Vehicle automatic driving control method fusing belief update and drift diffusion model

Abstract

The application provides a vehicle automatic driving control method fusing belief update and drift diffusion model, adopts a pre-trained multi-modal neural network to realize accurate inference on priori belief of pedestrians, introduces Bayes theorem and surprise amount calculation, quantifies difference between vehicle behavior and priori belief expected by pedestrians in real time, solves the problem of insufficient perception on internal state of pedestrians, finally, adopts a pre-trained drift diffusion model to simultaneously output decision tendency and decision time of pedestrians, provides key timing information for vehicle control, solves safety risks caused by prediction lag, avoids unnecessary parking and deceleration, and can timely trigger emergency avoidance, thereby ensuring safety and significantly improving traffic efficiency.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and more specifically, to an autonomous driving control method for vehicles that integrates belief updates and drift diffusion models. Background Technology

[0002] With the development of autonomous driving technology, the interaction between autonomous vehicles and pedestrians in scenarios without clear right-of-way (such as intersections without traffic lights) has become a key challenge for the industry. Existing control strategies for autonomous vehicles are mostly based on explicit traffic rules or predictive models of pedestrian external behaviors (such as position and speed).

[0003] However, in the real world, pedestrian behavior is highly unpredictable, especially in scenarios without clear right-of-way. Pedestrians are prone to complex behaviors such as hesitation, backing away, or sudden crossing due to unexpected vehicle actions (such as abnormal deceleration). Current autonomous driving systems struggle to deeply understand the underlying decision-making logic behind these pedestrian behaviors, leading to the following two major shortcomings:

[0004] 1. Predictions based on external trajectories have lag and bias. When pedestrian behavior is sudden, the system may not be able to react in time, leading to the risk of collision.

[0005] 2. To avoid risks, the system often adopts an overly conservative strategy (such as stopping too early or too frequently), which not only reduces traffic efficiency, but its ambiguous intention may also further exacerbate pedestrian confusion, creating a vicious cycle. Summary of the Invention

[0006] The technical problem this invention aims to solve is how to provide an autonomous driving control method that can understand the internal decision-making logic of pedestrians, thereby achieving safer and more efficient human-vehicle interaction.

[0007] This invention provides a vehicle autonomous driving control method that integrates belief update and drift diffusion models, comprising:

[0008] Step 1: Real-time collection of environmental information, pedestrian movement information, and vehicle operation information by onboard sensors to form an observation dataset; calculation of the current observation vector based on pedestrian movement information and vehicle operation information. ;

[0009] Step 2: By inputting the observation dataset from the previous time step into a pre-trained multimodal neural network, the prior beliefs used to characterize pedestrians' predictions of vehicle behavior are calculated and output.

[0010] Step 3: Use Bayes' theorem to calculate the vector representing the current observation. The degree of surprise at the inconsistency with prior beliefs, and the type of surprise is indicated;

[0011] Step 4: Input the type of surprise and the amount of surprise into the pre-trained drift-diffusion model, calculate and output the pedestrian decision and decision time;

[0012] Step 5: The vehicle generates a vehicle control strategy based on the pedestrian's decision and the decision time, and generates and executes control commands based on the vehicle control strategy.

[0013] Compared with existing technologies, this application has the following advantages: It uses a multimodal neural network to accurately infer the prior beliefs of pedestrians, digitizes the pedestrians' internal decision-making data, and then introduces Bayes' theorem and surprise quantity calculation to quantify the difference between vehicle behavior and the pedestrians' expected prior beliefs in real time, solving the problem of insufficient perception of the pedestrians' internal state. Finally, it uses a pre-trained drift-diffusion model to simultaneously output the pedestrian's decision and decision time, providing key time-series information for vehicle control, solving the safety risks caused by prediction lag, avoiding unnecessary stopping and deceleration, and triggering emergency avoidance in a timely manner, thus significantly improving the traffic efficiency of vehicles while ensuring safety.

[0014] In one possible implementation, the environmental information includes the interaction scene, traffic density, and road lane line information;

[0015] The pedestrian movement information includes the pedestrian's coordinates. Walking speed acceleration Body posture and direction of gaze;

[0016] The vehicle operation information includes the vehicle's coordinates. ,speed acceleration Yaw rate and turn signal status.

[0017] In one possible implementation, the current observation vector in step 1 Including the relative distance between pedestrians and vehicles Relative velocity Potential collision time Velocity change range and the magnitude of relative distance change The calculation formula is:

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] ;

[0023] Get the current time Current observation vector .

[0024] Compared with existing technologies, by clearly defining the key parameters of the current observation vector, the dynamic characteristics of the interaction scene can be fully captured, providing a sufficient information basis for subsequent pedestrian decision-making and decision time calculation, thereby improving the accuracy and reliability of the drift diffusion model.

[0025] In one possible implementation, by inputting the observation dataset from the previous time step into a pre-trained multimodal neural network, the expression used to characterize the prior belief of pedestrians in predicting vehicle behavior is calculated and output as follows:

[0026] ;

[0027] ;

[0028] In the formula, This represents the set of predictions made by pedestrians regarding vehicle behavior. Indicates the first Such vehicle behavior, Indicated to The historical observation dataset includes environmental information and vehicle operation information from time 1 to the previous time. Indicates in A priori beliefs about time.

[0029] Compared with existing technologies, data fusion through multimodal neural networks can more comprehensively model the cognitive state of pedestrians in different scenarios, fully consider the complex interactions between environmental factors and vehicle behavior, and improve the accuracy of prior belief inference.

[0030] In one possible implementation, step 3 specifically includes:

[0031] Step 301, set vehicle behavior as ,based on Prior beliefs at any given moment and current observation vector The current observation vector is calculated using Bayes' theorem. The prediction probability is calculated using the following formula:

[0032] ;

[0033] In the formula, These represent the parameters of a multimodal neural network.

[0034] Step 302, calculate the probability at time [time value missing] based on the predicted probability. The amount of surprise is calculated using the following formula:

[0035] ;

[0036] Step 303: Using the vehicle operation information at the current moment and the current observation vector... Add a surprise type label.

[0037] Compared with existing technologies, the probability-based surprise calculation formula can objectively and quantitatively assess the degree of difference between the current observation and the pedestrian's expectations, providing a reliable driving signal for the drift diffusion model and ensuring the accuracy of decision prediction.

[0038] In one possible implementation, vehicle behavior includes deceleration, acceleration, starting, and stopping.

[0039] In one possible implementation, the types of surprise include abnormal deceleration, abnormal acceleration, trajectory deviation, and start-stop hesitation.

[0040] In one possible implementation, the pre-training process of the drift diffusion model in step 4 specifically includes:

[0041] A. Obtain a historical interaction dataset, which includes multiple historical interaction samples. Each historical interaction sample includes: surprise level. Surprise type Pedestrian decision-making and decision time ;

[0042] B, Construct a drift-diffusion model, using the quantity of surprise. Surprise type Pedestrian decision-making based on input and decision time To output a training drift-diffusion model, the drift-diffusion model models the pedestrian's decision-making process as an evidence accumulation process. The drift-diffusion model is described by stochastic differential equations:

[0043] ;

[0044] In the formula, Indicates entering the initial time; This represents the pedestrian's decision to either proceed or yield. Indicates surprise. Represents the drift rate function. The parameters of the drift rate function are... Where is the diffusion coefficient. ; For standard Wiener process;

[0045] The parameter vector of the drift-diffusion model Includes: drift rate function Decision threshold Starting point deviation Non-decision time ;

[0046] C, given parameter vector The maximum likelihood estimation method is used to fit the parameters of the drift-diffusion model. The Fokker-Planck equation corresponding to the stochastic differential equation is solved numerically to obtain the parameter vector. and input driver First arrival time probability density function Obtain historical interaction samples joint likelihood :

[0047] ;

[0048] D, the log-likelihood form obtained by multiplying over all historical interaction samples, is expressed as:

[0049] ;

[0050] E is obtained by fitting the drift rate function through maximum likelihood estimation. Decision threshold Starting point deviation Non-decision time The optimal estimate:

[0051] ;

[0052] In the formula, This represents the optimal parameter vector.

[0053] Compared with existing technologies, by training the drift diffusion model based on real data, the model can accurately capture the decision-making patterns of pedestrians, while predicting pedestrian decisions and decision times, providing a more comprehensive and reliable basis for vehicle control.

[0054] In one possible implementation, the drift rate function Represented as a linear function of surprise quantity and surprise type:

[0055] ;

[0056] in, , Represents a learnable parameter associated with the type of surprise;

[0057] The decision threshold Represented as a scene variable function: ;

[0058] The starting point deviation Represented as the belief distribution function: ;

[0059] in, Indicates the initial time The resulting a priori beliefs; Represents the belief distribution function Parameters; Represents scene variable function Parameters;

[0060] Non-decision time satisfy .

[0061] Compared with existing technologies, by incorporating both the type and intensity of surprise into the drift rate calculation, differentiated responses to abnormal behaviors of different natures are achieved, enabling the drift diffusion model to simulate pedestrian decision-making processes more precisely and improve prediction accuracy.

[0062] In one possible implementation, the vehicle control strategy in step 5 includes a forced yield strategy, a traffic hold strategy, a gradual yield strategy, and an emergency avoidance strategy.

[0063] Compared with existing technologies, by designing multi-level and differentiated control strategies, vehicles can select the most appropriate interaction strategy based on pedestrians' real-time cognitive state and decision-making tendencies, thereby optimizing traffic efficiency while ensuring safety. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating the process of this application;

[0065] Figure 2 This is a schematic diagram illustrating pedestrian decision-making based on prior belief output generated from acquired images using the methods described in this application. Detailed Implementation

[0066] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0067] In the description of the embodiments of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application based on the specific circumstances.

[0068] In the embodiments of this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0069] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0070] See Figure 1 As shown in the figure, this application discloses a vehicle autonomous driving control method that integrates belief update and drift diffusion models, aiming to solve two core problems raised in the background art: insufficient perception of pedestrians' internal state and the imbalance between safety and efficiency caused by prediction lag. Specifically, it includes the following steps:

[0071] Step 1: The vehicle uses its onboard sensors to collect environmental information, pedestrian movement information, and vehicle operation information in real time to form an observation dataset. Based on the pedestrian movement information and vehicle operation information, a current observation vector is calculated. ;

[0072] This application embodiment uses an onboard LiDAR (such as Velodyne VLP-16), millimeter-wave radar (such as Continental ARS408), a forward-facing camera (such as Sony IMX490), and a GPS / IMU integrated navigation system to collect the current time. The system collects environmental information, pedestrian movement information, and vehicle operation information; it uses LiDAR point cloud data and PointPillars algorithm for pedestrian detection and tracking, and YOLOv7 for pedestrian posture and gaze direction recognition based on images captured by cameras; millimeter-wave radar provides speed compensation and obstacle confirmation.

[0073] The environmental information includes the interaction scenario, traffic density, and road lane information. By acquiring information such as the interaction scenario and traffic density, the vehicle can understand the macro background of the current interaction, which is an important context for judging the pedestrian's expected behavior. For example, at intersections with high traffic density, pedestrians may be more inclined to wait cautiously.

[0074] The pedestrian movement information includes the pedestrian's coordinates. Walking speed acceleration And body posture and gaze direction identified through computer vision algorithms; gaze direction is particularly crucial, as it can partially reflect the degree of pedestrian attention to vehicles and is the data starting point for this method to go beyond simple trajectory prediction and gain insight into the internal state.

[0075] The vehicle operation information includes the vehicle's coordinates. ,speed acceleration Yaw rate and turn signal status.

[0076] Construct the current observation vector based on pedestrian movement information and vehicle operation information. The current observation vector is generated every 100ms. Including the relative distance between pedestrians and vehicles Relative velocity Potential collision time Velocity change range and the magnitude of relative distance change The calculation formula is:

[0077] ;

[0078] ;

[0079] ;

[0080] ;

[0081] ;

[0082] Get the current time Current observation vector Current observation vector The design specifically captures the core of dynamic safety in human-vehicle interaction, including relative distance. and relative velocity It is the basic geometric measure of collision risk; potential collision time Combining distance and speed into an intuitive time-based risk indicator; speed variation range and the magnitude of relative distance change This is used to capture the dynamic trends of interactions, directly serving to detect whether vehicle behavior is "unexpected by pedestrians"; for example, a sudden negative... (Sudden deceleration) may surprise pedestrians.

[0083] Step 2: By inputting the observation dataset from the previous time step into a pre-trained multimodal neural network, the prior beliefs used to characterize pedestrian predictions of vehicle behavior are calculated and output; the expression is:

[0084] ;

[0085] ;

[0086] In the formula, This represents the set of predictions made by pedestrians regarding vehicle behavior. Indicates the first Such vehicle behavior, Indicated to The historical observation dataset includes environmental information and vehicle operation information from time 1 to the previous time. Indicates in Pedestrian prediction at any time The probability distribution of prior beliefs; pedestrians' predicted vehicle behaviors include deceleration, acceleration, starting, and stopping.

[0087] The multimodal neural network in this application adopts a hierarchical encoder structure, and designs an environmental modal encoder, a vehicle motion modal encoder, and a pedestrian subjective prediction modal encoder according to the data characteristics of the environment, vehicles, and pedestrians. The environmental modal encoder uses a multilayer perceptron structure, including fully connected layers, ReLU activation functions, and dropout layers, to map environmental information to a 64-dimensional feature space. The vehicle motion modal encoder adopts a two-layer LSTM network with a hidden layer dimension of 128, which is used to capture the temporal dependencies in vehicle motion information. The pedestrian subjective prediction modal encoder combines a convolutional neural network and a long short-term memory network. First, it extracts local features through convolutional layers, and then models temporal relationships through LSTM layers. Next, it integrates the feature vectors output by the environmental modal encoder, the vehicle motion modal encoder, and the pedestrian subjective prediction modal encoder through a fusion layer. The fusion layer adopts a fully connected network structure, concatenating the features of the three modalities and mapping them to a 256-dimensional joint feature space. Finally, it generates probability distributions for multiple predicted vehicle behaviors through the softmax function of the output layer. In this embodiment, the pedestrian predicted vehicle behaviors include continuing forward, slowing down to yield, stopping completely, and passing through the intersection at a constant speed.

[0088] In the pre-training stage of the multimodal neural network, this embodiment first performs large-scale data collection and processing. The data sources include publicly available human-vehicle interaction datasets and self-collected datasets, totaling more than 5,000 sets of valid interaction samples. Each sample contains data from three modalities: environmental information, vehicle motion sequences, and pedestrian behavior characteristics.

[0089] During data processing, environmental information undergoes one-hot encoding and normalization, transforming scene type, traffic density, and road conditions into standardized feature vectors. For vehicle operation information, position, velocity, acceleration, and yaw rate data are collected continuously over 3 seconds at a frequency of 10Hz, forming a time-series sequence. For pedestrian subjective prediction modalities, computer vision algorithms extract pedestrian gaze direction, body orientation, and historical trajectory features. The gaze direction uses one-hot encoding, the body orientation is converted into continuous angle values, and the historical trajectory retains the position sequence from the past 2 seconds. All training samples are manually labeled by a professional annotation team, with annotations indicating the pedestrian's expected category of vehicle behavior, including continuing forward, slowing down to yield, stopping completely, and passing through the intersection at a constant speed. To ensure annotation quality, cross-validation by multiple independent annotators is employed, and the final annotation results must achieve a Kappa coefficient greater than 0.85 for consistency.

[0090] The training process employs a phased strategy. First, each modal encoder is pre-trained individually, followed by end-to-end joint training. The single-modal pre-training phase uses a high learning rate to quickly converge the parameters of each dedicated encoder. The joint training phase uses the AdamW optimizer with an initial learning rate of 0.001, coupled with a cosine annealing learning rate scheduler to dynamically adjust the learning rate during training.

[0091] The loss function employs weighted cross-entropy loss, assigning higher weights to classes with fewer samples based on the class distribution of the training data to address the data imbalance problem. Simultaneously, an L2 regularization term is included in the loss function to control model complexity and prevent overfitting. An early stopping strategy is used during training; training automatically terminates when the validation set performance shows no improvement for 20 consecutive epochs, preserving the optimal model parameters.

[0092] To meet practical deployment requirements, the trained model is compressed and accelerated. Knowledge distillation is employed, using a pre-trained large-scale teacher model to guide the training of a lightweight student model, reducing the number of model parameters while maintaining performance. Simultaneously, inference engines such as TensorRT are used to quantize and optimize the model, supporting FP16 mixed-precision computation and significantly improving inference speed.

[0093] The optimized model is deployed via onboard computing units and integrated into the real-time processing pipeline of the autonomous driving system. During actual operation, interaction data is continuously collected and the model is updated periodically, with incremental learning used to continuously improve the model's adaptability and generalization ability. This complete pre-training process ensures that the multimodal neural network can accurately understand pedestrians' expectations of vehicle behavior in different traffic scenarios, providing a reliable prior information foundation for subsequent belief updates and decision predictions.

[0094] Step 3: Use Bayes' theorem to calculate the vector representing the current observation. The degree of surprise at the inconsistency with prior beliefs, and the type of surprise is indicated; specifically including:

[0095] Step 301, set vehicle behavior as Based on the prior beliefs of the previous moment and the current observation vector The current observation vector is calculated using Bayes' theorem. The prediction probability is calculated using the following formula:

[0096] ;

[0097] In the formula, These represent the parameters of a multimodal neural network.

[0098] Step 302, calculate the probability at time [time value missing] based on the predicted probability. The amount of surprise is calculated using the following formula:

[0099] ;

[0100] According to information theory, the value of surprise measures the amount of unexpected information carried by the current observation vector; the lower the probability, the greater the surprise.

[0101] Step 303: Using the vehicle operation information at the current moment and the current observation vector... The type of surprise is labeled; in this embodiment, the types of surprise in vehicle behavior include abnormal deceleration, abnormal acceleration, trajectory deviation, and hesitation in starting or stopping; for example, the magnitude of the vehicle's speed change at the current moment. If the acceleration is less than 0 and exceeds the preset comfortable deceleration threshold, the type of surprise is marked as abnormal deceleration; or the magnitude of the change in velocity at the current moment. When the acceleration exceeds the preset acceleration threshold, the type of surprise is marked as abnormal acceleration; when the deviation between the vehicle's yaw rate and the expected path exceeds the preset tolerance threshold, the type of surprise is marked as trajectory cheap; when the vehicle's acceleration repeatedly switches between start and stop, it is marked as start-stop hesitation.

[0102] By measuring the difference between vehicle behavior and pedestrian expectations in real time and quantitatively, the surprise level is... The magnitude of the surprise directly reflects the degree to which a pedestrian's cognition has been subverted, while the type of surprise indicates the source of the subversion. This not only tells us that the pedestrian is surprised, but also what kind of surprise it is, thus laying the foundation for more refined decision-making and prediction.

[0103] Step 4: Input the type and magnitude of surprise into the pre-trained drift-diffusion model, and output the pedestrian decision and decision time; the pre-training process of the drift-diffusion model specifically includes:

[0104] A. Obtain a historical interaction dataset, which includes multiple historical interaction samples. Each historical interaction sample includes: surprise level. Surprise type Pedestrian decision-making and decision time ;

[0105] B, Construct a drift-diffusion model, using the quantity of surprise. Surprise type Pedestrian decision-making based on input and decision time To output a training drift-diffusion model, the drift-diffusion model models the pedestrian's decision-making process as an evidence accumulation process. The drift-diffusion model is described by stochastic differential equations:

[0106] ;

[0107] In the formula, Indicates entering the initial time; This represents the pedestrian's decision to either proceed or yield. Indicates surprise. Represents the drift rate function. The parameters of the drift rate function are... Where is the diffusion coefficient. ; For standard Wiener process;

[0108] The parameter vector of the drift-diffusion model Includes: drift rate function Decision threshold Starting point deviation Non-decision time The drift rate function Represented as a linear function of surprise quantity and surprise type:

[0109] ;

[0110] in, , Represents a learnable parameter associated with the type of surprise;

[0111] The decision threshold Represented as a scene variable function: ;

[0112] The starting deviation Represented as the belief distribution function: ;

[0113] in, Indicates the initial time The resulting a priori beliefs; Represents the belief distribution function Parameters; Represents scene variable function Parameters;

[0114] Non-decision time satisfy .

[0115] C, given parameter vector The maximum likelihood estimation method is used to fit the parameters of the drift-diffusion model. The Fokker-Planck equation corresponding to the stochastic differential equation is solved numerically to obtain the parameter vector. and input driver First arrival time probability density function Obtain historical interaction samples joint likelihood :

[0116] ;

[0117] D, the log-likelihood form obtained by multiplying over all historical interaction samples, is expressed as:

[0118] ;

[0119] E is obtained by fitting the drift rate function through maximum likelihood estimation. Decision threshold Starting point deviation Non-decision time The optimal estimate:

[0120] .

[0121] Different types of surprise, combined with different amounts of surprise, differentially affect the speed (drift rate) and direction of evidence accumulation.

[0122] Step 5: The vehicle generates a vehicle control strategy based on the pedestrian's decision and decision time. The vehicle then generates and executes control commands based on this strategy, aiming to dynamically balance safety and efficiency. The vehicle control strategies include a forced yield strategy, a pass-keeping strategy, a gradual yield strategy, and an emergency avoidance strategy. The forced yield strategy involves the vehicle slowing down in advance, including slowing to a stop. Specifically, when the predicted pedestrian decision is "to cross" and the decision time is less than a preset first threshold, it indicates that the pedestrian is determined to cross quickly. To ensure absolute safety, the vehicle should clearly warn the pedestrian and decisively slow down or stop, conveying an undisputed intention to yield.

[0123] The traffic-keeping strategy involves vehicles maintaining a constant speed or slightly decelerating to pass through the interaction zone on a stable trajectory, avoiding unnecessary stops and improving traffic efficiency. Specifically, when the predicted pedestrian decision is "yield" and the decision time is higher than a preset second threshold, it indicates that the pedestrian has clearly agreed to wait and there is no imminent risk. Vehicles can maintain or fine-tune their speed to pass through the interaction zone efficiently on a stable trajectory. This directly avoids unnecessary stops and traffic flow interruptions caused by excessive conservatism, significantly improving traffic efficiency.

[0124] The gradual yielding strategy involves slowing down slightly to create more time for observation and interaction. Specifically, when the prediction result is uncertain or the decision time is between a preset first threshold and a second threshold, the vehicle slows down gently and gradually. This aims to create additional time for observation and interaction, while releasing clear friendly signals to pedestrians through smooth and predictable vehicle dynamics, prompting them to make clear decisions, thereby effectively breaking the interaction deadlock of "vehicle hesitation causing pedestrian hesitation".

[0125] The emergency avoidance strategy is to immediately execute maximum safety braking or choose to detour, prioritizing pedestrian safety. Specifically, when it is predicted that the pedestrian's decision time is lower than the preset emergency threshold and accompanied by a high-risk type of surprise, the highest priority response is immediately triggered. This strategy prioritizes ensuring the absolute safety of pedestrians, and the execution methods include activating maximum safety braking force or performing an emergency turn and detour if there is a safe space.

[0126] like Figure 2 The diagram illustrates the application of this embodiment of the invention in a typical traffic light-free intersection scenario. It shows that as a vehicle approaches the intersection, onboard sensors (including LiDAR, cameras, millimeter-wave radar, etc.) detect in real time that a pedestrian is positioned on one side of the intersection and intends to cross. Based on the analysis of historical observation data using a multimodal neural network, it is inferred that the pedestrian's prior belief regarding vehicle behavior is "pedestrian priority," meaning the pedestrian expects the vehicle to slow down or stop to yield.

[0127] Subsequently, the system uses the observation vector at the current moment. The system calculates the surprise level and identifies the surprise type. In this embodiment, since the vehicle approaches the intersection at a relatively high speed without slowing down significantly, the system determines that the current vehicle behavior is significantly different from the pedestrian's expectations, and the surprise type is marked as "abnormal failure to slow down", with a high surprise level.

[0128] The surprise quantity and surprise type are fed into a pre-trained drift-diffusion model. The drift-diffusion model calculates the pedestrian's decision tendency as "give way" based on the input signal and predicts that their decision time is short. Based on this, a vehicle control strategy is generated: since the pedestrian is predicted to decide to cross quickly and there is a safety risk, a gradual yielding strategy is selected to control the vehicle to decelerate smoothly, convey a clear intention to yield to the pedestrian, and avoid sudden braking that may cause discomfort or rear-end collision risk.

[0129] Figure 2 This demonstrates how the method described in this application achieves safe, efficient, and explainable human-vehicle interaction in real-world road conditions. Especially in scenarios without clear right-of-way, the system can make intelligent control decisions that both ensure safety and improve traffic efficiency by understanding the pedestrian's cognitive state and predicting their decision-making sequence.

[0130] In the description of the embodiments of this application, it should be noted that the terms "inner" and "outer" and other terms indicating direction or positional relationship are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or component must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this application.

[0131] In the description of this application, the references to terms such as "an embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0132] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A vehicle autonomous driving control method integrating belief update and drift diffusion models, characterized in that, include: Step 1: Real-time collection of environmental information, pedestrian movement information, and vehicle operation information by onboard sensors to form an observation dataset; calculation of the current observation vector based on pedestrian movement information and vehicle operation information. ; Step 2: By inputting the observation dataset from the previous time step into a pre-trained multimodal neural network, the prior beliefs used to characterize pedestrians' predictions of vehicle behavior are calculated and output. Step 3: Use Bayes' theorem to calculate the vector representing the current observation. The degree of surprise at the inconsistency with prior beliefs, and the type of surprise is indicated; Specifically, it includes: Step 301, set vehicle behavior as ,based on Prior beliefs at any given moment and current observation vector The current observation vector is calculated using Bayes' theorem. The prediction probability is calculated using the following formula: ； In the formula, The parameters represent those of a multimodal neural network. Indicates in A priori beliefs of time This represents the set of predictions made by pedestrians about vehicle behavior. Step 302, calculate the probability at time [time value missing] based on the predicted probability. The amount of surprise is calculated using the following formula: ； Step 303: Using the vehicle operation information at the current moment and the current observation vector... Add a surprise type label; Surprise types include abnormal deceleration, abnormal acceleration, trajectory deviation, and start-stop hesitation; Step 4: Input the type of surprise and the amount of surprise into the pre-trained drift-diffusion model, calculate and output the pedestrian decision and decision time; Step 5: The vehicle generates a vehicle control strategy based on the pedestrian's decision and the decision time, and generates and executes control commands based on the vehicle control strategy.

2. The vehicle autonomous driving control method based on the fusion of belief update and drift diffusion models according to claim 1, characterized in that, The environmental information includes the interaction scenario, traffic density, and road lane information; The pedestrian movement information includes the pedestrian's coordinates. Walking speed acceleration Body posture and direction of gaze; The vehicle operation information includes the vehicle's coordinates. ,speed acceleration Yaw rate and turn signal status.

3. The vehicle autonomous driving control method based on the fusion of belief update and drift diffusion models according to claim 2, characterized in that, Current observation vector in step 1 Including the relative distance between pedestrians and vehicles Relative velocity Potential collision time Velocity change range and the magnitude of relative distance change The calculation formula is: ； ； ； ； ； Get the current time Current observation vector .

4. The vehicle autonomous driving control method based on the fusion of belief update and drift diffusion models according to claim 3, characterized in that, By inputting the observation dataset from the previous time step into a pre-trained multimodal neural network, the expression for representing the prior belief of pedestrians in predicting vehicle behavior is calculated and output as follows: ； ； In the formula, This represents the set of predictions made by pedestrians regarding vehicle behavior. Indicates the first Such vehicle behavior, Indicated to The historical observation dataset at any given time includes environmental information and vehicle operation information up to the previous time point. Indicates in A priori beliefs about time.

5. The vehicle autonomous driving control method based on the fusion of belief update and drift diffusion models according to claim 1, characterized in that, Vehicle behavior includes deceleration, acceleration, starting, and stopping.

6. The vehicle autonomous driving control method based on the fusion of belief update and drift diffusion models according to claim 1, characterized in that, The pre-training process of the drift-diffusion model in step 4 specifically includes: A. Obtain a historical interaction dataset, which includes multiple historical interaction samples. Each historical interaction sample includes: surprise level. Surprise type Pedestrian decision-making and decision time ; B, Construct a drift-diffusion model, using the quantity of surprise. Surprise type Pedestrian decision-making based on input and decision time To output a training drift-diffusion model, the drift-diffusion model models the pedestrian's decision-making process as an evidence accumulation process. The drift-diffusion model is described by stochastic differential equations: ； In the formula, Indicates entering the initial time; This represents the pedestrian's decision to either proceed or yield. Indicates surprise. Represents the drift rate function. The parameters of the drift rate function are... The diffusion coefficient is... ; For standard Wiener process; The parameter vector of the drift-diffusion model Includes: drift rate function Decision threshold Starting point deviation Non-decision time ; C, given parameter vector The maximum likelihood estimation method is used to fit the parameters of the drift-diffusion model. The Fokker-Planck equation corresponding to the stochastic differential equation is solved numerically to obtain the parameter vector. and input driver First arrival time probability density function Obtain historical interaction samples joint likelihood : ； D, the log-likelihood form obtained by multiplying over all historical interaction samples, is expressed as: ； E is obtained by fitting the drift rate function through maximum likelihood estimation. Decision threshold Starting point deviation Non-decision time The optimal estimate: ； In the formula, This represents the optimal parameter vector.

7. The vehicle autonomous driving control method based on the fusion of belief update and drift diffusion models according to claim 6, characterized in that, The drift rate function Represented as a linear function of surprise quantity and surprise type: ； in, , Represents a learnable parameter associated with the type of surprise; The decision threshold Represented as a scene variable function: ; The starting deviation Represented as the belief distribution function: ; in, Indicates the initial time The resulting a priori beliefs; Represents the belief distribution function Parameters; Represents scene variable function Parameters; Non-decision time satisfy .

8. The vehicle autonomous driving control method based on the fusion of belief update and drift diffusion models according to claim 1, characterized in that, The vehicle control strategies in step 5 include the forced yield strategy, the traffic hold strategy, the gradual yield strategy, and the emergency avoidance strategy.

Images