Automatic driving decision-making system based on information entropy-Bayesian estimation

Abstract

The invention provides an automatic driving decision-making system based on information entropy-Bayesian estimation. The automatic driving decision-making system comprises an environment sensing module, a hybrid quantification module and a decision-making optimization module. The environment sensing module is used for collecting and processing sensor data and completing target detection and classification; the hybrid quantization module performs reliability quantization on the recognition result based on information entropy and Bayesian uncertainty estimation, generates uncertainty evaluation results of different targets and scenes, and transmits the uncertainty evaluation results to the decision and control module; and the decision and control module dynamically adjusts the strategy of the automatic driving system according to the uncertainty measurement of each target. The automatic driving decision-making system based on information entropy-Bayesian estimation is used for quantifying the target recognition reliability in an automatic driving scene. The output confidence distribution of the target recognition model is evaluated by calculating the information entropy, and the uncertainty of the recognition model is estimated by using a Bayesian uncertainty estimation (such as Monte Carlo Dropout) method.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to an autonomous driving decision system based on information entropy-Bayesian estimation. Background Art

[0002] Autonomous driving technology is rapidly becoming an important part of intelligent transportation systems, aiming to improve road safety, traffic efficiency and user experience by reducing human intervention. Current autonomous driving systems rely on a variety of sensors (such as cameras, lidar and radar) to obtain information about the surrounding environment, and use machine learning and deep learning algorithms to process and analyze these perception data to achieve target detection, object recognition, path planning and decision control. However, in complex and changing real-world traffic scenarios, these technologies still face major challenges in target recognition and decision reliability.

[0003] The reliability of environmental perception and decision-making systems is one of the key bottlenecks in the current development of autonomous driving technology. In the process of autonomous driving, the collection and analysis of sensor data inevitably contain noise, uncertainty and errors. Especially in bad weather, poor light or limited field of view, it is difficult for traditional target recognition algorithms to guarantee high accuracy and low false alarm rate. At the same time, due to the lack of comprehensive quantitative assessment of uncertainty, existing systems often cannot make the best decision quickly when faced with complex scenarios, which may lead to dangerous driving behavior or traffic accidents.

[0004] At present, an autonomous driving decision system in the prior art applies the Bayesian deep learning method to the autonomous driving system to quantify the uncertainty of target detection and recognition. Bayesian deep learning estimates the confidence interval of the model output by introducing a probability model in the training process of the neural network. This method judges the reliability of the model prediction by analyzing the output probability distribution. Bayesian deep learning models usually show good uncertainty quantification capabilities in tasks such as image classification and target detection. The Bayesian method is used to quantify the confidence interval of the model output, evaluate the reliability of target recognition, and provide richer probability information for decision-making in complex environments and provide a probability distribution of recognition results, which helps to reduce the frequency of high-risk decisions. However, it only relies on the probability distribution of the model output, lacks multi-level and multi-dimensional information comprehensive judgment, and cannot make full use of other sensor data or external information to improve decision-making.

[0005] At present, another autonomous driving decision system in the prior art is a method that combines deep learning with information entropy to better deal with the uncertainty problem in autonomous driving. Information entropy, as the basis of uncertainty measurement, can be used to evaluate the confusion and confidence of target recognition. By analyzing the entropy value of the model prediction results, the system can judge the reliability of the recognition results. For example, in an object detection task, if the prediction results of multiple candidate areas have a high information entropy, it means that the system has a high uncertainty in the current recognition, which can prompt the system to make a more cautious decision. Use information entropy as an uncertainty quantification tool to provide a reliability measure for target recognition. Evaluate the confidence of the decision based on the output results of the deep learning model. However, only based on information entropy to quantify uncertainty, without combining other uncertainty estimation methods (such as Bayesian estimation), the comprehensiveness and accuracy of the reliability measurement are still limited, and lack of deep integration with multi-sensor data fusion, it is impossible to fully utilize the advantages of various perception data.

[0006] The shortcomings of the above-mentioned autonomous driving decision-making system in the prior art include:

[0007] The uncertainty quantification method is single and cannot fully evaluate the reliability of target recognition. Most existing technologies use a single uncertainty quantification method, such as a confidence-based probability model (deep learning confidence output) or Bayesian uncertainty estimation. A single method can only reflect the reliability of target recognition in a specific scenario and is difficult to cope with the diverse and complex actual traffic scenarios. This single uncertainty estimation may cause the system to make wrong decisions in high uncertainty or unclear environments, thereby increasing risks.

[0008] Lack of multi-level and multi-dimensional information integration. Existing technologies often lack effective integration of multi-level information (such as data from different sensors, historical decision data, and environmental perception data). Multi-sensor data fusion technology mainly focuses on the fusion accuracy of data, while ignoring the interaction and dynamic adjustment capabilities between multi-level information. In addition, many methods rely only on a single data source or fixed model, resulting in insufficient analysis capabilities for complex scenes.

[0009] Insufficient real-time dynamic decision-making capabilities. Many methods in existing autonomous driving technologies lack the ability to adjust decisions in real time and dynamically. Most systems use preset strategies or fixed learning models. When the environment changes or data becomes abnormal, it is difficult to adjust decisions in time, which can easily lead to decision delays or misjudgments. In addition, although some algorithms based on deep learning and reinforcement learning can optimize decisions during the training phase, they lack adaptability in practical applications. Summary of the invention

[0010] An embodiment of the present invention provides an autonomous driving decision system based on information entropy-Bayesian estimation to effectively improve the safety of the autonomous driving system.

[0011] In order to achieve the above object, the present invention adopts the following technical scheme.

[0012] An automatic driving decision system based on information entropy-Bayesian estimation, comprising: an environment perception module, a hybrid quantization module and a decision optimization module;

[0013] The environment perception module is used to collect and process data from different sensors, perform target detection and classification after preprocessing the collected data, obtain target recognition results, and transmit the target recognition results to the hybrid quantization module;

[0014] The hybrid quantization module is used to quantify the reliability of the target recognition result output by the environmental perception module using information entropy and Bayesian uncertainty estimation method, to form uncertainty evaluation results for different targets and scenes, and transmit the uncertainty evaluation results to the decision and control module;

[0015] The decision and control module is used to dynamically adjust the decision and control strategy of the autonomous driving system according to the final uncertainty measurement value of each target transmitted by the hybrid quantization module.

[0016] Preferably, the system further comprises:

[0017] The feedback learning module is used to collect the data generated during the operation of the autonomous driving vehicle and the mixed quantitative results of the system, conduct continuous learning and optimization, and improve the performance of the autonomous driving decision-making system in different scenarios.

[0018] Preferably, the system further comprises:

[0019] The risk warning and protection module is used to provide warning and protection when high uncertainty or potential risks are identified, ensuring the safety of the autonomous driving decision-making system.

[0020] Preferably, the environment perception module comprises: a data acquisition submodule, a data preprocessing submodule and a feature extraction and target detection submodule;

[0021] The data acquisition submodule is used to collect environmental data in real time through the sensor array, transmit the collected raw data to the data preprocessing submodule, collect the RGB image sequence around the autonomous driving vehicle through the camera, and output the image data set I = {I t}, where t represents the time series, and the point cloud data P of the autonomous driving vehicle is collected by the lidar i =(x i ,yi , z i ), where p i The three-dimensional coordinates of the points scanned by the laser radar are collected through the millimeter-wave radar to collect the distance, speed and angle information of the autonomous driving vehicle to form radar target data R = (d i , v i ,θ i ), where d i is the target distance, v i is the relative speed, θ i is the angle;

[0022] The data preprocessing submodule is used to preprocess the raw data transmitted from the data acquisition submodule, and transmit the preprocessed data to the feature extraction and target detection submodule; the image preprocessing uses a Gaussian filter to smooth the image, the point cloud data preprocessing uses a random sampling consistency RANSAC algorithm to segment the ground and non-ground points, and the radar target data is denoised, including using a Kalman filter to estimate the target position and speed;

[0023] The feature extraction and target detection submodule is used to use the deep learning model to perform target detection and classification on the preprocessed data to obtain target recognition results, wherein the targets include pedestrians, vehicles and traffic signs, and the target recognition results include the category, position, speed and direction information of the target;

[0024] Image object detection uses a convolutional neural network for object detection. The input image I is used to extract the feature map F through multi-layer convolution operations, and finally the detection box and category are output: F = CNN(I){(x, y, w, h, c)}, where (x, y) is the center of the detection box, w and h are the width and height, and c is the category confidence.

[0025] Point cloud data target detection uses a clustering and shape matching-based method to detect objects, and clusters different targets through the density clustering DBSCAN algorithm, as shown in formula (4):

[0026] DBSCAN(P,∈,MinPts)→{P cluster} (4)

[0027] Among them, ∈ is the maximum neighborhood distance, and MinPts is the minimum number of points to form a cluster.

[0028] Preferably, the hybrid quantization module comprises: an information entropy calculation submodule, a Bayesian uncertainty estimation submodule and a hybrid quantization result integration submodule;

[0029] The information entropy calculation submodule is used to calculate the confidence distribution of each target according to the target detection result output by the environment perception module, and to measure the confusion and uncertainty of target recognition according to the entropy value by calculating the entropy value of the confidence distribution. n}Calculate information entropy H(p): where p i is the confidence of the i-th category;

[0030] The Bayesian uncertainty estimation submodule is used to estimate the uncertainty of the target recognition results output by the environmental perception module to obtain the probability distribution of the model prediction, analyze the multiple prediction results of each target, calculate the Bayesian uncertainty metric, evaluate the recognition reliability of the model under different circumstances, and use Monte Carlo dropout to calculate the mean μ and variance σ of multiple prediction results through multiple forward propagations during the neural network reasoning process. 2 , as shown in formula (5):

[0031]

[0032] in, is the prediction result of the tth forward propagation, and T is the total number of forward propagations;

[0033] The hybrid quantization result integration submodule is used to integrate the results of information entropy and Bayesian uncertainty estimation, calculate the hybrid quantization index according to the preset weight or dynamically adjusted rules, determine the final uncertainty measurement value of each target, and transmit the final uncertainty measurement value of each target to the decision and control module. The hybrid quantization index Q = α·H(p)+β·σ 2 , where α and β are preset or dynamically adjusted weight coefficients.

[0034] Preferably, the decision and control module comprises: an information entropy calculation submodule, a Bayesian uncertainty estimation submodule and a hybrid quantization result integration submodule;

[0035] The strategy generation submodule is used to generate a driving strategy of the autonomous driving system according to the final uncertainty measurement value of each target transmitted from the hybrid quantization module, generate a conservative driving strategy in a high uncertainty scenario, and generate an aggressive driving strategy in a low uncertainty scenario. The driving strategy is generated according to the hybrid quantization index Q and a preset threshold, as shown in formula (6):

[0036]

[0037] Among them, Q threshold1 and Q threshold2 is the threshold value;

[0038] The real-time adjustment submodule is used to monitor environmental changes and sensor data in real time, dynamically adjust driving strategies, and automatically switch to safety mode when potential risks or increased uncertainties are detected;

[0039] The control instruction execution submodule is used to convert the driving strategy generated by the strategy generation submodule into a specific control instruction, and send the instruction to each actuator of the vehicle through the electronic control unit to complete the actual operation of the vehicle. The formula is as follows:

[0040] δ=f(S, environmental information), a=g(S, distance of obstacle ahead), b=h(S, current vehicle speed).

[0041] The steering wheel angle δ is calculated by a function f, the input of which includes the system state S and environmental information. The system state S includes the vehicle's position information and current direction, and the environmental information includes the road condition and lane line position. The acceleration a is calculated by a function g, the input of g is the system state S and the distance to the obstacle ahead, and the braking intensity b is calculated by a function h, the input of h is the system state S and the current vehicle speed.

[0042] Preferably, the feedback learning module includes: a data collection and analysis submodule, a model optimization submodule and an adaptive adjustment submodule;

[0043] The data collection and analysis submodule is used to continuously collect vehicle sensor data, environmental data and decision data. A historical database is formed and the performance of the mixed quantitative results in each scenario is analyzed and compared with the actual driving situation; I is the input information collected at time t, P is the strategy decision at time t, indicating the driving strategy selected by the autonomous driving system at that moment, and R t is the actual result after executing the strategy at time t, S t is the system status at time t, including the current position, speed and direction information of the vehicle, which is used to record the real-time status of the vehicle;

[0044] The model optimization submodule is used to continuously optimize the calculation method and weight coefficient of information entropy and Bayesian uncertainty estimation based on historical data and feedback results, and improve the adaptability and accuracy of the autonomous driving system model in specific scenarios through online learning or incremental training, as shown in formula (7):

[0045]

[0046] in, is the loss function, is the mixed quantitative index for prediction, Q * is a true reliability indicator;

[0047] The adaptive adjustment submodule is used to update the decision and control strategies in real time according to the continuously optimized autonomous driving system model and hybrid quantization system.

[0048] Preferably, the risk warning and protection module includes: a risk monitoring submodule, a warning prompt submodule and a protection strategy execution submodule;

[0049] The risk monitoring submodule is used to monitor the mixed quantitative indicators. When the uncertainty exceeds the set threshold, it is determined that there is a potential risk. The risk assessment threshold R is set according to the mixed quantitative result Q. threshold , when Q ≥ R threshold When the warning signal is triggered;

[0050] The warning prompt submodule is used to provide risk warning prompts to the driver or passengers, prompting the automatic driving system to perform self-protection actions;

[0051] The protection strategy execution submodule is used to automatically execute the protection strategy in a high-risk scenario, trigger the full system protection program in an emergency state, and calculate the emergency operation parameters after triggering the warning, as shown in formula (8):

[0052] b emergency =max(b current , b max ) (8)

[0053] Among them, b current is the current braking intensity, b max It is the maximum braking intensity in an emergency.

[0054] It can be seen from the technical solution provided by the present invention that the present invention proposes a hybrid quantization system based on information entropy and Bayesian uncertainty estimation, which is used to quantify the reliability of target recognition in autonomous driving scenarios; the output confidence distribution of the target recognition model is evaluated by calculating the information entropy, and the uncertainty of the recognition model is estimated by using the Bayesian uncertainty estimation (such as Monte Carlo Dropout) method.

[0055] The present invention proposes a real-time decision optimization module based on hybrid quantization results, which dynamically adjusts the vehicle's driving strategy (aggressive mode, robust mode, conservative mode) according to different reliability indicators. The decision optimization module can dynamically adjust control parameters (such as steering wheel angle, acceleration, braking intensity, etc.) according to the hybrid quantization results in different scenarios, thereby improving the flexibility and safety of the system.

[0056] Additional aspects and advantages of the present invention will be given in part in the following description, which will become obvious from the following description, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0058] Figure 1 A structural diagram of an autonomous driving decision system based on information entropy-Bayesian estimation provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0060] It will be understood by those skilled in the art that, unless expressly stated, the singular forms "one", "said", and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, the "connection" or "coupling" used herein may include wireless connection or coupling. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0061] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as those generally understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with the meanings in the context of the prior art, and will not be interpreted with idealized or overly formal meanings unless defined as herein.

[0062] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each embodiment does not constitute a limitation on the embodiments of the present invention.

[0063] Aiming at the problem of single uncertainty quantification method, this paper proposes a hybrid quantification system based on information entropy and Bayesian uncertainty estimation. This method combines multiple uncertainty assessment methods to provide a more comprehensive target recognition reliability assessment, thereby improving the adaptability of the system in complex traffic environments. By building a multi-level reliability quantification framework and integrating data and dynamic feedback from multiple sensors, richer environmental perception and information fusion are achieved, making up for the shortcomings of existing technologies in multi-dimensional information integration.

[0064] The embodiment of the present invention proposes a hybrid quantization system based on information entropy and Bayesian uncertainty estimation, which evaluates the output of the environmental perception module in real time and optimizes the vehicle's decision-making strategy based on the quantification results. The system implements a closed loop of the entire process from perception, quantification to decision-making and learning, and adopts a modular design, making the system easy to expand and maintain, and the data transmission and functional separation between modules are clear and explicit.

[0065] The structure diagram of an automatic driving decision system based on information entropy-Bayesian estimation provided by an embodiment of the present invention is as follows: Figure 1 As shown, it includes: environment perception module, hybrid quantification module, decision optimization module, risk warning module and feedback learning module.

[0066] 1: Environmental perception module, which is used to collect and process data from different sensors, pre-process the collected data, perform target detection and classification, obtain target recognition results, and transmit the target recognition results to the hybrid quantification module. It realizes functions such as environmental modeling, target detection and recognition, and obstacle positioning, and provides basic information for subsequent reliability quantification and decision-making. The above sensors include cameras, lidars, millimeter-wave radars, and ultrasonic sensors. The environmental perception module includes three sub-modules: data acquisition sub-module, data pre-processing sub-module, and feature extraction and target detection sub-module.

[0067] The data acquisition submodule is used to collect environmental data in real time through a sensor array (such as cameras, lidar, and millimeter-wave radar, etc.), and transmit the collected raw data to the data preprocessing submodule. Each sensor is responsible for collecting a specific type of data. The camera data acquisition can capture the RGB (red (R), green (G), blue (B)) image sequence around the autonomous driving vehicle, and output the image data set I = {I t}, where t represents the time series. LiDAR data collects point cloud data of autonomous driving vehicles P = p i =(x i ,y i , z i ), where p iThe millimeter wave radar collects and captures the distance, speed and angle information of the autonomous driving vehicle to form radar target data R = (d i , v i ,θ i ), where d i is the target distance, v i is the relative speed, θ i is the angle.

[0068] The data preprocessing submodule is used to preprocess the raw data transmitted from the data acquisition submodule and transmit the preprocessed data to the feature extraction and target detection submodule. The above preprocessing includes removing noise, correcting errors and unifying data formats, performing preprocessing such as denoising and color enhancement on camera images, and performing ground segmentation and point cloud registration on point cloud data.

[0069] The image preprocessing uses a Gaussian filter to smooth the image, and the noise removal formula (1) is:

[0070]

[0071] Where σ is the standard deviation of the filter and k is the filter radius.

[0072] Point cloud preprocessing uses the RANSAC (random sampling consensus) algorithm to segment ground and non-ground points. The specific steps are to randomly select 3 points, construct a plane equation ax+by+cz+d=0, calculate the distance from other points to the plane, and satisfy The points are considered as internal points, repeated multiple times, and the plane with the most internal points is selected as the ground.

[0073] The radar target data is denoised, including using the Kalman filter to estimate the target position and velocity, as shown in formulas (2) and (3):

[0074] x k|k =x k|k―1 +K k (z k ―Hx k|k―1 ) (2)

[0075] K k =P k|k―1 H T (HP k|k―1 H T +R) ―1 (3)

[0076] Among them, x k|k is the state estimation, K k is the Kalman gain, H is the observation matrix, P k|k―1is the prediction error covariance matrix, and R is the measurement noise covariance.

[0077] The feature extraction and target detection submodule is used to detect and classify the preprocessed data using deep learning models (such as convolutional neural networks, YOLO, etc.), obtain target recognition results, and provide input data for the subsequent reliability quantification module. The above targets include pedestrians, vehicles, and traffic signs. The target recognition results include information such as the target category, location, speed, and direction.

[0078] Image object detection uses convolutional neural network (CNN) for object detection. Input image I, extract feature map F through multi-layer convolution operation, and finally output detection box and category: F = CNN (I) {(x, y, w, h, c)}, where (x, y) is the center of the detection box, w, h are the width and height, and c is the category confidence;

[0079] Point cloud target detection uses a clustering and shape matching-based method to detect objects. Through the DBSCAN (density clustering) algorithm, clustering forms different targets, as shown in formula (4):

[0080] DBSCAN(P,∈,MinPts)→{P cluster} (4)

[0081] Among them, ∈ is the maximum neighborhood distance, and MinPts is the minimum number of points to form a cluster.

[0082] 2: The hybrid quantization module is used to quantify the reliability of the target recognition results output by the environment perception module using information entropy and Bayesian uncertainty estimation methods, to form uncertainty assessment results for different targets and scenes, and to transmit the uncertainty assessment results to the decision and control module. The hybrid quantization module includes an information entropy calculation submodule, a Bayesian uncertainty estimation submodule, and a hybrid quantization result integration submodule.

[0083] The information entropy calculation submodule is used to calculate the confidence distribution of each target based on the target detection results output by the environment perception module. By calculating the entropy value of the confidence distribution, the confusion and uncertainty of target recognition are measured according to the entropy value. The higher the entropy value, the greater the uncertainty of target recognition. For each target detection result, the category confidence distribution p = {p1, p2, ..., p n}Calculate information entropy H(p): where p i is the confidence of the ith category.

[0084] The Bayesian uncertainty estimation submodule is used to use Bayesian neural networks or Monte Carlo dropout techniques to estimate the uncertainty of the target recognition results output by the environmental perception module, obtain the probability distribution of the model prediction, analyze the multiple prediction results for each target, calculate the Bayesian uncertainty metric, and evaluate the recognition reliability of the model under different circumstances. During the neural network inference process, Monte Carlo dropout calculates the mean μ and variance σ of multiple prediction results through multiple forward propagations (each time Dropout is enabled) 2 , as shown in formula (5):

[0085]

[0086] in, is the prediction result of the tth forward propagation, and T is the total number of forward propagations.

[0087] The hybrid quantization result integration submodule is used to integrate the results of information entropy and Bayesian uncertainty estimation, calculate the hybrid quantization index according to the preset weights or dynamically adjusted rules, determine the final uncertainty measurement value of each target, and transmit the final uncertainty measurement value of each target to the decision and control module to provide reliability input for the decision and control module. The above hybrid quantization index Q = α·H(p)+β·σ 2 , where α and β are preset or dynamically adjusted weight coefficients.

[0088] 3: Decision and control module, which is used to dynamically adjust the decision and control strategies of the autonomous driving system according to the final uncertainty measurement value of each target transmitted by the hybrid quantization module to achieve safe and efficient driving behavior. The decision and control module includes a strategy generation submodule, a real-time adjustment submodule and a control instruction execution submodule.

[0089] In the "strategy generation submodule", the autonomous driving system generates a corresponding driving strategy based on the final uncertainty measurement value of each target transmitted by the hybrid quantization module to achieve a balance between safety and efficiency. Depending on the level of uncertainty, the system adopts different driving strategies:

[0090] High uncertainty scenarios: When the uncertainty in the environment is high, such as interference with sensor data, complex road conditions, or other vehicles behaving erratically, the system generates conservative driving strategies, such as slowing down, stopping, or maintaining a larger distance between vehicles, to prioritize safety and avoid potential risks.

[0091] Low uncertainty scenario: When the system can identify the target with high confidence and the environment is relatively simple, such as clear roads and sparse traffic environments, the system generates aggressive driving strategies, such as moderate acceleration and lane change to overtake, to improve driving efficiency and driving smoothness.

[0092] In addition, in order to cope with scenarios with medium uncertainty or fluctuating complexity, the system generates robust driving strategies that balance safety and efficiency. For example, when encountering temporary obstacles or unpredictable traffic participants, the system will choose to maintain a moderate speed and steering range to ensure overall stability and adaptability. This robust strategy achieves a dynamic balance between safety and efficiency, allowing the system to flexibly adjust in different driving scenarios and improve the reliability of autonomous driving.

[0093] The driving strategy is generated according to the hybrid quantitative index Q and the pre-set threshold, as shown in formula (6):

[0094]

[0095] Among them, Q threshold1 and Q threshold2 Threshold

[0096] The real-time adjustment submodule is used to monitor environmental changes and sensor data in real time, dynamically adjust driving strategies, and automatically switch to safety mode when potential risks or increased uncertainty are detected to ensure driving safety;

[0097] The control command execution submodule is used to convert the generated driving strategy into specific control commands (such as steering wheel angle, acceleration, braking intensity, etc.), and send commands to various actuators of the vehicle (such as steering system, power system, braking system, etc.) through the electronic control unit (ECU) to complete the actual operation of the vehicle. The formula is as follows:

[0098] δ = f (S, environmental information), a = g (S, distance to the obstacle ahead), b = h (S, current vehicle speed)

[0099] The steering wheel angle δ is calculated by a function f, whose inputs include the system state (S) and environmental information. The system state may include the vehicle's location information, current direction, etc., and the environmental information may include road conditions, lane line position, etc. The acceleration a is calculated by a function g, whose inputs are the system state S and the distance to the obstacle ahead. The system state includes information about the vehicle's current state, and the distance to the obstacle ahead helps determine whether acceleration or deceleration is needed. The braking intensity b is calculated by a function h, whose inputs are the system state S and the current vehicle speed. The current vehicle speed is used to determine the necessity and intensity of braking to ensure safety.

[0100] 4: Feedback learning module, which is used to collect the data generated during driving and the mixed quantitative results of the system, conduct continuous learning and optimization, and improve the performance of the system in different scenarios. It includes data collection and analysis submodule, model optimization submodule and adaptive adjustment submodule.

[0101] The data collection and analysis submodule is used to continuously collect vehicle sensor data, environmental data, and decision-making data. Form a historical database and analyze the performance of the mixed quantitative results in each scenario, and compare it with the actual driving situation (such as whether a misjudgment or accident occurred). I (Input): Input information collected at time t, which may include environmental perception data, such as raw information from sensors such as cameras, lidars, and radars. P (Policy): Policy decision at time t, indicating the driving strategy selected by the autonomous driving system at that moment. R (Result): The actual result after executing the strategy at time t, such as whether the vehicle reaches the target safely, whether emergency braking is performed, etc. S (System State): The system state at time t, including the vehicle's current position, speed, direction, and other information, used to record the real-time status of the vehicle.

[0102] The model optimization submodule is used to continuously optimize the calculation method and weight coefficient of information entropy and Bayesian uncertainty estimation based on historical data and feedback results, and improve the adaptability and accuracy of the model in specific scenarios through online learning or incremental training, as shown in formula (7):

[0103]

[0104] in, is the loss function, is the mixed quantitative index for prediction, Q * is a true reliability indicator.

[0105] The adaptive adjustment submodule is used to update the decision and control strategies in real time according to the continuously optimized model and hybrid quantization system, improve the robustness of the system, and realize the adaptive adjustment capability of the system so that it can maintain good performance in unknown or unseen scenarios.

[0106] 5: Risk warning and protection module, which is used to proactively provide warning and protection when high uncertainty or potential risks are identified to ensure the safety of the system, including risk monitoring submodule, warning prompt submodule and protection strategy execution submodule. The risk monitoring submodule is used to monitor the mixed quantitative indicators. When the uncertainty exceeds the set threshold, it determines the existence of potential risks and detects abnormal situations or special scenarios (such as the sudden appearance of pedestrians, vehicle breakdown, etc.). According to the mixed quantitative result Q, the risk assessment threshold R is set. threshold , when Q ≥ R threshold When an alarm occurs, a warning signal is triggered.

[0107] The early warning submodule is used to provide risk warning prompts to the driver or passengers (such as through visual, auditory or tactile feedback) and prompt the system to take self-protection actions (such as deceleration and emergency braking);

[0108] The protection strategy execution submodule is used to automatically execute protection strategies in high-risk scenarios, such as deceleration, parking, emergency avoidance, etc., to ensure the safety of the vehicle and passengers and, when necessary, trigger the full system protection program in an emergency state to ensure the safety of the system in extreme situations. After the warning is triggered, the emergency operation parameters (such as emergency braking force) are calculated, as shown in formula (8):

[0109] b emergency =max(b current , b max ) (8)

[0110] Among them, b current is the current braking intensity, b max It is the maximum braking intensity in an emergency.

[0111] Example 1. In an urban driving environment, an autonomous vehicle enters a complex intersection. The system's environmental perception module collects real-time data from multiple cars, pedestrians, and non-motor vehicles. The information entropy calculation module performs entropy calculations on these data and obtains a high entropy value, indicating that the recognition result has a high uncertainty. The Bayesian uncertainty estimation module further confirms the high uncertainty of the perception module. The hybrid quantization model gives a high comprehensive quantitative index Q, based on which the decision optimization module adopts a conservative deceleration strategy and continuously monitors environmental changes until the reliability of the recognition result is improved.

[0112] Example 2. In a highway scenario, an autonomous vehicle is driving at night under poor lighting conditions. The perception module recognizes an object ahead, but the information entropy calculation module shows a low entropy value (i.e., high reliability), while the result given by the Bayesian uncertainty estimation module shows high uncertainty. The hybrid quantization model obtains a medium-valued comprehensive quantization index Q.

[0113] In summary, the embodiments of the present invention have the following beneficial effects compared with the prior art:

[0114] More accurate reliability quantification: This invention can more comprehensively evaluate the reliability of target recognition in autonomous driving scenarios by introducing a hybrid quantification model that combines information entropy calculation and Bayesian uncertainty estimation. Compared with a single uncertainty measurement method, the hybrid quantification model uses multiple indicators (such as confidence distribution and model prediction uncertainty) to construct a more accurate reliability measurement, improving adaptability to complex driving environments and recognition accuracy.

[0115] More flexible decision optimization: The real-time decision optimization module of the present invention dynamically adjusts the driving strategy based on the hybrid quantization results, and can adopt aggressive, robust or conservative driving modes in different scenarios. This multi-level decision mechanism is more flexible and adaptable than the traditional fixed strategy decision method, which helps to enhance the safety and responsiveness of the vehicle in scenarios with high uncertainty.

[0116] The present invention enhances the real-time decision-making ability of the system by calculating mixed quantitative indicators in real time and dynamically adjusting driving strategies, ensuring that the best decision can be made quickly in scenarios with high uncertainty, effectively improving the safety and stability of the autonomous driving system. In general, the present invention aims to significantly improve the performance of existing technologies in autonomous driving environments through a multi-dimensional, multi-level and real-time dynamic reliability evaluation method.

[0117] Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0118] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present invention or certain parts of the embodiments.

[0119] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0120] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. An automatic driving decision system based on information entropy-Bayesian estimation, characterized in that: include: Environmental perception module, hybrid quantization module and decision optimization module; The environment perception module is used to collect and process data from different sensors, perform target detection and classification after preprocessing the collected data, obtain target recognition results, and transmit the target recognition results to the hybrid quantization module; The hybrid quantization module is used to quantify the reliability of the target recognition result output by the environmental perception module using information entropy and Bayesian uncertainty estimation method, to form uncertainty evaluation results for different targets and scenes, and transmit the uncertainty evaluation results to the decision and control module; The decision and control module is used to dynamically adjust the decision and control strategy of the autonomous driving system according to the final uncertainty measurement value of each target transmitted by the hybrid quantization module.

2. The system according to claim 1, characterized in that The system further comprises: The feedback learning module is used to collect the data generated during the operation of the autonomous driving vehicle and the mixed quantitative results of the system, conduct continuous learning and optimization, and improve the performance of the autonomous driving decision-making system in different scenarios.

3. The system according to claim 2, characterized in that The system further comprises: The risk warning and protection module is used to provide warning and protection when high uncertainty or potential risks are identified, ensuring the safety of the autonomous driving decision-making system.

4. The system according to claim 1, 2 or 3, characterized in that: The environmental perception module includes: a data acquisition submodule, a data preprocessing submodule and a feature extraction and target detection submodule; The data acquisition submodule is used to collect environmental data in real time through the sensor array, transmit the collected raw data to the data preprocessing submodule, collect the RGB image sequence around the autonomous driving vehicle through the camera, and output the image data set I = {I t }, where t represents the time series, and the point cloud data P of the autonomous driving vehicle is collected by the lidar i =(x i ,y i , z i ), where p i The three-dimensional coordinates of the points scanned by the laser radar are collected through the millimeter-wave radar to collect the distance, speed and angle information of the autonomous driving vehicle to form radar target data R = (d i , v i ,θ i ), where d i is the target distance, v i is the relative speed, θ i is the angle; The data preprocessing submodule is used to preprocess the raw data transmitted from the data acquisition submodule, and transmit the preprocessed data to the feature extraction and target detection submodule; the image preprocessing uses a Gaussian filter to smooth the image, the point cloud data preprocessing uses a random sampling consistency RANSAC algorithm to segment the ground and non-ground points, and the radar target data is denoised, including using a Kalman filter to estimate the target position and speed; The feature extraction and target detection submodule is used to use the deep learning model to perform target detection and classification on the preprocessed data to obtain target recognition results, wherein the targets include pedestrians, vehicles and traffic signs, and the target recognition results include the category, position, speed and direction information of the target; Image object detection uses a convolutional neural network for object detection. The input image I is used to extract the feature map F through multi-layer convolution operations, and finally the detection box and category are output: F = CNN(I){(x, y, w, h, c)}, where (x, y) is the center of the detection box, w and h are the width and height, and c is the category confidence. Point cloud data target detection uses a clustering and shape matching-based method to detect objects, and clusters different targets through the density clustering DBSCAN algorithm, as shown in formula (4): DBSCAN(P,∈,MinPts)→{P cluster } (4) Among them, v is the maximum neighborhood distance, and MinPts is the minimum number of points to form a cluster.

5. The system according to claim 4, characterized in that The hybrid quantization module includes: an information entropy calculation submodule, a Bayesian uncertainty estimation submodule and a hybrid quantization result integration submodule; The information entropy calculation submodule is used to calculate the confidence distribution of each target according to the target detection result output by the environment perception module, and to measure the confusion and uncertainty of target recognition according to the entropy value by calculating the entropy value of the confidence distribution. n }Calculate information entropy H(p): where p i is the confidence of the i-th category; The Bayesian uncertainty estimation submodule is used to estimate the uncertainty of the target recognition results output by the environmental perception module to obtain the probability distribution of the model prediction, analyze the multiple prediction results of each target, calculate the Bayesian uncertainty metric, evaluate the recognition reliability of the model under different circumstances, and use Monte Carlo dropout to calculate the mean μ and variance σ of multiple prediction results through multiple forward propagations during the neural network reasoning process. 2 , as shown in formula (5): in, is the prediction result of the tth forward propagation, and T is the total number of forward propagations; The hybrid quantization result integration submodule is used to integrate the results of information entropy and Bayesian uncertainty estimation, calculate the hybrid quantization index according to the preset weight or dynamically adjusted rules, determine the final uncertainty measurement value of each target, and transmit the final uncertainty measurement value of each target to the decision and control module. The hybrid quantization index Q = α·H(p)+β·σ 2 , where α and β are preset or dynamically adjusted weight coefficients.

6. The system according to claim 5, characterized in that The decision-making and control module includes: an information entropy calculation submodule, a Bayesian uncertainty estimation submodule and a hybrid quantization result integration submodule; The strategy generation submodule is used to generate a driving strategy of the autonomous driving system according to the final uncertainty measurement value of each target transmitted from the hybrid quantization module, generate a conservative driving strategy in a high uncertainty scenario, and generate an aggressive driving strategy in a low uncertainty scenario. The driving strategy is generated according to the hybrid quantization index Q and a preset threshold, as shown in formula (6): Among them, Q threshold1 and Q threshold2 is the threshold value; The real-time adjustment submodule is used to monitor environmental changes and sensor data in real time, dynamically adjust driving strategies, and automatically switch to safety mode when potential risks or increased uncertainties are detected; The control instruction execution submodule is used to convert the driving strategy generated by the strategy generation submodule into a specific control instruction, and send the instruction to each actuator of the vehicle through the electronic control unit to complete the actual operation of the vehicle. The formula is as follows: δ=f(S, environmental information), a=g(S, distance of obstacle ahead), b=h(S, current vehicle speed). The steering wheel angle δ is calculated by a function f, the input of which includes the system state S and environmental information. The system state S includes the vehicle's position information and current direction, and the environmental information includes the road condition and lane line position. The acceleration a is calculated by a function g, the input of g is the system state S and the distance to the obstacle ahead, and the braking intensity b is calculated by a function h, the input of h is the system state S and the current vehicle speed.

7. The system according to claim 6, characterized in that The feedback learning module includes: a data collection and analysis submodule, a model optimization submodule and an adaptive adjustment submodule; The data collection and analysis submodule is used to continuously collect vehicle sensor data, environmental data and decision data. Form a historical database and analyze the performance of the hybrid quantization results in each scenario and compare it with the actual driving situation; I t is the input information collected at time t, P t is the strategy decision at time t, indicating the driving strategy selected by the autonomous driving system at that moment, R t is the actual result after executing the strategy at time t, S t is the system status at time t, including the current position, speed and direction information of the vehicle, which is used to record the real-time status of the vehicle; The model optimization submodule is used to continuously optimize the calculation method and weight coefficient of information entropy and Bayesian uncertainty estimation based on historical data and feedback results, and improve the adaptability and accuracy of the autonomous driving system model in specific scenarios through online learning or incremental training, as shown in formula (7): in, is the loss function, is the mixed quantitative index for prediction, Q * is a true reliability indicator; The adaptive adjustment submodule is used to update the decision and control strategies in real time according to the continuously optimized autonomous driving system model and hybrid quantization system.

8. The system according to claim 7, characterized in that The risk warning and protection module includes: a risk monitoring submodule, a warning prompt submodule and a protection strategy execution submodule; The risk monitoring submodule is used to monitor the mixed quantitative indicators. When the uncertainty exceeds the set threshold, it is determined that there is a potential risk. The risk assessment threshold R is set according to the mixed quantitative result Q. threshold , when Q ≥ R threshold When the warning signal is triggered; The warning prompt submodule is used to provide risk warning prompts to the driver or passengers, prompting the automatic driving system to perform self-protection actions; The protection strategy execution submodule is used to automatically execute the protection strategy in a high-risk scenario, trigger the full system protection program in an emergency state, and calculate the emergency operation parameters after triggering the warning, as shown in formula (8): b emergency =max(b current , b max ) (8) Where b current is the current braking intensity, b max It is the maximum braking intensity in an emergency.

Images