Method, device and equipment for determining automatic driving decision, medium and product
Through federated learning and quantum key encryption technology, combined with neural structure search, the problem of privacy protection of autonomous driving systems in the process of intelligent development is solved, and efficient and intelligent autonomous driving decisions are achieved.
Patent Information
- Application Number
- CN202510213673.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-26
- Publication Date
- 2025-08-01
AI Technical Summary
In the process of developing efficient and intelligent autonomous driving systems, how to protect user privacy and prevent personal information from being leaked.
Using a federated learning method, the model is trained locally and updated and tuned in the central server, combining quantum key encryption and neural structure search technology to protect data privacy and improve model performance.
While protecting user privacy, the decision-making accuracy and adaptability of the autonomous driving system are improved to ensure the efficient operation of the model in different environments.
Smart Images
Figure CN120396995A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of autonomous driving technology, and in particular, to a method, apparatus, device, medium, and product for determining autonomous driving decisions. Background Art
[0002] Autonomous vehicles use various sensors, such as cameras, millimeter-wave radars, lidars, etc., to collect data on the surrounding environment, which is called sensor training data. This data includes not only the vehicle's immediate position and speed, but also information about surrounding objects, pedestrians, road signs, and traffic conditions. By analyzing this data, the autonomous driving system can make judgments and decisions, such as when to accelerate, decelerate, turn, or avoid obstacles. This process requires the system to accurately understand its surrounding environment and make quick responses, which requires continuous training and optimization of algorithms through a large amount of actual driving data. During the process of collecting, processing, and analyzing this data, the autonomous driving system may come into contact with a large amount of personal privacy information. For example, the vehicle's driving route may reveal the user's residence, workplace, and even social habits. If this information is obtained by an unauthorized third party, it may cause serious privacy leakage problems. Therefore, how to protect the privacy of users while developing efficient and intelligent autonomous driving technology has become a key issue that the industry and researchers need to solve. Summary of the Invention
[0003] The purpose of this application is to provide a method, apparatus, device, medium, and product for determining autonomous driving decisions, which can protect the privacy of users while developing efficient and intelligent autonomous driving technology.
[0004] To achieve the above purpose, this application provides the following solutions:
[0005] In a first aspect, this application provides a method for determining autonomous driving decisions, including:
[0006] At the current iteration number, for any autonomous vehicle, local training is performed on the optimal global model at the current iteration number according to the sensor training data of the autonomous vehicle at the current iteration number to obtain the local model of the autonomous vehicle at the current iteration number, and the local model parameters of the autonomous vehicle at the current iteration number are sent to the central server;
[0007] The central server updates the optimal global model at the current iteration number according to the local model parameters of each autonomous vehicle at the current iteration number to obtain the global model parameters at the next iteration number, and sends the global model at the next iteration number to each autonomous vehicle;
[0008] Based on the running results of the global model on the sensor verification data of each autonomous driving vehicle in the next iteration, optimize the global model parameters in the next iteration to obtain the optimal global model in the next iteration, send the optimal global model in the next iteration to each autonomous driving vehicle, then update the iteration number, enter the next iteration, and stop until the iteration stop condition is reached. Generate autonomous driving decisions according to the optimal global model in the last iteration number.
[0009] Optionally, send the local model parameters of the autonomous driving vehicle in the current iteration number to the central server, specifically including:
[0010] Encrypt the local model parameters of the autonomous driving vehicle in the current iteration number to obtain the encrypted local model parameters of the autonomous driving vehicle in the current iteration number;
[0011] Send the encrypted local model parameters of the autonomous driving vehicle in the current iteration number to the central server.
[0012] Optionally, the central server updates the optimal global model in the current iteration number according to the local model parameters of each autonomous driving vehicle in the current iteration number to obtain the global model parameters in the next iteration number, specifically including:
[0013] The central server decrypts the encrypted local model parameters of the autonomous driving vehicle in the current iteration number to obtain the decrypted local model parameters of the autonomous driving vehicle in the current iteration number;
[0014] Update the optimal global model in the current iteration number according to the decrypted local model parameters of each autonomous driving vehicle in the current iteration number to obtain the global model parameters in the next iteration number.
[0015] Optionally, the step of optimizing the global model parameters in the next iteration number based on the running results of the global model on the sensor verification data of each autonomous driving vehicle in the next iteration number to obtain the optimal global model in the next iteration number specifically includes:
[0016] Based on the running results of the global model on the sensor verification data of each autonomous driving vehicle in the next iteration number, determine the performance index values of the global model in the next iteration number under the verification data of each autonomous driving vehicle;
[0017] Determine the performance index value of the global model in the next iteration number according to the performance index values of the global model in the next iteration number under the verification data of each autonomous driving vehicle;
[0018] Taking the numerical values of the performance metrics of the global model in the next iteration as the evaluation criteria, the neural architecture search technology is used to optimize the parameters of the global model in the next iteration to obtain the optimal global model in the next iteration.
[0019] Optionally, encrypt the local model parameters of the autonomous vehicle in the current iteration, specifically:
[0020] Use a quantum key to encrypt the local model parameters of the autonomous vehicle in the current iteration.
[0021] Optionally, before locally training the optimal global model in the current iteration according to the sensor training data of the autonomous vehicle in the current iteration to obtain the local model of the autonomous vehicle in the current iteration, it further includes:
[0022] Preprocess the sensor training data of the autonomous vehicle.
[0023] In a second aspect, the present application provides a device for determining autonomous driving decisions, including:
[0024] A model training module, configured to, in the current iteration, for any autonomous vehicle, locally train the optimal global model in the current iteration according to the sensor training data of the autonomous vehicle in the current iteration to obtain the local model of the autonomous vehicle in the current iteration, and send the local model parameters of the autonomous vehicle in the current iteration to the central server;
[0025] A model update module, configured to the central server updates the optimal global model in the current iteration according to the local model parameters of each autonomous vehicle in the current iteration to obtain the global model parameters in the next iteration, and sends the global model in the next iteration to each autonomous vehicle; a model optimization module, configured to optimize the global model parameters in the next iteration according to the running results of the global model in the next iteration on the sensor verification data of each autonomous vehicle to obtain the optimal global model in the next iteration, send the optimal global model in the next iteration to each autonomous vehicle, then update the iteration number, enter the next iteration, until the iteration stop condition is reached, and generate an autonomous driving decision according to the optimal global model in the last iteration number.
[0026] In a third aspect, the present application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, where the processor executes the computer program to implement the method for determining autonomous driving decisions described in any one of the above.
[0027] Fourthly, the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the method for determining an autonomous driving decision described in any one of the above is implemented.
[0028] Fifthly, the present application provides a computer program product, including a computer program. When the computer program is executed by a processor, the method for determining an autonomous driving decision described in any one of the above is implemented.
[0029] According to the specific embodiments provided by the present application, the present application has the following technical effects:
[0030] The present application provides a method, device, equipment, medium and product for determining an autonomous driving decision. After each vehicle trains a model locally using local data and then sends it to the central server, the central server updates the global model according to the local models sent by each vehicle, and trains the model using the federated learning method, which can protect the privacy of users while developing efficient and intelligent autonomous driving technologies. Description of the Drawings
[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.
[0032] Figure 1 It is a schematic flowchart of a method for determining an autonomous driving decision provided by an embodiment of the present application;
[0033] Figure 2 It is a schematic diagram of the principle of a method for determining an autonomous driving decision provided by an embodiment of the present application;
[0034] Figure 3 It is a schematic structural diagram of a computer device provided by an embodiment of the present application. Detailed Embodiments
[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.
[0036] To make the above objects, features and advantages of the present application more obvious and understandable, the present application will be further described in detail below with reference to the drawings and specific embodiments.
[0037] In view of the problems in the related art that the surrounding environment recognition of the autonomous driving system is inaccurate, and personal privacy information may be leaked during model training, which affects the privacy protection of the driver during driving, and the adaptive ability of the system in abnormal situations is weak, prone to incorrect operations and false alarms, seriously affecting the accuracy of the system during use. The present application aims to solve the problem of protecting the privacy of users while developing efficient and intelligent autonomous driving technology. In an exemplary embodiment, as Figure 1 and Figure 2 shown, a method for determining autonomous driving decisions is provided, including the following steps, where:
[0038] Step 201: At the current iteration number, for any autonomous driving vehicle, based on the sensor training data of the autonomous driving vehicle at the current iteration number, locally train the optimal global model at the current iteration number to obtain the local model of the autonomous driving vehicle at the current iteration number, and send the local model parameters of the autonomous driving vehicle at the current iteration number to the central server. The sensor training data of the autonomous driving vehicle is data measured by a variety of sensors on the autonomous driving vehicle, including cameras, lidar, and millimeter-wave radars, as well as road sign and traffic signal data under different weather conditions. Specific data includes: Environmental perception data: including information such as the position, shape, and speed of objects around the vehicle. Vehicle state data: such as the speed, acceleration, and steering wheel angle of the vehicle. Road information data: including road signs, traffic signals, lane lines, etc. Weather and lighting condition data: environmental information under different weathers (such as rain, snow, fog) and lighting (such as day, night). Data of pedestrians and other traffic participants: the position and actions of pedestrians, and the dynamic information of other vehicles.
[0039] Among them, the camera is mainly used to capture high-resolution images and video streams, specifically including: Visual images: used to identify and classify road signs, traffic lights, pedestrians, vehicles, and other obstacles. Lane detection: detecting the position and direction of lane lines through image processing technology to help the vehicle stay in the correct lane. Object recognition: identifying and tracking traffic participants such as pedestrians, vehicles, and bicycles around. Lighting and color information: helping the system adjust the perception algorithm under different lighting conditions to improve the recognition accuracy.
[0040] Lidar and millimeter-wave radar are mainly used to obtain high-precision distance and speed information. Lidar is used to obtain 3D point cloud data: generating a high-precision three-dimensional environmental model, accurately measuring the distance, shape, and position of objects. Used for obstacle detection: identifying and locating obstacles on the road, such as vehicles, pedestrians, curbs, etc. Used for terrain and environmental modeling: helping the vehicle understand complex terrain and environmental structures, such as slopes, potholes, etc.
[0041] Millimeter-wave radar: Used for speed measurement: Precisely measures the relative speed of surrounding objects to help predict their movement trajectories. Used for distance measurement: Can still work stably under different weather conditions (such as rain, fog) and provide reliable distance information. Used for object detection: Identifies and tracks fast-moving objects, such as vehicles traveling at high speeds.
[0042] Step 202: The central server updates the optimal global model at the current iteration based on the local model parameters of each autonomous vehicle at the current iteration to obtain the global model parameters at the next iteration, and sends the global model at the next iteration to each autonomous vehicle.
[0043] Step 203: According to the running results of the global model at the next iteration on the sensor verification data of each autonomous vehicle, optimize the global model parameters at the next iteration to obtain the optimal global model at the next iteration, send the optimal global model at the next iteration to each autonomous vehicle, then update the iteration number, enter the next iteration, until the iteration stop condition is reached, and generate autonomous driving decisions based on the optimal global model at the last iteration.
[0044] Implementing the above steps 201 to 203, the present application can develop efficient and intelligent autonomous driving technologies while protecting user privacy.
[0045] In another exemplary embodiment of the present application, before locally training the optimal global model at the current iteration based on the sensor training data of the autonomous vehicle at the current iteration to obtain the local model of the autonomous vehicle at the current iteration, it further includes:
[0046] Preprocess the sensor training data of the autonomous vehicle. Specifically: The sensor training data collected by the i-th autonomous vehicle at the current iteration t is D i,t , and it is processed through a composite preprocessing function on D i,t , denoted as This function not only includes basic data cleaning, formatting, and data augmentation (by simulating different lighting and weather conditions to increase the diversity and robustness of the data), but also includes more complex feature extraction (using networks such as convolutional networks and transformers to extract multi-modal features).
[0047] The specific steps of this function are:
[0048] First, for D i,tPerform data cleaning and formatting, including using a median filter to remove random noise in the images captured by the camera, adopting the K-nearest neighbor algorithm to fill in the missing values in the point cloud data obtained by the lidar, and normalizing all sensor training data. Standardize the image pixel values to the range of [0, 1], and standardize the radar distance data to have a mean of 0 and a variance of 1. Then, perform time synchronization to ensure that data from different sensors are aligned in time through a well-known linear interpolation method.
[0049] Secondly, perform data augmentation operations, including randomly adjusting the brightness range from 0.5 to 1.5 times, the contrast range from 0.8 to 1.2 times on the images, adding rain and snow effects with fixed parameters, and performing rotations of ±15 degrees, horizontal flips, and scaling (range from 0.9 to 1.1 times). These augmentation steps are carried out after data cleaning and formatting to ensure data diversity and model robustness.
[0050] Finally, perform feature extraction. Use a convolutional neural network to extract spatial features from the camera image data, and at the same time, adopt a well-known Transformer network to perform context modeling on the point cloud data of the lidar and millimeter-wave radar to capture long-range dependencies. Integrate the features from the convolutional neural network and the Transformer network through a feature concatenation method by a well-known method to generate a unified high-dimensional feature representation D ′ i,t 。
[0051] In another exemplary embodiment of the present application, according to the sensor training data of the autonomous vehicle at the current iteration number, locally train the optimal global model at the current iteration number to obtain the local model of the autonomous vehicle at the current iteration number, which specifically includes:
[0052] Step 2001: Forward propagation: Use the preprocessed data D ′ i,i as the input, and calculate the output result (autonomous driving decision result, such as lane keeping instructions, pedestrian detection results, and traffic sign recognition information, for real-time control and path planning of the vehicle) through the neural network model.
[0053] Step 2002: Calculate the main loss function L main (D ′ i,t , θ) value. The main loss function is used to optimize main tasks such as lane keeping, pedestrian detection, and traffic sign recognition.
[0054] Among them, the main loss functions include lane keeping, pedestrian detection, and traffic sign recognition. First, for the lane keeping task, the system evaluates the performance by measuring the distance of the vehicle deviating from the center line of the lane and the angular error between the vehicle's driving direction and the lane direction, and calculates the corresponding loss using these error values to guide the vehicle to drive stably within the lane. Second, in the pedestrian detection task, the system evaluates the accuracy of the model in identifying and locating pedestrians. By comparing the actual detection results with the true annotations, the classification error and the localization error are calculated to quantify the loss and ensure that the model can accurately identify and avoid pedestrians. For the traffic sign recognition task, the system calculates the corresponding loss value by detecting the recognition accuracy and localization accuracy of the model for different traffic signs to ensure that the vehicle can correctly understand and comply with traffic rules. Combining the losses of the above tasks, by setting different weight coefficients, the loss values of lane keeping, pedestrian detection, and traffic sign recognition are weighted and summed to form the overall main loss function.
[0055] Step 2003: Calculate the regularization term or the custom loss term R j (θ): For example, use L2 regularization to control the model complexity, or add an adaptive loss under specific scenarios. In addition, under harsh weather conditions (such as rain and snow), the system will introduce an adaptive loss term for the pedestrian detection task to increase the penalty for the pedestrian localization error under harsh weather.
[0056] Step 2004: Calculate the composite loss function L * : Calculate the value of the main loss L main (D′ i,t ,θ) and the regularization term or the custom loss R j (θ) are weighted and summed according to the coefficient λ to obtain the composite loss L * to optimize multiple objectives simultaneously, that is, the main tasks, such as lane keeping, pedestrian detection, and traffic sign recognition:
[0057]
[0058] Among them, j represents the index of each regularization term or custom loss term, θ is the local model parameter, and λ is the coefficient for adjusting the importance of each item, which is used to balance the weights of the main loss function and the regularization term or custom loss term.
[0059] Step 2005: Backpropagation and parameter update: With the goal of minimizing the composite loss function, adopt the backpropagation algorithm, calculate the gradient according to the composite loss L * and use an optimization algorithm (such as the Adam optimizer) to update the model parameter θ.
[0060] In this embodiment, traditional loss functions are considered in the local model training stage, and regularization terms and custom loss terms are also introduced to improve the generalization ability of the model and its adaptability to specific tasks.
[0061] In another exemplary embodiment of the present application, the local model parameters of the autonomous vehicle at the current iteration are sent to the central server, which specifically includes:
[0062] Encrypt the local model parameters of the autonomous vehicle at the current iteration to obtain the encrypted local model parameters of the autonomous vehicle at the current iteration;
[0063] Send the encrypted local model parameters of the autonomous vehicle at the current iteration to the central server.
[0064] In another exemplary embodiment of the present application, encrypting the local model parameters of the autonomous vehicle at the current iteration specifically includes:
[0065] Use the quantum key K QKD to encrypt the local model parameters of the autonomous vehicle at the current iteration. Specifically:
[0066] Step 1: Quantum key generation: Through quantum key distribution (QKD) technology, the communicating party A (central server) generates a series of random quantum states and sends them to the communicating party B (autonomous vehicle) through a quantum channel.
[0067] Step 2: Quantum state measurement: After receiving the random quantum states, the communicating party B selects a measurement basis for measurement and records the results.
[0068] Step 3: Basis information exchange: The communicating parties A and B exchange measurement basis information through a classical channel, filter out the measurement results using the same basis, and generate a shared secure key K QKD .
[0069] Step 4: Encrypted model update: Use the shared key K QKD to encrypt the local model parameters to ensure that they are not eavesdropped or tampered with during transmission.
[0070] In another exemplary embodiment of the present application, the central server updates the optimal global model at the current iteration based on the local model parameters of each autonomous vehicle at the current iteration to obtain the global model parameters at the next iteration, which specifically includes:
[0071] The central server decrypts the encrypted local model parameters of the autonomous vehicle at the current iteration to obtain the decrypted local model parameters of the autonomous vehicle at the current iteration;
[0072] Update the optimal global model at the current iteration based on the decrypted local model parameters of each autonomous vehicle at the current iteration to obtain the global model parameters at the next iteration.
[0073] In another exemplary embodiment of the present application, the central server decrypts the encrypted local model parameters of the autonomous vehicles at the current iteration, specifically:
[0074] The central server uses the corresponding quantum key K QKD To decrypt the received decrypted local model parameters.
[0075] Considering the performance differences of different autonomous vehicles in different environments and tasks, a dynamic weight allocation method is adopted to optimize the global model. The contribution of each autonomous vehicle is no longer equal, but is dynamically adjusted according to its data quality, model performance, and task relevance. Therefore, in another exemplary embodiment of the present application, the optimal global model at the current iteration is updated based on the decrypted local model parameters of each autonomous vehicle at the current iteration to obtain the global model parameters at the next iteration, and the global model at the next iteration is sent to each autonomous vehicle, specifically:
[0076] First, evaluate the contribution of each autonomous vehicle from the following three aspects:
[0077] Data quality: The integrity, accuracy, and representativeness of the data collected by the autonomous vehicle.
[0078] Model performance: The performance of the local model of the autonomous vehicle on the validation set, such as indicators like accuracy and recall rate.
[0079] Task relevance: The consistency and relevance of the tasks processed by the autonomous vehicle to the global model goal.
[0080] Consistency evaluation: The degree of matching (i.e., the repetition degree) between the local task and the global task goal. The higher the matching degree, the higher the score. For example, if the path planning and traffic sign recognition of vehicle A completely match the path planning and traffic sign recognition tasks of the global model, it is considered a complete match. If only one core task of the task processed by a certain vehicle is consistent with the global task, it represents a low degree of matching.
[0081] Relevance evaluation: Evaluate the actual contribution degree of the local task of each vehicle to achieving the global model goal. The higher the contribution degree, the higher the relevance. For example, obstacle avoidance is directly related to driving safety and is considered a high contribution. Although vehicle status monitoring helps the overall system, its impact on specific decisions is limited, so it is considered a low contribution.
[0082] Then, for dynamic weight allocation, based on the aforementioned evaluation results, the decrypted weight w of each autonomous vehicle at the current iteration t is i,t adjusted to ensure that and reflects the actual contribution of each vehicle, where w i,t represents the weight of the i-th autonomous vehicle in the global model update. Next, the central server collects the encrypted model updates Δθ of each vehicle through the quantum key distribution (QKD) process i,t , and uses the key K QKD to decrypt it, thereby obtaining the decrypted local model parameter updates. Finally, the decrypted updates are weighted and aggregated according to the dynamic weights: According to the formula
[0083]
[0084] By means of weighted summation, the local model updates of each autonomous vehicle are fused into a comprehensive global model update. Among them, Δθ i,t represents the decrypted local model parameters of the i-th autonomous vehicle at the current iteration t, and N represents the total number of autonomous vehicles participating in the training.
[0085] The purpose of this aggregation process is to fuse the model updates of all participating autonomous vehicles to form a comprehensive and improved global model. Through quantum key encryption and decryption, the security of the model update process is ensured, preventing potential eavesdropping and data leakage during transmission. After the model update aggregation, through global model distribution, this step ensures that all autonomous vehicles can update their local models in a timely manner, making the model more accurately reflect the cumulative learning results. This method not only improves the performance of the model but also maintains the consistency and synchronization of the model, thereby enhancing the efficiency and effectiveness of the entire autonomous vehicle.
[0086] In another exemplary embodiment of the present application, according to the running results of the global model on the sensor verification data of each autonomous vehicle at the next iteration, the global model parameters at the next iteration are optimized to obtain the optimal global model at the next iteration, specifically including:
[0087] According to the running results of the global model on the sensor verification data of each autonomous vehicle at the next iteration, determine the numerical values P of the performance indicators (such as accuracy, recall, etc.) of the global model on the verification data of each autonomous vehicle at the next iteration i (a);
[0088] According to the numerical values of the performance indicators of the global model on the verification data of each autonomous vehicle at the next iteration, determine the numerical value P of the performance indicator of the global model at the next iteration global (a);
[0089] Taking the numerical value of the performance metric of the global model in the next iteration as the evaluation criterion, the neural architecture search (NAS) technology is used to optimize the parameters of the global model in the next iteration, and the optimal global model in the next iteration is obtained.
[0090] Because different vehicles may encounter different environmental conditions, and each vehicle can only access local and limited data. In order to better find the optimal autonomous driving model in the federated learning framework that ensures data privacy, the technology of self-evolving model architecture is introduced. Specifically, it combines the NAS technology with an adaptive adjustment strategy.
[0091] In another exemplary embodiment of the present application, taking the numerical value of the performance metric of the global model in the next iteration as the evaluation criterion, using the neural architecture search technology to optimize the parameters of the global model in the next iteration, and obtaining the optimal global model in the next iteration, specifically includes:
[0092] Step A: Define the search space A as the variable part of the model architecture (such as the number of layers, the number of autonomous driving vehicles in each layer, the type of activation function, etc.) and the global model architecture a of the initial architecture in the next iteration.
[0093] In another exemplary embodiment of the present application, the numerical value P global (a) of the performance metric of the global model in the next iteration is used as the evaluation criterion, and the neural architecture search technology is used to optimize the parameters of the global model in the next iteration, and the optimal global model in the next iteration is obtained, specifically including:
[0094] Step B: Perform neural network search (NAS).
[0095] First, define the state space S as the current network configuration, including the number of layers, the number of autonomous driving vehicles in each layer, the type of activation function, etc. Define the action space A as possible operations, such as adding layers, reducing layers, modifying activation functions, etc. Define the reward function R(s,a) as based on the performance of the new architecture on the validation set (such as accuracy improvement or loss reduction), and give rewards according to the performance of the network on the validation set after taking action a (such as accuracy improvement or loss reduction).
[0096] Then, optimize the policy π through reinforcement learning to maximize the expected return Q to find the optimal architecture a ′The learning optimization strategy π is a strategy used to guide the neural architecture search process. Through reinforcement learning, the strategy is trained to select optimal actions (such as increasing the number of layers, modifying activation functions, etc.) in the search space to find the best model architecture. Among them, the search space is the set of all possible neural network architectures, and the state space is the current state of these neural network architectures.
[0097] In this way, the system can learn more effectively from each autonomous vehicle and improve the overall performance and adaptability of the global model.
[0098]
[0099] Among them, Q π (s,a) represents the expected return of taking action a in state s under policy π. s represents the current network configuration state, including the number of layers, the number of autonomous vehicles in each layer, the type of activation function, etc. a represents the actions taken, such as adding layers, reducing layers, modifying activation functions, etc. E is the expectation operator representing the expectation (average value) of future rewards. k is an index variable representing the number of subsequent time steps starting from the current time step t. S t represents the current network configuration state at time step t, and A t represents the set of actions and policies taken at time step t.
[0100] γ represents the discount factor used to adjust the current value of future rewards. R t+k+1 represents the reward obtained at time t + k + 1. Based on the performance of the network on the validation set, the optimization strategy π represents finding the optimal strategy for generating a high-performance network structure by iteratively optimizing the strategy π to maximize the Q function.
[0101] Step C: Evaluate the candidate architectures.
[0102] Processing method: Deploy the candidate architecture a ′ to individual autonomous vehicles, collect the performance of the new architecture a ′ in the actual driving environment, and calculate the performance metric P candidate (a ′ ) to evaluate the new architecture.
[0103] Step D: Select the optimal architecture and update the global model.
[0104] According to the performance metric P candidate (a ′ ) select the architecture a * with the best performance, and update the global model parameters. Among them, It represents the defined architecture search space. By iteratively optimizing the policy π to maximize the Q function, the optimal policy for generating high-performance network structures can be found. For the adaptive adjustment policy, in addition to updating parameters based on the loss function, the dynamic adjustment mechanism of the network structure needs to be considered in detail. Considering that the adjustment of the network structure not only affects the complexity of the model but also may affect its communication efficiency in a distributed environment, an evaluation function E(a) that comprehensively considers performance and efficiency is introduced to evaluate the overall quality of the architecture, which is defined as: E(a) = λ1·Performance(a) - λ2·Complexity(a) - λ3·CommunicationCost(a).
[0105] Among them, λ1, λ2, and λ3 are the coefficients for adjusting the importance of each index respectively; Performance(a) represents the performance of the model (such as accuracy) under the given architecture a, that is, the proportion of correct samples in the total samples; Complexity(a) represents the complexity of the model (such as the number of parameters and the amount of computation), that is, the total sum of trainable parameters and the number of floating-point operations in the forward propagation process; CommunicationCost(a) includes the transmission volume and transmission frequency of model parameters, representing the communication cost of model updates in the federated learning environment.
[0106] Based on the above evaluation function, the adjustment strategy of the network structure can be achieved by solving the following optimization problem To solve this optimization problem, strategies such as gradient-based methods or evolutionary algorithms can be adopted.
[0107] Step E: Apply the adaptive adjustment strategy.
[0108] To enable the model to adapt to the changing data distribution in the federated learning environment, an adaptive adjustment strategy is introduced. This strategy dynamically adjusts the network architecture according to the performance feedback of the model on different autonomous vehicles.
[0109] After each round of federated learning, evaluate the performance P of the global model a * on each autonomous vehicle i (a * ). According to the global performance metric P global (a * ), decide whether to keep the current architecture unchanged or update the model according to the new architecture found by NAS.
[0110] Suppose the performance on the i-th autonomous vehicle is P i (a), and the global performance is P global (a), then the performance feedback can be expressed as:
[0111]
[0112] Among them, in the federated learning environment, the evaluation of model performance not only needs to consider the performance of individual autonomous vehicles, but also the comprehensive performance of the entire network. Therefore, a global performance evaluation index P is introduced. global (a): Global performance index, which synthesizes the model performance of all autonomous vehicles.
[0113] Decide whether to update the model architecture according to the global performance:
[0114]
[0115] NAS(a): The new architecture found through the NAS process.
[0116] Dynamic adjustment: If it is decided to update the architecture, the new architecture a found by NAS is adopted. * Perform model update to ensure that the model can adapt to the changing data distribution and driving environment.
[0117] In this way, the autonomous driving system under federated learning can not only improve the model performance through the cooperation of autonomous vehicles while protecting privacy, but also ensure that the model architecture always adapts to the current data and task requirements through NAS and adaptive adjustment strategies, so as to maintain high efficiency and accuracy in a complex and changing environment.
[0118] This application uses the method of federated learning to locally train the model, protect the privacy of passengers and drivers, optimize the loss function during the training process, and introduce regularization terms and custom loss terms to improve the generalization ability of the model and its adaptability to specific tasks; in addition, an adaptive adjustment strategy is introduced to enable the autonomous driving system to adapt to the changing data distribution under federated learning, accurately understand its surrounding environment, and make quick responses, while developing efficient and intelligent autonomous driving technologies, protecting the privacy of users. This application realizes the self-evolution of the system through processes such as data collection, data processing, model training, network adjustment, and model evaluation. This self-evolving model architecture can continuously adapt to new road environments and driving conditions while ensuring data privacy and security.
[0119] Based on the same inventive concept, the embodiment of this application also provides an apparatus for determining an autonomous driving decision for implementing the method for determining an autonomous driving decision involved above. The solution provided by this apparatus for solving problems is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the apparatus for determining an autonomous driving decision provided below can refer to the limitations on the method for determining an autonomous driving decision in the above text, and will not be elaborated here.
[0120] In an exemplary embodiment, an apparatus for determining an autonomous driving decision is provided, including:
[0121] A model training module, which is used to, for any autonomous vehicle at the current iteration number, locally train the optimal global model at the current iteration number according to the sensor training data of the autonomous vehicle at the current iteration number, obtain the local model of the autonomous vehicle at the current iteration number, and send the local model parameters of the autonomous vehicle at the current iteration number to the central server;
[0122] A model updating module, which is used for the central server to update the optimal global model at the current iteration number according to the local model parameters of each autonomous vehicle at the current iteration number, obtain the global model parameters at the next iteration number, and send the global model at the next iteration number to each autonomous vehicle; A model tuning module, which is used to tune the global model parameters at the next iteration number according to the running results of the global model at the next iteration number on the sensor verification data of each autonomous vehicle, obtain the optimal global model at the next iteration number, send the optimal global model at the next iteration number to each autonomous vehicle, then update the iteration number, enter the next iteration, until the iteration stop condition is reached, and generate an autonomous driving decision according to the optimal global model at the last iteration number.
[0123] In an exemplary embodiment, the apparatus for determining an autonomous driving decision further includes:
[0124] A data collection module: introducing a variety of sensors, including cameras, lidars, and millimeter-wave radars, to collect a more rich and diverse data set, including road sign and traffic signal data under different weather conditions.
[0125] A data preprocessing module: adopting advanced preprocessing methods to more effectively extract useful information from the collected data and reduce noise.
[0126] The model training module of the present application: collecting richer data and locally training the model based on federated learning, optimizing the loss function, introducing a regularization term and a custom loss term to improve the generalization ability of the model and its adaptability to specific tasks. The model updating module: adopting a method of dynamic weight allocation to optimize the process of global model updating. The model tuning module: optimizing the model by introducing an adaptive adjustment strategy so that it can adapt to the continuously changing data distribution under federated learning. The model evaluation module: evaluating the model performance of all autonomous vehicles based on global performance evaluation indicators.
[0127] In an exemplary embodiment, a computer device is provided. The computer device can be a server or a terminal, and its internal structure diagram can be as Figure 3As shown in the figure. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the determined data for autonomous driving decisions. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements a method for determining autonomous driving decisions.
[0128] Those skilled in the art can understand that Figure 3 the structure shown in the figure is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the above method embodiments are implemented.
[0129] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by the processor, the above method embodiments are implemented.
[0130] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by the processor, the above method embodiments are implemented.
[0131] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with relevant regulations.
[0132] In the present application, all actions of obtaining signals, information, or data are carried out on the premise of complying with the corresponding data protection regulations and policies of the country where the location is located and obtaining authorization from the owner of the corresponding device.
[0133] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.
[0134] The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processors, graphics processors, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.
[0135] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.
[0136] Specific examples are used in this article to elaborate on the principles and implementation manners of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present application.
Claims
1. A method for determining an autonomous driving decision, characterized in that, The method for determining the autonomous driving decision includes: At the current iteration number, for any autonomous driving vehicle, based on the sensor training data of the autonomous driving vehicle at the current iteration number, locally train the optimal global model at the current iteration number to obtain the local model of the autonomous driving vehicle at the current iteration number, and send the local model parameters of the autonomous driving vehicle at the current iteration number to the central server; The central server updates the optimal global model at the current iteration number according to the local model parameters of each autonomous driving vehicle at the current iteration number to obtain the global model parameters at the next iteration number, and sends the global model at the next iteration number to each autonomous driving vehicle; According to the running results of the global model at the next iteration number on the sensor verification data of each autonomous driving vehicle, optimize the global model parameters at the next iteration number to obtain the optimal global model at the next iteration number, send the optimal global model at the next iteration number to each autonomous driving vehicle, then update the iteration number, enter the next iteration, until the iteration stop condition is reached, and generate the autonomous driving decision according to the optimal global model at the last iteration number.
2. The method for determining an autonomous driving decision according to claim 1, wherein Sending the local model parameters of the autonomous driving vehicle at the current iteration number to the central server specifically includes: Encrypt the local model parameters of the autonomous driving vehicle at the current iteration number to obtain the encrypted local model parameters of the autonomous driving vehicle at the current iteration number; Send the encrypted local model parameters of the autonomous driving vehicle at the current iteration number to the central server.
3. The method for determining an autonomous driving decision according to claim 2, wherein The central server updates the optimal global model at the current iteration number according to the local model parameters of each autonomous driving vehicle at the current iteration number to obtain the global model parameters at the next iteration number, specifically including: The central server decrypts the encrypted local model parameters of the autonomous driving vehicle at the current iteration number to obtain the decrypted local model parameters of the autonomous driving vehicle at the current iteration number; Update the optimal global model at the current iteration number according to the decrypted local model parameters of each autonomous driving vehicle at the current iteration number to obtain the global model parameters at the next iteration number.
4. The method for determining an autonomous driving decision according to claim 1, wherein The optimizing the global model parameters at the next iteration number according to the running results of the global model at the next iteration number on the sensor verification data of each autonomous driving vehicle to obtain the optimal global model at the next iteration number specifically includes: Determine the performance index values of the global model at the next iteration number under the verification data of each autonomous driving vehicle according to the running results of the global model at the next iteration number on the sensor verification data of each autonomous driving vehicle; Determine the performance index value of the global model at the next iteration number according to the performance index values of the global model at the next iteration number under the verification data of each autonomous driving vehicle; Taking the performance index value of the global model at the next iteration number as the evaluation criterion, use the neural architecture search technology to optimize the global model parameters at the next iteration number to obtain the optimal global model at the next iteration number.
5. The method for determining an autonomous driving decision according to claim 2, wherein Encrypt the local model parameters of the autonomous vehicle at the current iteration number, specifically: Use a quantum key to encrypt the local model parameters of the autonomous vehicle at the current iteration number.
6. The method for determining an autonomous driving decision according to claim 1, wherein Before locally training the optimal global model at the current iteration number based on the sensor training data of the autonomous vehicle at the current iteration number to obtain the local model of the autonomous vehicle at the current iteration number, it further includes: Preprocess the sensor training data of the autonomous vehicle.
7. A device for determining an autonomous driving decision, characterized in that, The determination device for autonomous driving decisions includes: A model training module, which is used to, at the current iteration number, for any autonomous vehicle, locally train the optimal global model at the current iteration number based on the sensor training data of the autonomous vehicle at the current iteration number to obtain the local model of the autonomous vehicle at the current iteration number, and send the local model parameters of the autonomous vehicle at the current iteration number to the central server; A model update module, which is used for the central server to update the optimal global model at the current iteration number based on the local model parameters of each autonomous vehicle at the current iteration number to obtain the global model parameters at the next iteration number, and send the global model at the next iteration number to each autonomous vehicle; A model tuning module, which is used to tune the global model parameters at the next iteration number according to the running results of the global model at the next iteration number on the sensor verification data of each autonomous vehicle to obtain the optimal global model at the next iteration number, send the optimal global model at the next iteration number to each autonomous vehicle, then update the iteration number, enter the next iteration, until the iteration stop condition is reached, and generate an autonomous driving decision according to the optimal global model at the last iteration number.
8. A computer device, comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the method for determining an autonomous driving decision according to any one of claims 1-6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for determining an autonomous driving decision according to any one of claims 1-6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for determining an autonomous driving decision according to any one of claims 1-6.