish the target correspondence. The fusion of the prediction layer specifically includes: A multi-dimensional evaluation function is constructed to calculate the scores of the predicted trajectories output by the two models in terms of road structure compliance, traffic rule compliance, and kinematic rationality. The scores are then weighted and summed to obtain a comprehensive score. The predicted trajectory with the higher comprehensive score is selected, or the two trajectories are weighted and fused. The expression is as follows:
[0015] in, This represents the evaluation of the predicted trajectory against road constraints. Indicates an assessment of traffic rules. This represents the evaluation of physical constraint factors. 、 、 These are the weighting coefficients; The integration at the planning level specifically includes: The collision risk value, safe distance from obstacles, trajectory smoothness, and deviation from the global path of the planned trajectories generated by the two models are calculated respectively, and the final trajectory is selected according to the preset safety priority rules.
[0016] As a further aspect of the present invention: the security assessment mechanism in step S4 is as follows: It is implemented based on a joint training framework, whose loss function includes the main task loss of the end-to-end model, the fusion loss to promote consistency of the outputs of the two models, and the safety rule loss to strengthen safety constraints.
[0017] As a further aspect of the present invention: the joint training framework for the loss function is expressed as follows:
[0018] in, For end-to-end learning model loss, Used to optimize fusion consistency or fusion accuracy. Used to constrain security compliance 、 These are the weighting coefficients.
[0019] The second aspect of the technical solution: A hierarchical fusion system employing an autonomous driving dual-model hierarchical fusion safety decision-making method as described in any of the above, comprising: The multimodal perception module is used to simultaneously process camera images, LiDAR point cloud data, and vehicle status interface input; A dual-track decision module is used to run the end-to-end learning model and the rule-driven model in parallel; The layered fusion module is used to perform dynamic selection and fusion logic of the dual model outputs at the three levels of perception, prediction, and planning. The safety monitoring and fallback module is used to implement a safety assessment mechanism and trigger emergency control commands when necessary.
[0020] The third aspect of the technical solution: A device comprising: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method as described in any one of the above statements.
[0021] The fourth aspect of the technical solution: a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method as described in any one of the above.
[0022] Compared with the prior art, the present invention has the following technical advantages: The above technical solution involves simultaneously processing camera, LiDAR, and vehicle status information through a multimodal perception step to form a unified perception input. Subsequently, in the dual-track decision-making step, data is input in parallel to a data-driven end-to-end learning model and a rule-driven model based on explicit rules and geometric constraints. Both models independently reason and output their respective perception, prediction, and planning results. Next, in the hierarchical fusion step, the outputs of the two models are dynamically compared and fused at the perception, prediction, and planning levels, based on geometric rationality verification, multi-dimensional safety scoring, and trajectory safety assessment, respectively, to generate the final execution trajectory. Finally, through an independent safety fallback step, a real-time safety assessment mechanism is introduced to trigger a preset emergency strategy when both models are unreliable, ensuring the system has fail-safe capabilities. Attached Figure Description
[0023] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings: Figure 1 This is a schematic diagram illustrating the steps of the hierarchical fusion security decision-making method according to an embodiment of this application; Figure 2 This is a flowchart illustrating the hierarchical fusion security decision-making method according to an embodiment of this application. Detailed Implementation
[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] Please refer to Figure 1 and Figure 2 In this embodiment of the invention, an autonomous driving dual-model hierarchical fusion safety decision-making method includes the following steps: Step S1, Multimodal Perception: Simultaneously acquire and process camera images, LiDAR point cloud data, and vehicle status interface input information to form multimodal sensor data; In this embodiment, the vehicle status interface input information in step S1 includes one or more of vehicle positioning information, inertial measurement information, and / or high-precision map prior information.
[0026] Step S11, Time Synchronization and Alignment: Timestamp alignment of data from the camera, LiDAR, and vehicle status interface; when there are differences in sampling frequency, interpolation / buffering queues can be used to align vehicle status information to the perception frame time.
[0027] Step S12, Spatial calibration and coordinate unification: Unify the point cloud and image features to the vehicle coordinate system or map coordinate system (e.g., use the extrinsic calibration matrix and pose information to complete the coordinate transformation); optionally, combine the high-precision map prior to complete the geometric constraint extraction of lane / road boundaries.
[0028] Step S13, Multimodal Preprocessing and Feature Construction: Distortion correction and scale normalization are performed on the image; noise reduction / ground segmentation / voxelization or columnar coding are performed on the point cloud; vehicle state information is normalized and encoded to form a feature set for subsequent fusion (which can be referred to as fusion features). This is used for constructing sparse BEV features in subsequent end-to-end models.
[0029] Step S2, Dual-track decision-making: The multimodal sensor data is input in parallel into a dual model, which is a data-driven end-to-end learning model and a rule-driven model based on explicit rules and geometric constraints. The two models perform inference independently and output their respective perception results, prediction results and planning results. In this embodiment, the end-to-end learning model in step S2 adopts a sparse bird's-eye view feature construction method based on an attention mechanism, and only performs feature transformation on key regions to reduce computational load. The expression is:
[0030] in, For view transformation functions, Represents the characteristic coordinates, Representing the characteristics of integration, It's an attention score. For the threshold, This represents the number of candidate features. The perception, prediction, and planning results are output synchronously through a multi-task head, expressed as follows:
[0031] in, This indicates a multi-task header used for multi-task decoupling processing of the sparse bird's-eye view features and generating corresponding task outputs, wherein... This represents the perceptual result output by the end-to-end learning model. This represents the prediction result output by the end-to-end learning model. This represents the planning result output by the end-to-end learning model.
[0032] In this embodiment, in step S2, the rule-driven model sequentially executes tasks based on sensor data perception, physical motion model prediction, and optimization algorithm planning. The prediction task uses a constant velocity model or a constant acceleration model to extrapolate the dynamic target trajectory, expressed as:
[0033] in, For the first The current location of each target. For velocity vectors, For predicting time.
[0034] Specifically, the dual-track decision-making process is implemented by two parallel and independent reasoning paths: The first path is end-to-end learning model inference. The multimodal sensor data generated in step S1 (such as image features, point cloud voxel features, and vehicle state codes) is input into an end-to-end model built on a deep neural network. This model first generates unified fused features through a cross-modal attention fusion module. Subsequently, the model employs a sparse bird's-eye view (BEV) encoding strategy: calculating the attention score for each preset grid cell in the BEV space. The view transformation function is applied only to the K key grids whose scores exceed a preset threshold, thereby obtaining sparse BEV features. This design significantly reduces computational complexity. Finally, the sparse BEV feature is fed into a multi-task head, which includes a parallel perception branch (outputting 3D bounding boxes of the target), a prediction branch (outputting the probabilities of future trajectories of multiple targets), and a planning branch (outputting candidate trajectories and scores), achieving synchronous output.
[0035] The second path is rule-driven model inference. This model strictly follows a pre-defined, interpretable chain of algorithm modules. First, in the perception module, target detection and tracking are performed on the LiDAR point cloud based on geometric clustering, and semantic segmentation of the image is performed based on traditional computer vision or lightweight neural networks. The results of both are fused using Kalman filtering to output a target list. Second, in the prediction module, kinematic extrapolation is performed on each tracked target using a constant velocity (CV) model or a constant acceleration (CA) model. Finally, in the planning module, based on the drivable area provided by the high-precision map, traffic rules (such as traffic lights and stop lines), and the perception / prediction results, a collision-free, smooth trajectory from the current location to the local navigation target point is searched using algorithms such as A, HybridA, or an optimizer (such as iLQR).
[0036] Step S3, Layered Selection and Fusion Mechanism: In the perception layer, prediction layer and planning layer of the dual model, the outputs of the end-to-end learning model and the rule-driven model are dynamically compared, evaluated and fused to generate the final execution trajectory. The fusion is based on verifiable geometric rationality, multidimensional safety score or trajectory safety assessment. Step S31, the fusion of the perception layer specifically includes: First, establish the correspondence between targets in the perception results of the two models. Then, perform geometric verification on the target detection results output by the end-to-end learning model, including size, position and motion state. If the verification is successful and the confidence level is higher than the first threshold, the end-to-end result is adopted. Otherwise, the result of the rule-driven model is adopted. The Hungarian algorithm or the nearest neighbor matching algorithm is used to establish the target correspondence. Step S32, the fusion of the prediction layer specifically includes: A multi-dimensional evaluation function is constructed to calculate the scores of the predicted trajectories output by the two models in terms of road structure compliance, traffic rule compliance, and kinematic rationality. The scores are then weighted and summed to obtain a comprehensive score. The predicted trajectory with the higher comprehensive score is selected, or the two trajectories are weighted and fused. The expression is as follows:
[0037] in, This represents the evaluation of the predicted trajectory against road constraints. Indicates an assessment of traffic rules. This represents the evaluation of physical constraint factors. 、 、 These are the weighting coefficients; Step S33, the integration of the planning layer specifically includes: The collision risk value, safe distance from obstacles, trajectory smoothness, and deviation from the global path of the planned trajectories generated by the two models are calculated respectively, and the final trajectory is selected according to the preset safety priority rules.
[0038] Specifically, the layered integration mechanism is implemented at three levels as follows: 1. Perceptual Layer Fusion: First, using the Hungarian algorithm, an optimal matching relationship is established between the end-to-end model detection set and the rule-based model detection set based on the position and motion state (e.g., velocity, heading) of the target boxes. For each matched target pair, the geometric rationality of the detection results output by the end-to-end model is verified, including: (a) Size rationality: Check whether its 3D size conforms to the prior range of common vehicles, pedestrians, etc.; (b) Location validity: Check whether its location is within the drivable area; (c) Motion Continuity: Check whether the instantaneous velocity and acceleration are within physically feasible thresholds. If all geometric verifications pass and the target confidence given by the end-to-end model is higher than the preset threshold, the result of the end-to-end model is adopted as the final perceived state of the target; otherwise, the result of the rule model is adopted. Unmatched targets will directly retain the result of their source model.
[0039] 2. Prediction Layer Fusion: For each dynamic target, multiple possible trajectories (and their probabilities) predicted by the end-to-end model and a single trajectory predicted by the rule-based model are obtained separately. Two (or more) trajectories are evaluated using a multi-dimensional evaluation function. This function calculates a comprehensive score for each candidate trajectory T. Finally, the trajectory with the highest comprehensive score is selected as the final prediction result for that target, or a fused trajectory is obtained by interpolation between two high-scoring trajectories.
[0040] 3. Planning Layer Fusion: Obtain multiple candidate trajectories and their scores from the end-to-end model planning branch output, as well as the optimal trajectory output by the rule-driven model. Perform a quantitative safety assessment on each candidate trajectory, with assessment metrics including at least: (a) Collision risk: Based on the prediction results, calculate the minimum distance and probability of collision between the trajectory and all dynamic / static obstacles in the next few seconds; (b) Comfort: Calculate the longitudinal acceleration, lateral acceleration, and jerk of the trajectory; (c) Navigation target achievement: Calculate the deviation distance between the trajectory endpoint and the local navigation target point.
[0041] Establish a safety-first decision rule: first, exclude any trajectory with a collision risk exceeding the absolute safety threshold; among the remaining safe trajectories, select the trajectory that best balances comfort and navigation success as the final execution trajectory. Generally, when the rule-based model trajectory is safe and reasonable, it is adopted first to enhance interpretability; however, when the end-to-end model provides a significantly better (e.g., more efficient, more human-like) and safer solution, the end-to-end result is adopted.
[0042] Step S4, Safety Backup: Introduce a safety assessment mechanism independent of the dual models to perform real-time safety judgment on the final execution trajectory. When it is determined that neither of the dual model outputs meets the preset safety conditions, trigger emergency braking or the minimum risk strategy and output a safe trajectory; otherwise, output the vehicle control command corresponding to the final execution trajectory.
[0043] In this embodiment, the security assessment mechanism in step S4 is as follows: It is implemented based on a joint training framework, whose loss function includes the main task loss of the end-to-end model, the fusion loss to promote consistency of the outputs of the two models, and the safety rule loss to strengthen safety constraints.
[0044] In this embodiment, the joint training framework for the loss function is expressed as follows:
[0045] in, For end-to-end learning model loss, Used to optimize fusion consistency or fusion accuracy. Used to constrain security compliance 、 These are the weighting coefficients.
[0046] Specifically, the safety fallback step operates as an independent and highest-priority monitoring layer, and its implementation includes a real-time trajectory safety checker and a pre-set emergency strategy library.
[0047] The core of the real-time trajectory safety checker is a lightweight, physics- and rule-based safety assessment model. It performs rapid simulation verification of the "final execution trajectory" generated in step S3 every control cycle (e.g., 10ms). The verification criteria are more stringent than those in step S3, and mainly include: (1) Whether the trajectory encroaches on any absolutely prohibited area (such as the oncoming lane or curb) at any point; (2) Whether the instantaneous distance to the nearest obstacle is less than the emergency braking distance (considering the system's maximum deceleration); (3) Whether the vehicle's yaw rate or lateral acceleration is close to the instability threshold.
[0048] This safety assessment mechanism is reinforced during the model development phase through a joint training framework. A total loss function for the end-to-end learning model is defined, which injects safety prior knowledge into the end-to-end model by encoding safety rules (such as maintaining a safe distance and adhering to speed limits) as differentiable loss functions and penalizing unsafe behaviors generated by the model during training.
[0049] When the security checker determines that the current "final execution trajectory" does not meet any of the above preset security conditions, it considers the dual-model fusion decision unreliable and immediately triggers a safety fallback. The system will ignore the dual-model output and instead select and execute a preset policy from the emergency policy library. This policy library includes at least: (a) Emergency braking strategy: Generate a longitudinal deceleration trajectory to stop according to the maximum comfortable deceleration or the maximum deceleration allowed by the system; (b) Minimum Risk Response (MRC): Near the current location, the system identifies the safest roadside parking area using a high-precision map and generates a trajectory leading to and stopping at that area. After the safe trajectory is generated, the system directly outputs the corresponding vehicle control commands (accelerator, brake, steering) to ensure the vehicle enters a controllable and safe state.
[0050] The second aspect of the technical solution: A hierarchical fusion system employing an autonomous driving dual-model hierarchical fusion safety decision-making method as described in any of the above, comprising: The multimodal perception module is used to simultaneously process camera images, LiDAR point cloud data, and vehicle status interface input; A dual-track decision module is used to run the end-to-end learning model and the rule-driven model in parallel; The layered fusion module is used to perform dynamic selection and fusion logic of the dual model outputs at the three levels of perception, prediction, and planning. The safety monitoring and fallback module is used to implement a safety assessment mechanism and trigger emergency control commands when necessary.
[0051] The third aspect of the technical solution: A device comprising: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor performs the method as described in any one of the above statements.
[0052] The fourth aspect of the technical solution: a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method as described in any one of the above.
[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention. The scope of the invention is defined by the appended claims and their equivalents, all of which should be included within the scope of protection of the invention.
Claims
1. A dual-model hierarchical fusion safety decision-making method for autonomous driving, characterized in that, Includes the following steps: Simultaneously acquire and process camera images, LiDAR point cloud data, and vehicle status interface input information to form multimodal sensor data; The multimodal sensor data is input in parallel into a dual model consisting of a data-driven end-to-end learning model and a rule-driven model based on explicit rules and geometric constraints. The two models perform inference independently and output their respective perception results, prediction results, and planning results. In the perception layer, prediction layer and planning layer of the dual model, the outputs of the end-to-end learning model and the rule-driven model are dynamically compared, evaluated and fused to generate the final execution trajectory. The fusion is based on verifiable geometric rationality, multidimensional safety score or trajectory safety assessment. A safety assessment mechanism independent of the dual models is introduced to perform real-time safety judgment on the final execution trajectory. When it is determined that neither of the dual model outputs meets the preset safety conditions, emergency braking or the minimum risk strategy is triggered and a safe trajectory is output; otherwise, the vehicle control command corresponding to the final execution trajectory is output.
2. The autonomous driving dual-model hierarchical fusion safety decision-making method according to claim 1, characterized in that, The vehicle status interface input information in step S1 includes one or more of the following: vehicle positioning information, inertial measurement information, and / or high-precision map prior information.
3. The autonomous driving dual-model hierarchical fusion safety decision-making method according to claim 1, characterized in that, In step S2, the end-to-end learning model employs an attention-based sparse bird's-eye view feature construction method, performing feature transformation only on key regions to reduce computational load. The expression is as follows: in, For view transformation functions, Represents the characteristic coordinates, Representing the characteristics of integration, It's an attention score. For the threshold, This represents the number of candidate features. The perception, prediction, and planning results are output synchronously through a multi-task head, expressed as follows: in, This indicates a multi-task header used for multi-task decoupling processing of the sparse bird's-eye view features and generating corresponding task outputs, wherein... This represents the perceptual result output by the end-to-end learning model. This represents the prediction result output by the end-to-end learning model. This represents the planning result output by the end-to-end learning model.
4. The autonomous driving dual-model hierarchical fusion safety decision-making method according to claim 1, characterized in that, In step S2, the rule-driven model sequentially executes tasks based on sensor data perception, physical motion model prediction, and optimization algorithm planning. The prediction task uses a constant velocity model or a constant acceleration model to extrapolate the dynamic target trajectory, expressed as: in, For the first The current location of each target. For velocity vectors, For predicting time.
5. The autonomous driving dual-model hierarchical fusion safety decision-making method according to claim 1, characterized in that, The fusion of the perception layer in step S3 specifically includes: First, establish the correspondence between targets in the perception results of the two models. Then, perform geometric verification on the target detection results output by the end-to-end learning model, including size, position and motion state. If the verification is successful and the confidence level is higher than the first threshold, the end-to-end result is adopted. Otherwise, the result of the rule-driven model is adopted. The Hungarian algorithm or the nearest neighbor matching algorithm is used to establish the target correspondence. The fusion of the prediction layer specifically includes: A multi-dimensional evaluation function is constructed to calculate the scores of the predicted trajectories output by the two models in terms of road structure compliance, traffic rule compliance, and kinematic rationality. The scores are then weighted and summed to obtain a comprehensive score. The predicted trajectory with the higher comprehensive score is selected, or the two trajectories are weighted and fused. The expression is as follows: in, This represents the evaluation of the predicted trajectory against road constraints. Indicates an assessment of traffic rules. This represents the evaluation of physical constraint factors. 、 、 These are the weighting coefficients; The integration at the planning level specifically includes: The collision risk value, safe distance from obstacles, trajectory smoothness, and deviation from the global path of the planned trajectories generated by the two models are calculated respectively, and the final trajectory is selected according to the preset safety priority rules.
6. The autonomous driving dual-model hierarchical fusion safety decision-making method according to claim 1, characterized in that, The security assessment mechanism in step S4 is as follows: It is implemented based on a joint training framework, whose loss function includes the main task loss of the end-to-end model, the fusion loss to promote consistency of the outputs of the two models, and the safety rule loss to strengthen safety constraints.
7. The autonomous driving dual-model hierarchical fusion safety decision-making method according to claim 6, characterized in that, The joint training framework for the loss function is expressed as follows: in, For end-to-end learning model loss, Used to optimize fusion consistency or fusion accuracy. Used to constrain security compliance 、 These are the weighting coefficients.
8. A hierarchical fusion system employing the dual-model hierarchical fusion safety decision-making method for autonomous driving as described in any one of claims 1 to 7, characterized in that, include: The multimodal perception module is used to simultaneously process camera images, LiDAR point cloud data, and vehicle status interface input; A dual-track decision module is used to run the end-to-end learning model and the rule-driven model in parallel; The layered fusion module is used to perform dynamic selection and fusion logic of the dual model outputs at the three levels of perception, prediction, and planning. The safety monitoring and fallback module is used to implement a safety assessment mechanism and trigger emergency control commands when necessary.
9. A device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method as described in any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.