A vehicle trajectory prediction method, system, device and storage medium

Abstract

The application discloses a vehicle trajectory prediction method, system, device and storage medium, and relates to the technical field of intelligent driving. The method comprises the following steps: acquiring a first motion state of a vehicle; constructing a first network model, and obtaining a first trajectory prediction result based on the first motion state of the vehicle; acquiring a first fusion feature of a driver; processing the first fusion feature of the driver according to a second network model to obtain a second trajectory prediction result; performing uncertainty evaluation on the first trajectory prediction result to obtain first uncertainty data, and performing uncertainty evaluation on the second trajectory prediction result to obtain second uncertainty data; acquiring a first vehicle speed of the vehicle, and determining a prediction strategy according to the first uncertainty data, the second uncertainty data and the first vehicle speed of the vehicle; and determining a target trajectory prediction result of the vehicle according to the prediction strategy. The method realizes full-time-domain and high-robustness trajectory prediction, and fills the gap of single model in complex scene adaptation.

Description

Technical Field

[0001] This application relates to the field of intelligent driving technology, and in particular to a vehicle trajectory prediction method, system, device and storage medium. Background Technology

[0002] Currently, physics-based methods are one of the mainstream approaches for vehicle trajectory prediction. These methods take the vehicle's current motion state as input and extrapolate its trajectory over a short period using pre-defined physical formulas. Their core principle is to leverage the inertial characteristics and mechanical constraints of vehicle motion. In scenarios with high determinism (such as uniform straight-line driving or gentle turns), they can output high-precision prediction results and possess clear physical interpretability, making them easy to implement in engineering.

[0003] However, trajectory prediction methods based on physical models rely solely on the vehicle's instantaneous motion state. When the prediction time window is extended or the scenario complexity increases, the prediction error of the physical model will accumulate sharply, and may even result in a complete deviation from the actual driving behavior, making it difficult to meet the reliability requirements of intelligent driving systems for long-term trajectory prediction. Summary of the Invention

[0004] This application provides a vehicle trajectory prediction method, system, device, and storage medium. By integrating short-term physical models with long-term multimodal deep learning, it achieves high-precision trajectory prediction across the entire time domain, combining instantaneous accuracy with long-term predictability. The online uncertainty gating mechanism and deep perception of driver emotions in this solution enhance the safety, reliability, and scene understanding capabilities of autonomous driving.

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] Firstly, this application provides a vehicle trajectory prediction method, which includes:

[0007] The process involves: acquiring the vehicle's first motion state; constructing a first network model and obtaining a first trajectory prediction result based on the vehicle's first motion state; acquiring the driver's first fusion feature; processing the driver's first fusion feature to obtain a second trajectory prediction result; performing uncertainty assessment on the first trajectory prediction result to obtain first uncertainty data; performing uncertainty assessment on the second trajectory prediction result to obtain second uncertainty data; acquiring the vehicle's first speed; determining a prediction strategy based on the first uncertainty data, the second uncertainty data, and the vehicle's first speed; and determining the vehicle's target trajectory prediction result based on the prediction strategy.

[0008] In some possible implementations, constructing the first network model includes:

[0009] An initial network model is constructed, which is coupled with lateral motion to obtain the first network model.

[0010] In some possible implementations, obtaining the first trajectory prediction result based on the vehicle's first motion state includes:

[0011] The first motion state is input into the first network model to obtain a first initial trajectory prediction result; the first initial trajectory prediction result is corrected to obtain a first trajectory prediction result.

[0012] In some possible implementations, the correction of the first initial predicted trajectory result includes:

[0013] Obtain the lateral acceleration measured by the vehicle sensors, and calculate the difference between the model-predicted lateral acceleration and the measured lateral acceleration corresponding to the first initial trajectory prediction result;

[0014] The difference is processed by a Kalman filter to obtain a real-time correction value for the tire lateral stiffness coefficient;

[0015] The real-time correction value of the tire side stiffness coefficient is substituted into the second-order correction equation that includes tire nonlinear characteristics, and the vehicle motion state parameters at each time step in the first initial trajectory prediction result are recalculated to obtain the first trajectory prediction result; wherein, the vehicle motion state parameters include the instantaneous velocity of the vehicle in the X direction, the instantaneous velocity in the Y direction, and the yaw rate of the vehicle body in the geodetic coordinate system.

[0016] In some possible implementations, obtaining the driver's first fusion feature includes:

[0017] The driver's intention signal is obtained; the driver's intention signal is input into the first network sub-model to obtain the first sub-feature; the driver's psychological and physiological signals are obtained; the driver's psychological and physiological signals are input into the second network sub-model to obtain the second sub-feature; the traffic scene is obtained; the traffic scene is input into the third network sub-model to obtain the third sub-feature; the first sub-feature, the second sub-feature, and the third sub-feature are fused to obtain the driver's first fused feature.

[0018] In some possible implementations, acquiring the driver's intention signal includes:

[0019] The turn signal switch status signal is collected by the vehicle controller local area network within a preset time period. The turn signal switch status signal includes the activation duration, activation frequency and the vehicle speed corresponding to the activation.

[0020] The lateral offset and time-series change rate of the vehicle relative to the lane centerline are obtained through a lane line recognition sensor; real-time change data of the steering wheel angle are collected through a steering wheel angle sensor, including the peak angle, the angle change cycle, and the angle holding time; the turn signal switch status signal, the lateral offset and time-series change rate, and the real-time change data of the steering wheel angle are integrated to obtain the driver's intention signal.

[0021] In some possible implementations, determining the prediction strategy based on the first uncertainty data, the second uncertainty data, and the vehicle's first speed includes:

[0022] A gating network model is constructed, and the first uncertainty data, the second uncertainty data, and the vehicle's first speed are input into the gating network model to obtain gating coefficients. If the gating coefficients are higher than a first threshold, the first trajectory prediction result is determined as the prediction strategy. If the gating coefficients are lower than a second threshold, the second trajectory prediction result is determined as the prediction strategy. If the gating coefficients are higher than the second threshold and lower than the first threshold, the first trajectory prediction result and the second trajectory prediction result are combined to obtain a third trajectory prediction result, and the third trajectory prediction result is determined as the prediction strategy.

[0023] In some possible implementations, processing the driver's first fused features to obtain the second trajectory prediction result includes:

[0024] Monte Carlo sampling is performed on the driver's first fusion feature to generate a trajectory set containing a preset number of candidate trajectories. The probability of the candidate trajectories in the trajectory set is calculated, and the candidate trajectory with the highest probability is determined as the second trajectory prediction result.

[0025] In some possible implementations, the combination of the first trajectory prediction result and the second trajectory prediction result to obtain the third trajectory prediction result includes:

[0026] The tail of the first trajectory prediction result is connected to the head of the second trajectory prediction result to obtain the third initial trajectory prediction result; the third initial trajectory prediction result is then smoothed to obtain the third trajectory prediction result.

[0027] In some possible implementations, the smoothing process for the third initial trajectory prediction result includes:

[0028] A moving average filtering algorithm is used, with a preset length of sliding window, to filter the position coordinate sequence of the third initial trajectory prediction result point by point. A cubic spline interpolation algorithm is then used on the filtered trajectory sequence to add interpolation nodes at trajectory connection points and curvature abrupt change points, so that the first and second derivatives of the trajectory are continuous. The curvature change rate of the smoothed trajectory is calculated. If there are road segments where the curvature change rate exceeds the preset safety threshold, the density of interpolation nodes is readjusted until the curvature change rate of the trajectory meets the safety constraint, thus obtaining the third trajectory prediction result.

[0029] In some possible implementations, the uncertainty assessment of the first trajectory prediction result to obtain the first uncertainty data includes:

[0030] To assess the uncertainty of the first trajectory prediction result, the state covariance matrix of the Kalman filter is used. The trace of the state covariance matrix is ​​obtained from the state covariance matrix and is determined as the first uncertainty data. The calculation formula is as follows:

[0031]

[0032] in, The trace of the state covariance matrix. The element in the i-th row and i-th column of the state covariance matrix represents the overall variance of the state estimate, used to quantify the uncertainty of the first trajectory prediction result. n is the dimension of the state covariance matrix. .

[0033] In some possible implementations, the uncertainty assessment of the second trajectory prediction result to obtain second uncertainty data includes:

[0034] To assess the uncertainty of the second trajectory prediction results, a Monte Carlo technique is used for estimation. First, the covariance matrix of the trajectory state is calculated for each time step of the trajectory cluster. Based on the covariance matrix of the trajectory state, the trace of the covariance matrix of the trajectory state is obtained. The trace of the covariance matrix of the trajectory state is determined as the second uncertainty data. The calculation formula is as follows:

[0035]

[0036] in, Let be the trace of the covariance matrix of the trajectory state. The element in the i-th row and i-th column of the covariance matrix of the trajectory state is used to quantify the uncertainty of the second trajectory prediction result at time t, and m is the dimension of the covariance matrix of the trajectory state. .

[0037] Secondly, this application provides a vehicle trajectory prediction system, which includes:

[0038] The first prediction module is used to obtain the first motion state of the vehicle; construct a first network model, and obtain the first trajectory prediction result based on the first motion state of the vehicle;

[0039] The second prediction module is used to obtain the driver's first fusion features; and to process the driver's first fusion features to obtain the second trajectory prediction result.

[0040] The gating module is used to evaluate the uncertainty of the first trajectory prediction result to obtain the first uncertainty data, evaluate the uncertainty of the second trajectory prediction result to obtain the second uncertainty data, obtain the first vehicle speed, and determine the prediction strategy based on the first uncertainty data, the second uncertainty data and the first vehicle speed.

[0041] The target prediction module is used to determine the target trajectory prediction result of the vehicle according to the prediction strategy.

[0042] Thirdly, this application provides a computing device, including a memory and a processor;

[0043] The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of the first aspects.

[0044] Fourthly, this application provides a computer-readable storage medium for storing a computer program for performing the method as described in any one of the first aspects.

[0045] Fifthly, this application provides a computer program product comprising one or more computer instructions, wherein when the computer instructions are executed by a computer, the computer performs the method as described in any one of the first aspects.

[0046] As can be seen from the above technical solution, this application has at least the following beneficial effects:

[0047] In this application, a first motion state of the vehicle is obtained; a first network model is constructed, and a first trajectory prediction result is obtained based on the first motion state of the vehicle; a first fusion feature of the driver is obtained; the first fusion feature of the driver is processed according to a second network model to obtain a second trajectory prediction result; uncertainty assessment is performed on the first trajectory prediction result to obtain first uncertainty data, and uncertainty assessment is performed on the second trajectory prediction result to obtain second uncertainty data; a first vehicle speed is obtained, and a prediction strategy is determined based on the first uncertainty data, the second uncertainty data, and the first vehicle speed; the target trajectory prediction result of the vehicle is determined according to the prediction strategy. In the prior art, vehicle trajectory prediction mostly relies on a single physical model or a learning model. Although physical models are accurate, they are difficult to cope with complex interactive scenarios, while learning models have strong coverage but suffer from a lack of physical consistency and difficulty in quantifying uncertainty, leading to deviations or risks in trajectory prediction under complex traffic conditions. Therefore, this application, by combining a first network model and a second network model, uses a physical model to ensure short-term prediction accuracy, leverages a learning model to cover complex scenarios, and allows the system to intelligently select strategies through uncertainty assessment, achieving full-time-domain, highly robust trajectory prediction and filling the gap in the adaptability of single models in complex scenarios.

[0048] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0049] Figure 1 A schematic diagram illustrating an application scenario provided in an embodiment of this application;

[0050] Figure 2 A flowchart illustrating a vehicle trajectory prediction method provided in this application embodiment;

[0051] Figure 3 A schematic diagram of a vehicle trajectory prediction device provided in an embodiment of this application;

[0052] Figure 4 This is a schematic diagram of a computing device provided in an embodiment of this application. Detailed Implementation

[0053] The terms "first," "second," and "third," etc., used in this application specification and accompanying drawings are used to distinguish different objects, not to limit a specific order.

[0054] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0055] To ensure clarity and conciseness in the description of the following embodiments, a brief introduction to the related technologies is given first:

[0056] Currently, vehicle trajectory prediction is a core technology in the field of intelligent driving, used to predict the future trajectory of vehicles and assist in decision-making and planning. Traditional methods mostly rely on a single model, such as a physical model. Such models can respond with high accuracy based on the current state of the vehicle in a short time domain (e.g., within 0.5 seconds), but their core drawback is the lack of foresight. They cannot understand and predict nonlinear, long-term trajectory changes driven by the driver's long-term intentions (e.g., an impending lane change) or complex traffic games (e.g., no left-turn space at an intersection). As the time series lengthens, their prediction errors accumulate rapidly and diverge exponentially, leading to significant errors from minor inaccuracies. Furthermore, while based on short-term extrapolations of vehicle state, they struggle to incorporate complex scenario information such as driver intentions and multi-vehicle interactions.

[0057] For example, pure learning models excel at learning complex driving patterns from large datasets and can make longer-term predictions. However, their shortcomings are also significant: a lack of physical interpretability and safety boundaries: as "black box" models, they may generate physically unrealistic trajectories that violate vehicle dynamics constraints (e.g., instantaneous steering, overload skidding); insufficient generalization ability and reliability in edge scenarios: the model's performance is heavily dependent on the coverage of the training data. For "out-of-distribution scenarios" that are not present or rare in the training data, the reliability of its predictions drops sharply, potentially leading to catastrophic consequences.

[0058] Complex traffic scenarios, such as multi-vehicle interactions and sudden driver maneuvers, place stringent demands on predicted trajectory results. A single model struggles to handle all these scenarios, and most models only output the most probable trajectory, failing to capture and express the multimodal characteristics of future behavior. For example, at an intersection, vehicles may go straight, turn left, or turn right; predicting only one possibility constitutes a serious underestimation of traffic risk.

[0059] Most existing technologies focus only on the driver's explicit intentions, such as turn signals and large steering wheel movements, while neglecting the implicit states that are more indicative of risk, namely the driver's emotional and physiological states (such as fatigue, distraction, tension, road rage, etc.). These implicit states significantly affect the stability of driving behavior and risk preference, and are key factors leading to "human-caused accidents." Because existing models lack awareness of this dimension, the predictive systems are inherently inadequate in their ability to warn of sudden risks.

[0060] Therefore, some studies attempt to fuse physical and data models, but their fusion strategies are often static or based on simple rules, such as fixed time slices or linear weighting. This one-size-fits-all approach cannot dynamically adjust the confidence levels of the two models according to the real-time changing scene determinism. It cannot answer a core question: "In the current frame, should we trust the accurate physical deduction more, or the forward-looking learning model more?" This static fusion is essentially a compromise, not a true intelligent arbitration, and it is difficult to achieve optimal prediction results in all scenarios.

[0061] In summary, current technology urgently needs a new trajectory prediction framework that can dynamically and adaptively integrate short-term physical accuracy with long-term multimodal predictability, deeply understand the complete driving scenario including driver emotions, and quantify the uncertainty of its own predictions to guide safety decisions.

[0062] In view of this, embodiments of this application provide a vehicle trajectory prediction method. This method involves: acquiring a first motion state of the vehicle; constructing a first network model; obtaining a first trajectory prediction result based on the first motion state of the vehicle; acquiring a first fusion feature of the driver; processing the first fusion feature of the driver according to a second network model to obtain a second trajectory prediction result; performing uncertainty assessment on the first trajectory prediction result to obtain first uncertainty data; performing uncertainty assessment on the second trajectory prediction result to obtain second uncertainty data; acquiring a first vehicle speed; determining a prediction strategy based on the first uncertainty data, the second uncertainty data, and the first vehicle speed; and determining the target trajectory prediction result of the vehicle according to the prediction strategy. It is evident that this application, by combining a first network model and a second network model, uses a physical model to ensure short-term prediction accuracy, leverages a learning model to cover complex scenarios, and allows the system to intelligently select strategies through uncertainty assessment, achieving full-time-domain, highly robust trajectory prediction and filling the gap in adaptability of single models to complex scenarios.

[0063] To make the technical solution of this application clearer and easier to understand, the application scenarios of the technical solution of this application are described below with reference to the accompanying drawings. Figure 1 As shown in the figure, this figure is a schematic diagram of an application scenario provided by an embodiment of this application.

[0064] In this application scenario, vehicle A is the vehicle for which trajectory prediction is to be provided in this application. The scenario presents a multi-lane driving environment. There are other vehicles in the lane where vehicle A is located and in the surrounding area. The lane arrows and turn signs above indicate that vehicle A needs to achieve accurate trajectory prediction and assisted driving decision-making in a road scenario with multiple vehicles coexisting and turn guidance, by using the trajectory prediction method of this application, combined with information such as the movement status of surrounding vehicles and road guidance.

[0065] To make the technical solution of this application clearer and easier to understand, the following describes a vehicle trajectory prediction method provided by an embodiment of this application, in conjunction with the above application scenarios. For example... Figure 2 As shown in the figure, this is a flowchart of a vehicle trajectory prediction method provided in an embodiment of this application.

[0066] This method is applied to in-vehicle systems, and the vehicle trajectory prediction method includes:

[0067] S201, Obtain the vehicle's first motion state.

[0068] The first motion state refers to the physical parameters and driver operations collected in real time during vehicle operation within the first time window. Specifically, this includes data acquired through the vehicle controller area network (CAN bus) and inertial measurement unit (IMU), such as vehicle speed, distance from the vehicle's center of gravity to the front and rear axles, steering angles of the four wheels, and yaw angles, as well as driver operation signals like steering wheel angle, throttle opening, and brake pedal travel. These data collectively constitute a complete description of the vehicle's current motion state, providing the basic input for subsequent trajectory prediction.

[0069] S202. Construct the first network model and obtain the first trajectory prediction result based on the first motion state of the vehicle.

[0070] An initial network model is constructed, which is coupled with lateral motion to obtain the first network model.

[0071] The first motion state is input into the first network model to obtain the first initial trajectory prediction result;

[0072] The first initial predicted trajectory result is corrected to obtain the first trajectory prediction result. The lateral acceleration measured by the vehicle sensors is acquired, and the difference between the model-predicted lateral acceleration and the measured lateral acceleration corresponding to the first initial trajectory prediction result is calculated. This difference is processed using a Kalman filter to obtain a real-time correction value for the tire side stiffness coefficient. This real-time correction value for the tire side stiffness coefficient is substituted into a second-order correction equation incorporating tire nonlinear characteristics to recalculate the vehicle motion state parameters at each time step in the first initial trajectory prediction result, thus obtaining the first trajectory prediction result. The vehicle motion state parameters include the vehicle's instantaneous velocity in the X-direction, instantaneous velocity in the Y-direction, and yaw rate in the geodetic coordinate system. Specifically, this includes:

[0073] The initial network model is a bicycle model, a simplified vehicle dynamics model that simplifies the complex motion of a four-wheeled vehicle into a bicycle-like front and rear axle structure for analysis. Its core idea is to ignore the differences between the left and right wheels, treating the vehicle as a rigid body composed of two concentrated mass points at the front and rear. By simplifying the force and motion relationships between the front and rear axles, the planar motion state of the vehicle can be quickly calculated. This model is widely used in short-time-domain trajectory prediction because it balances computational efficiency and physical realism, and can well reflect the basic motion characteristics of vehicles in the low- to medium-speed range.

[0074] Then, the bicycle model is coupled with yaw motion. Yaw motion refers to the rotational motion of the bicycle around an axis perpendicular to the ground, which can be understood as a virtual axis pointing from the bottom of the vehicle to the top, similar to the "left and right shaking" motion of the vehicle when turning. The magnitude of this rotation is usually described by the yaw angle. The coupling between the bicycle model and yaw motion is mainly achieved through two aspects:

[0075] (1) Kinematic Relationship: The horizontal translational motion of a vehicle (i.e., changes in position in the X and Y directions) is closely linked to the yaw motion of the vehicle body (changes in the yaw angle) through a dynamic relationship. Changes in the yaw angle directly change the direction of the vehicle's translation. For example, when the vehicle body shakes to the left (the yaw angle increases), the vehicle's forward direction will deviate to the left. At the same time, the speed and direction of the vehicle's translation will also affect the speed of the yaw motion. For example, when the vehicle speed increases, the angular velocity of the yaw motion will be greater under the same steering operation.

[0076] (2) Closed-loop relationship of interaction: In the core computational logic of the model, the rotational speed of the yaw motion (i.e., the yaw angular velocity) depends on the steering angle of the front and rear wheels, the overall vehicle speed, and the lateral angle at the center of gravity of the vehicle body (the angle between the actual direction of vehicle movement and the orientation of the vehicle body). The vehicle's speed in the front-back and left-right directions depends on the current yaw angle and the lateral angle at the center of gravity. This cyclical relationship of "translational motion affecting rotational motion, and rotational motion reacting to translational motion" forms a tight closed loop, making the yaw motion deeply coupled with the overall vehicle motion, jointly determining the real-time trajectory of the vehicle. Thus, the first network model is obtained, and its core state differential equation is as follows:

[0077]

[0078]

[0079]

[0080]

[0081] in, Let be the instantaneous velocity of the vehicle in the X direction of the geodetic coordinate system. Let be the instantaneous velocity of the vehicle in the Y direction of the geodetic coordinate system. This refers to the yaw rate of the vehicle body, which is the rotational speed during yaw motion. This refers to the overall vehicle speed, specifically the forward speed at the vehicle's center of gravity. This is the vehicle body yaw angle, which is the angle between the vehicle body's longitudinal axis and the X-axis of the geodetic coordinate system. The sideslip angle is the angle between the direction of the vehicle's center of gravity velocity and the vehicle's longitudinal axis. This is the distance from the vehicle's center of gravity to the front axle. This is the distance from the vehicle's center of gravity to the rear axle. This refers to the front wheel steering angle. This is the rear wheel steering angle.

[0082] This application generates the trajectory point sequence P_short for the next 0.5 seconds by numerically integrating the above differential equation and iterating in a small step size Δt=0.01s for a total of N_short=50 steps, which is the first initial trajectory prediction result.

[0083] Since there are differences between the ideal model and the real vehicle (such as tire wear, changes in road surface adhesion coefficient, etc.), corrections are required through the following steps:

[0084] The ideal model, also known as the first network model, assumes that parameters such as tire lateral stiffness are fixed. However, in reality, these parameters change dynamically with driving conditions, leading to variations in the lateral acceleration predicted by the model (as determined by...). , and The derivation shows a discrepancy between the lateral acceleration measured by the vehicle's sensors (such as IMU) and the actual lateral acceleration measured by the sensor.

[0085] An integrated online parameter estimator is used, and the Kalman filter algorithm is employed to process the error signal: the input is the difference between the measured lateral acceleration and the model-predicted lateral acceleration; the output is the real-time estimated correction value of the tire lateral stiffness coefficient.

[0086] Substitute the corrected value of the tire side stiffness coefficient into the second-order correction term of a more advanced dynamic model (such as a supplementary equation considering tire nonlinear characteristics), and recalculate for each time step. , and The initial trajectory point sequence P_short is corrected point by point to eliminate the accumulated error caused by parameter mismatch. The corrected trajectory point sequence is the first trajectory prediction result, which has significantly improved physical consistency and consistency with the actual motion, providing a reliable short-time domain prediction basis for subsequent fusion decisions.

[0087] S203, Obtain the driver's first fusion feature.

[0088] The system acquires the driver's intention signal; collects turn signal switch status signals over a preset time period via the vehicle controller local area network, including the activation duration, activation frequency, and vehicle speed at which the turn signal is activated; acquires the lateral offset and time-series change rate of the vehicle relative to the lane centerline via a lane line recognition sensor; acquires real-time change data of the steering wheel angle via a steering wheel angle sensor, including the peak angle, angle change period, and angle holding duration; and integrates the turn signal switch status signal, lateral offset and time-series change rate, and real-time steering wheel angle change data to obtain the driver's intention signal.

[0089] The driver's intention signal is input into the first network sub-model to obtain the first sub-feature;

[0090] Acquire the driver's psychological and physiological signals; input the driver's psychological and physiological signals into the second network sub-model to obtain the second sub-features;

[0091] Obtain the traffic scene; input the traffic scene into the third network sub-model to obtain the third sub-feature;

[0092] The first, second, and third sub-features are fused to obtain the driver's first fused feature. Specifically:

[0093] Intent signals refer to signals emitted by the driver through vehicle operation that explicitly reflect their driving decisions, such as turn signals and steering wheel operations, used to determine the vehicle's short-term motion goals, such as maintaining lanes and changing lanes; the first network sub-model refers to a deep learning model used to process driver intent signals, whose core function is to extract key features from temporal intent signals and output intent feature vectors.

[0094] The driver's intention signal is input into the first network sub-model to obtain the first sub-feature, which is the intention signal obtained from the past second through onboard sensors, including the turn signal status, the lateral offset of the vehicle relative to the lane centerline, and the rate of change of the steering wheel angle. The first network sub-model adopts a hybrid architecture of a one-dimensional convolutional neural network (3-layer 1D-CNN) + a first recurrent neural network (Bidirectional Long Short-Term Memory, Bi-LSTM). The 1D-CNN (with a kernel size of 5 and the number of channels increasing from 16 to 64) extracts local patterns from the temporal signal, such as the duration of the turn signal and the abrupt changes in the steering wheel angle. The Bi-LSTM (128 hidden layer units) captures the long-distance temporal dependencies between features (such as the correlation between lateral offset and steering wheel operation). Finally, through fully connected layers and a softmax activation function, a probability vector of "keeping the lane", "changing lanes to the left", and "changing lanes to the right" is generated, which is the first sub-feature, quantifying the driver's intention tendency.

[0095] Psychological and physiological signals refer to implicit signals that imply the driver's emotions, fatigue, and other states, such as steering wheel torque fluctuations, pedal pressure / rate, and high-frequency steering wheel vibrations, reflecting the driver's irrational behavioral tendencies. The second network sub-model refers to a deep learning model used to process the driver's psychological and physiological signals. Its core function is to extract emotional and physiological state features from multi-source implicit signals and output psychological and physiological feature vectors.

[0096] The driver's psychological and physiological signals are input into the second network sub-model to obtain the second sub-feature. Implicit behavioral characteristics of the driver are captured through onboard sensors, including subtle fluctuations in steering wheel torque (high frequency when nervous, smoother when fatigued), the force and rate of pedal input (rapid pedal operation when agitated, unstable force when fatigued), and high-frequency vibration signals from the steering wheel reflecting hand stability (irregular vibration when nervous). The second network sub-model employs a "parallel 1D-CNN + LSTM + self-attention mechanism" architecture. The parallel 1D-CNN extracts frequency domain features for different signals separately (e.g., low-frequency signals reflect fatigue, high-frequency signals reflect nervousness). LSTM models the temporal evolution of emotional states, such as the gradual transition from normal to fatigue. The self-attention mechanism dynamically weights the contribution of different signals; for example, during emergency braking, the pedal signal has a higher weight than the steering wheel signal. Finally, a classification head generates the probability distribution of emotional states, i.e., the second sub-feature, quantifying the driver's psychological and physiological state.

[0097] Traffic scene refers to the real-time traffic environment in which a vehicle is located, including surrounding vehicles, pedestrians, traffic lights, road structures and other elements and their interactions; the third network sub-model refers to a deep learning model used to process traffic scene information. Its core function is to extract interaction features from the dynamic traffic environment and output a "scene interaction feature vector".

[0098] The traffic scene is input into the third network sub-model to obtain the third sub-feature. The complex environment is abstracted into a dynamic graph, where nodes include the vehicle itself, N key neighboring vehicles within the field of view, pedestrians, traffic lights, roadside units, etc. Edges are dynamically constructed based on the relative positions, speed differences, and line-of-sight occlusion relationships between nodes. For example, a "potential interaction edge" is established when the distance between a neighboring vehicle and the vehicle is less than 50 meters. The third network sub-model adopts a "Temporal-Graph Attention Network (T-GAT)". In the spatial dimension, the influence weights between nodes are learned through a graph attention network (GAT). For example, the influence weight of a traffic light on the vehicle is higher than that of a pedestrian in the distance. In the temporal dimension, the graph structure information of the past 5 frames is aggregated through LSTM to capture the dynamic evolution of the scene, such as the change of a neighboring vehicle from going straight to decelerating. The output is a highly condensed scene context feature vector, i.e., the third sub-feature, which quantifies the interaction relationships of the traffic environment.

[0099] When generating the first fused feature by fusing sub-features, the first, second, and third sub-features are concatenated to form the fused feature vector, as shown in the following formula:

[0100] F_joint=concat(P_intent,P_emotion,F_env)

[0101] Where F_joint is the first fusion feature, P_intent is the first sub-feature, P_emotion is the second sub-feature, F_env is the third sub-feature, and concat is the concatenation feature.

[0102] The state vector of the first trajectory prediction result at time t=0.5s can be selectively added as an initial condition. This first fusion feature F_joint contains the complete basis for predicting future trajectories. The concatenation method can be understood with an example: if the intention probability vector is 3-dimensional, the emotion probability distribution is 4-dimensional, and the scene interaction vector is 128-dimensional, concatenation will form a 3+4+128=135-dimensional fusion feature vector, thus integrating multi-dimensional information to provide comprehensive input for long-term trajectory prediction.

[0103] This fused feature vector integrates the driver's explicit intent, implicit state, and environmental constraints, providing a comprehensive basis for decision-making in subsequent long-term trajectory prediction. This not only overcomes the limitations of traditional trajectory prediction methods that rely on only a single signal, avoiding prediction bias caused by missing information, but also provides an extremely comprehensive and accurate basis for decision-making in subsequent long-term trajectory prediction. This significantly improves the accuracy and adaptability of vehicle trajectory prediction over long time domains, powerfully promoting the optimization and improvement of intelligent driving system performance.

[0104] S204. Process the first fusion feature of the driver to obtain the second trajectory prediction result.

[0105] Monte Carlo sampling is performed on the driver's first fusion feature to generate a trajectory set containing a preset number of candidate trajectories. The probability of the candidate trajectories in the trajectory set is calculated, and the candidate trajectory with the highest probability is determined as the second trajectory prediction result.

[0106] The first fusion feature integrates multi-dimensional information such as driver intent, psychophysiological state, and traffic scenario, and performs Monte Carlo sampling on it. Based on the probability distribution of the fusion feature, it can simulate and generate a large number of candidate trajectories. These candidate trajectories cover the driving possibilities under different driving situations, operational tendencies, and environmental interactions, breaking through the limitations of a single prediction path. It fully explores the potential changes in vehicle trajectory in complex driving scenarios, making the system's prediction of future driving trajectories more comprehensive and providing rich references for dealing with changing road conditions.

[0107] After generating a trajectory set containing a preset number of candidate trajectories, the probability of each candidate trajectory is calculated, and the trajectory with the highest probability is selected as the second prediction result. This process quantifies the rationality of different trajectories through probability, selecting the best from many possibilities. It combines the real driving logic reflected by multi-dimensional fusion features with the elimination of unreasonable and low-probability trajectories through probability screening, greatly improving the accuracy of trajectory prediction. This makes the prediction results more consistent with the reasonable path that the vehicle should follow in actual driving, enhancing the reliability of the intelligent driving system's decision-making.

[0108] During driving, the driver's state and the behavior of other road users are constantly and dynamically changing. The trajectory set generated by Monte Carlo sampling can respond in real time to the dynamic updates of fused features, such as changes in driver intent or sudden lane changes by neighboring vehicles, continuously producing candidate trajectories adapted to new situations. The probability filtering mechanism can also quickly focus on the most likely trajectory in the current scenario, ensuring that trajectory prediction keeps up with the dynamic changes in the scenario. This guarantees that the intelligent driving system always outputs reasonable and accurate trajectory predictions in complex and dynamic scenarios, helping the vehicle to drive smoothly and safely, and improving the system's adaptability to complex road conditions.

[0109] The second trajectory prediction results provide crucial information for the decision-making layer of the intelligent driving system. Whether it's path adjustment and obstacle avoidance strategy formulation during automatic cruise, or operation planning such as following other vehicles and changing lanes, reliable trajectory prediction can be used to plan safer and more efficient driving schemes in advance, coordinate vehicle acceleration, braking, steering and other controls, optimize the overall decision-making process of intelligent driving, and promote a more intelligent and smooth closed loop from perception to execution, thereby improving driving safety and comfort.

[0110] S205. An uncertainty assessment is performed on the first trajectory prediction result to obtain the first uncertainty data, and an uncertainty assessment is performed on the second trajectory prediction result to obtain the second uncertainty data.

[0111] In the field of intelligent vehicle trajectory prediction, uncertainty is a core indicator for measuring the reliability of prediction results and reflecting the risks of dynamic changes in future scenarios. Essentially, it represents the probability of deviation between the output of the prediction model and the vehicle's actual future trajectory, directly impacting the safety and robustness of the autonomous driving system's decision-making and planning. Existing technologies have significant limitations in their understanding and handling of uncertainty: traditional physical models, while relying on physical inertia and immediate control input in the short time domain and possessing high certainty, exhibit exponential error divergence and a sharp increase in uncertainty as the prediction time domain lengthens, due to nonlinear factors such as the driver's long-term intentions and traffic game theory. Effective quantification and response mechanisms are lacking. Data-driven models, while capable of learning complex driving patterns from massive amounts of data to support long-term predictions, are prone to generating trajectories that violate physical constraints due to their black-box nature. Furthermore, uncertainty increases sharply in edge scenarios not covered by training data, yet they cannot proactively identify their own reliable boundaries.

[0112] This application assesses the uncertainty of the first trajectory prediction result by using the state covariance matrix of a Kalman filter. The trace of this matrix physically represents the overall variance of the state estimate. Therefore, the "trace" of the state covariance matrix is ​​calculated to quantify the uncertainty, and the trace of the state covariance matrix is ​​determined as the first uncertainty data. The calculation formula is as follows:

[0113]

[0114] in, The trace of the state covariance matrix. The element in the i-th row and i-th column of the state covariance matrix represents the overall variance of the state estimate, used to quantify the uncertainty of the first trajectory prediction result. n is the dimension of the state covariance matrix. .

[0115] To assess the uncertainty of the second trajectory prediction results, a Monte Carlo technique is used for estimation. First, the covariance matrix of the trajectory state is calculated for each time step of the trajectory cluster. Based on the covariance matrix of the trajectory state, the trace of the covariance matrix of the trajectory state is obtained. The trace of the covariance matrix of the trajectory state is determined as the second uncertainty data. The calculation formula is as follows:

[0116]

[0117] in, Let be the trace of the covariance matrix of the trajectory state. The element in the i-th row and i-th column of the covariance matrix of the trajectory state is used to quantify the uncertainty of the second trajectory prediction result at time t, and m is the dimension of the covariance matrix of the trajectory state. .

[0118] Therefore This data is identified as the first uncertainty in the first trajectory prediction result. This will be identified as the second uncertainty data in the second trajectory prediction result.

[0119] S206. Obtain the vehicle's first speed, and determine a prediction strategy based on the first uncertainty data, the second uncertainty data, and the vehicle's first speed.

[0120] A gating network model is constructed by inputting the first uncertainty data, the second uncertainty data, and the vehicle's first speed into the gating network model to obtain the gating coefficients. The first speed is the vehicle speed collected during prediction. The formula for calculating the gating coefficients is as follows:

[0121]

[0122]

[0123]

[0124]

[0125] in, This is the input vector of the gated network model. This represents the first uncertainty data for the first trajectory prediction result. This is the second uncertainty data for the second trajectory prediction result. The first speed, This is the output vector of the first layer of the gated network model. This is the first weight matrix. For the first bias term, This is the output vector of the second layer of the gated network model. This is the second weight matrix. For the second bias term, The gating coefficient, This is the third weight matrix. For the third bias term, It is a linear rectified function. It is a sigmoid activation function.

[0126] After obtaining the gating coefficient, the following filtering judgment is performed:

[0127] If the gating coefficient is higher than the first threshold, the prediction result of the first trajectory is determined as the prediction strategy; wherein, the first threshold can be set to 0.8.

[0128] If the gating coefficient is lower than the second threshold, the second trajectory prediction result is determined as the prediction strategy; whereby the second threshold can be set to 0.2.

[0129] When the gating coefficient is higher than the second threshold and lower than the first threshold, a third initial trajectory prediction result is obtained by combining the first trajectory prediction result and the second trajectory prediction result, and the third trajectory prediction result is determined as the prediction strategy. Specifically, the tail of the first trajectory prediction result is connected to the head of the second trajectory prediction result to obtain the third initial trajectory prediction result; the third initial trajectory prediction result is then smoothed to obtain the third trajectory prediction result.

[0130] S207. Determine the target trajectory prediction result of the vehicle based on the prediction strategy.

[0131] Based on the actual operating scenario of the vehicle, such as real-time traffic conditions and its own motion status, the prediction strategy is selected and executed according to the following logic to output the target trajectory:

[0132] First, the current prediction strategy is determined by comparing the gating coefficient α with the first and second thresholds:

[0133] Prediction Strategy 1: Directly select the prediction result of the first trajectory as the target trajectory. This trajectory has high accuracy in the near time step (e.g., 1-3 seconds in the future), adapts to the continuous extrapolation of the vehicle's current motion state, and is suitable for stable environments and scenarios where large lane changes are not required.

[0134] Prediction Strategy Two: Select the filtered result from the long-term trajectory cluster of the second trajectory prediction results as the target trajectory. This trajectory covers a long-term time step (e.g., 5-10 seconds into the future), reflects the macro-path trend, and is suitable for scenarios such as high-speed driving and the need to predict intersections / merging areas.

[0135] Prediction Strategy 3: Trajectory Connection: Extract the tail segment of the first trajectory prediction result (such as the trajectory point in the last second, which contains recent movement trends) and concatenate it with the head segment of the second trajectory prediction result (such as the trajectory point in the first 2 seconds, which reflects the starting point of long-term planning) to obtain the third initial trajectory. When connecting, the timestamps and position coordinate systems must be aligned to ensure the continuity of physical meaning (such as using interpolation to correct jumps at connection points).

[0136] Smoothing: A moving average filtering algorithm is used, and a sliding window of preset length is set to filter the position coordinate sequence of the third initial trajectory prediction result point by point. A cubic spline interpolation algorithm is used on the filtered trajectory sequence to add interpolation nodes at trajectory connection points and curvature change points to make the first and second derivatives of the trajectory continuous. The curvature change rate of the smoothed trajectory is calculated. If there are road segments where the curvature change rate exceeds the preset safety threshold, the density of interpolation nodes is readjusted until the curvature change rate of the trajectory meets the safety constraint, and the third trajectory prediction result is obtained.

[0137] The coordinate sequence of the third initial trajectory is processed using a moving average (e.g., mean filtering with a window length of 3) or spline interpolation algorithm to eliminate abrupt changes caused by transitions, generating the third trajectory prediction result. This trajectory balances "short-term accuracy" and "long-term planning," making it suitable for scenarios with complex traffic flow and the need for dynamic path adjustments, such as turning at urban intersections and multi-vehicle interactions.

[0138] Dynamically adapt and predictive strategies based on real-time vehicle status and environmental information:

[0139] Example 1: When a vehicle is following another vehicle at low speed (speed v<10km / h) in a congested area, the first priority strategy is to rely on short-term accurate trajectories to avoid frequent lane change misjudgments.

[0140] Example 2: When a vehicle is cruising on a straight highway (speed v>80km / h), the second priority strategy is to rely on long-term trajectory planning to find a distant path.

[0141] Example 3: A vehicle is waiting to turn at an urban intersection (turn signal on, intersection curvature) When the current vehicle is in motion, strategy three is triggered, which integrates short and long-term trajectories to ensure the accuracy of current vehicle following while also planning the turning path in advance.

[0142] Through the above-mentioned "strategy matching-trajectory generation-dynamic adaptation" process, the final output is a target trajectory prediction result that fits the actual operation scenario of the vehicle, supporting decision-making for functions such as autonomous driving path planning and assisted driving warning.

[0143] Based on the above, this application provides a vehicle trajectory prediction method. The method involves: acquiring a first motion state of the vehicle; constructing a first network model; obtaining a first trajectory prediction result based on the first motion state of the vehicle; acquiring a first fusion feature of the driver; processing the first fusion feature of the driver according to a second network model to obtain a second trajectory prediction result; performing uncertainty assessment on the first trajectory prediction result to obtain first uncertainty data; performing uncertainty assessment on the second trajectory prediction result to obtain second uncertainty data; acquiring a first vehicle speed; determining a prediction strategy based on the first uncertainty data, the second uncertainty data, and the first vehicle speed; and determining the target trajectory prediction result of the vehicle based on the prediction strategy. It is evident that this application, by combining a first network model and a second network model, uses a physical model to ensure short-term prediction accuracy, leverages a learning model to cover complex scenarios, and allows the system to intelligently select strategies through uncertainty assessment, achieving full-time-domain, highly robust trajectory prediction and filling the gap in adaptability of single models to complex scenarios.

[0144] The above text combined Figures 1 to 2 The vehicle trajectory prediction method provided in the embodiments of this application has been described in detail. The system and equipment provided in the embodiments of this application will be described below with reference to the accompanying drawings.

[0145] This application also provides a vehicle trajectory prediction system, such as... Figure 3 As shown in the figure, this is a schematic diagram of a vehicle trajectory prediction system provided in an embodiment of this application. The system includes:

[0146] The first prediction module 301 is used to obtain the first motion state of the vehicle; construct a first network model, and obtain the first trajectory prediction result based on the first motion state of the vehicle;

[0147] The second prediction module 302 is used to acquire the driver's first fusion features; and to process the driver's first fusion features to obtain the second trajectory prediction result.

[0148] Gating module 303 is used to perform uncertainty assessment on the first trajectory prediction result to obtain first uncertainty data, and to perform uncertainty assessment on the second trajectory prediction result to obtain second uncertainty data; to obtain the first vehicle speed, and to determine the prediction strategy based on the first uncertainty data, the second uncertainty data and the first vehicle speed;

[0149] The target prediction module 304 is used to determine the target trajectory prediction result of the vehicle according to the prediction strategy.

[0150] In some possible implementations, the first prediction module 301 is specifically used to construct an initial network model, which is coupled with lateral motion to obtain a first network model.

[0151] In some possible implementations, the first prediction module 301 is specifically used to input the first motion state into the first network model to obtain a first initial trajectory prediction result; and to correct the first initial predicted trajectory result to obtain a first trajectory prediction result.

[0152] In some possible implementations, the first prediction module 301 is specifically used to acquire the lateral acceleration measured by the vehicle sensor and calculate the difference between the model-predicted lateral acceleration corresponding to the first initial trajectory prediction result and the measured lateral acceleration.

[0153] The difference is processed by a Kalman filter to obtain a real-time correction value for the tire lateral stiffness coefficient;

[0154] The real-time correction value of the tire side stiffness coefficient is substituted into the second-order correction equation that includes tire nonlinear characteristics, and the vehicle motion state parameters at each time step in the first initial trajectory prediction result are recalculated to obtain the first trajectory prediction result; wherein, the vehicle motion state parameters include the instantaneous velocity of the vehicle in the X direction, the instantaneous velocity in the Y direction, and the yaw rate of the vehicle body in the geodetic coordinate system.

[0155] In some possible implementations, the second prediction module 302 is specifically used to acquire the driver's intention signal; input the driver's intention signal into a first network sub-model to obtain a first sub-feature; acquire the driver's psychological and physiological signals; input the driver's psychological and physiological signals into a second network sub-model to obtain a second sub-feature; acquire the traffic scene; input the traffic scene into a third network sub-model to obtain a third sub-feature; and fuse the first sub-feature, the second sub-feature, and the third sub-feature to obtain the driver's first fused feature.

[0156] In some possible implementations, the second prediction module 302 is specifically used to collect turn signal switch status signals over a preset period of time via the vehicle controller local area network. The turn signal switch status signals include the activation duration, activation frequency, and the vehicle speed corresponding to the activation. The module also acquires the lateral offset and time-series change rate of the vehicle relative to the lane centerline via a lane line recognition sensor. Furthermore, it collects real-time change data of the steering wheel angle via a steering wheel angle sensor, including the peak angle, angle change period, and angle holding duration. Finally, it integrates the turn signal switch status signals, lateral offset and time-series change rate, and real-time steering wheel angle change data to obtain the driver's intention signal.

[0157] In some possible implementations, the gating module 303 is specifically used to construct a gating network model, inputting first uncertainty data, second uncertainty data, and the vehicle's first speed into the gating network model to obtain gating coefficients; when the gating coefficients are higher than a first threshold, the first trajectory prediction result is determined as the prediction strategy; when the gating coefficients are lower than a second threshold, the second trajectory prediction result is determined as the prediction strategy; when the gating coefficients are higher than the second threshold and lower than the first threshold, the first trajectory prediction result and the second trajectory prediction result are combined to obtain a third trajectory prediction result, and the third trajectory prediction result is determined as the prediction strategy.

[0158] In some possible implementations, the second prediction module 302 is specifically used to perform Monte Carlo sampling on the driver's first fused features to generate a trajectory set containing a preset number of candidate trajectories, perform probability calculation on the candidate trajectories in the trajectory set, and determine the candidate trajectory with the highest probability as the second trajectory prediction result.

[0159] In some possible implementations, the gating module 303 is further configured to connect the tail of the first trajectory prediction result with the head of the second trajectory prediction result to obtain a third initial trajectory prediction result; and to smooth the third initial trajectory prediction result to obtain a third trajectory prediction result.

[0160] In some possible implementations, the gating module 303 is specifically used to employ a moving average filtering algorithm, set a sliding window of a preset length, and filter the position coordinate sequence of the third initial trajectory prediction result point by point; apply a cubic spline interpolation algorithm to the filtered trajectory sequence, adding interpolation nodes at trajectory connection points and curvature abrupt change points to make the first and second derivatives of the trajectory continuous; calculate the curvature change rate of the smoothed trajectory, and if there are road segments where the curvature change rate exceeds a preset safety threshold, readjust the interpolation node density until the trajectory curvature change rate meets the safety constraint to obtain the third trajectory prediction result.

[0161] In some possible implementations, the gating module 303 is specifically used to evaluate the uncertainty of the first trajectory prediction result. It employs the state covariance matrix of a Kalman filter, obtains the trace of the state covariance matrix based on the state covariance matrix, and determines the trace of the state covariance matrix as the first uncertainty data. The calculation formula is as follows:

[0162]

[0163] in, The trace of the state covariance matrix. The element in the i-th row and i-th column of the state covariance matrix represents the overall variance of the state estimate, used to quantify the uncertainty of the first trajectory prediction result. n is the dimension of the state covariance matrix. .

[0164] In some possible implementations, the gating module 303 is specifically used to evaluate the uncertainty of the second trajectory prediction result. Monte Carlo techniques are used for estimation; first, the covariance matrix of the trajectory state is calculated for each time step of the trajectory cluster. The trace of the trajectory state covariance matrix is ​​obtained based on this covariance matrix, and the trace is determined as the second uncertainty data. The calculation formula is as follows:

[0165]

[0166] in, Let be the trace of the covariance matrix of the trajectory state. The element in the i-th row and i-th column of the covariance matrix of the trajectory state is used to quantify the uncertainty of the second trajectory prediction result at time t, and m is the dimension of the covariance matrix of the trajectory state. .

[0167] This application also provides a computing device. For example... Figure 4 As shown in the figure, this is a schematic diagram of a computing device provided in an embodiment of this application. The computing device 400 includes a bus 401, a processor 402, a communication interface 403, and a memory 404. The processor 402, the memory 404, and the communication interface 403 communicate with each other via the bus 401.

[0168] Bus 401 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0169] Processor 402 can be any one or more of the following processors: central processing unit (CPU), graphics processing unit (GPU), microprocessor (MP), or digital signal processor (DSP).

[0170] The communication interface 403 is used for communication with external devices. For example, if the computing device is a first switch, the communication interface 403 can be used for communication between the first switch and a first user terminal, or for communication between the first switch and a second switch.

[0171] Memory 404 may include volatile memory, such as random access memory (RAM). Memory 404 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0172] The memory 404 stores executable code, and the processor 402 executes the executable code to perform the aforementioned vehicle trajectory prediction method.

[0173] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium capable of being stored by a computing device, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct the computing device to execute the aforementioned vehicle trajectory prediction method.

[0174] This application also provides a computer program product comprising one or more computer instructions. When the computer instructions are loaded and executed on a computing device, all or part of the processes or functions described in this application are generated.

[0175] The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another website, computer, or data center via wired (e.g., coaxial cable, fiber optic) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0176] When the computer program product is executed by a computer, the computer performs any of the aforementioned vehicle trajectory prediction methods. The computer program product can be a software installation package; when any of the aforementioned vehicle trajectory prediction methods needs to be used, the computer program product can be downloaded and executed on the computer.

[0177] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0178] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be covered within the scope of protection of this application.

Claims

1. A vehicle trajectory prediction method, characterized in that, The method includes: The first motion state of the vehicle is obtained; the first motion state of the vehicle includes at least the vehicle speed, the distance from the vehicle's center of gravity to the front and rear axles, the steering angle of the four wheels, the body yaw angle, the steering wheel angle, the throttle opening and the brake pedal travel. Construct the first network model and obtain the first trajectory prediction result based on the vehicle's first motion state; The driver's first fusion feature is obtained through the driver's intention signal, the driver's psychological and physiological signals, and the traffic scene. The first fusion feature of the driver is processed to obtain the second trajectory prediction result; Uncertainty assessment is performed on the first trajectory prediction result to obtain the first uncertainty data, and uncertainty assessment is performed on the second trajectory prediction result to obtain the second uncertainty data; Obtain the vehicle's first speed, and determine the prediction strategy based on the first uncertainty data, the second uncertainty data, and the vehicle's first speed. The target trajectory prediction result of the vehicle is determined according to the prediction strategy; The step of determining the prediction strategy based on the first uncertainty data, the second uncertainty data, and the vehicle's first speed includes: A gating network model is constructed by inputting the first uncertainty data, the second uncertainty data, and the vehicle's first speed into the gating network model to obtain the gating coefficients. If the gating coefficient is higher than the first threshold, the first trajectory prediction result is determined as the prediction strategy; If the gating coefficient is lower than the second threshold, the second trajectory prediction result is determined as the prediction strategy; When the gating coefficient is higher than the second threshold and lower than the first threshold, the third trajectory prediction result is obtained by combining the first trajectory prediction result and the second trajectory prediction result, and the third trajectory prediction result is determined as the prediction strategy.

2. The method according to claim 1, characterized in that, The construction of the first network model includes: An initial network model is constructed, which is coupled with lateral motion to obtain the first network model.

3. The method according to claim 1, characterized in that, The first trajectory prediction result obtained based on the vehicle's first motion state includes: The first motion state is input into the first network model to obtain the first initial trajectory prediction result; The first initial trajectory prediction result is corrected to obtain the first trajectory prediction result.

4. The method according to claim 3, characterized in that, The step of correcting the first initial trajectory prediction result includes: Obtain the lateral acceleration measured by the vehicle sensors, and calculate the difference between the model-predicted lateral acceleration and the measured lateral acceleration corresponding to the first initial trajectory prediction result; The difference is processed by a Kalman filter to obtain a real-time correction value for the tire lateral stiffness coefficient; The real-time correction value of the tire side stiffness coefficient is substituted into the second-order correction equation that includes tire nonlinear characteristics, and the vehicle motion state parameters at each time step in the first initial trajectory prediction result are recalculated to obtain the first trajectory prediction result; wherein, the vehicle motion state parameters include the instantaneous velocity of the vehicle in the X direction, the instantaneous velocity in the Y direction, and the yaw rate of the vehicle body in the geodetic coordinate system.

5. The method according to claim 1, characterized in that, The acquisition of the driver's first fusion feature includes: Obtain the driver's intention signals; The driver's intention signal is input into the first network sub-model to obtain the first sub-feature; Acquire the driver's psychological and physiological signals; The driver's psychological and physiological signals are input into the second network sub-model to obtain the second sub-feature; Obtain traffic scenarios; The traffic scene is input into the third network sub-model to obtain the third sub-feature; The first sub-feature, the second sub-feature, and the third sub-feature are fused to obtain the driver's first fused feature.

6. The method according to claim 5, characterized in that, The acquisition of the driver's intention signal includes: The turn signal switch status signal is collected by the vehicle controller local area network within a preset time period. The turn signal switch status signal includes the activation duration, activation frequency and the vehicle speed corresponding to the activation. The lateral offset of the vehicle relative to the center line of the lane is obtained by lane line recognition sensors, and the time-series change rate is also used. The steering wheel angle sensor collects real-time data on steering wheel angle changes, including peak angle, angle change cycle, and angle holding time. By integrating the turn signal switch status signal, lateral offset and timing change rate, and real-time steering wheel angle change data, the driver's intention signal is obtained.

7. The method according to claim 1, characterized in that, The process of processing the driver's first fused features to obtain the second trajectory prediction result includes: Monte Carlo sampling is performed on the driver's first fusion feature to generate a trajectory set containing a preset number of candidate trajectories. The probability of the candidate trajectories in the trajectory set is calculated, and the candidate trajectory with the highest probability is determined as the second trajectory prediction result.

8. The method according to claim 1, characterized in that, The process of combining the first trajectory prediction result and the second trajectory prediction result to obtain the third trajectory prediction result includes: The tail of the first trajectory prediction result is connected to the head of the second trajectory prediction result to obtain the third initial trajectory prediction result. The third initial trajectory prediction result is smoothed to obtain the third trajectory prediction result.

9. The method according to claim 8, characterized in that, The smoothing process for the third initial trajectory prediction result includes: A moving average filtering algorithm is used, and a sliding window of preset length is set to filter the position coordinate sequence of the third initial trajectory prediction result point by point. A cubic spline interpolation algorithm is used on the filtered trajectory sequence to add interpolation nodes at trajectory connection points and curvature abrupt change points, so that the first and second derivatives of the trajectory are continuous; Calculate the rate of change of curvature of the smoothed trajectory. If there are road segments where the rate of change of curvature exceeds the preset safety threshold, readjust the interpolation node density until the rate of change of trajectory curvature meets the safety constraint, and obtain the third trajectory prediction result.

10. The method according to claim 1, characterized in that, The uncertainty assessment of the first trajectory prediction result to obtain the first uncertainty data includes: To assess the uncertainty of the first trajectory prediction result, the state covariance matrix of the Kalman filter is used. The trace of the state covariance matrix is ​​obtained from the state covariance matrix and is determined as the first uncertainty data. The calculation formula is as follows: in, The trace of the state covariance matrix. The element in the i-th row and i-th column of the state covariance matrix represents the overall variance of the state estimate, used to quantify the uncertainty of the first trajectory prediction result. n is the dimension of the state covariance matrix. .

11. The method according to claim 1, characterized in that, The uncertainty assessment of the second trajectory prediction result to obtain the second uncertainty data includes: To assess the uncertainty of the second trajectory prediction results, a Monte Carlo technique is used for estimation. First, the covariance matrix of the trajectory state is calculated for each time step of the trajectory cluster. Based on the covariance matrix of the trajectory state, the trace of the covariance matrix of the trajectory state is obtained. The trace of the covariance matrix of the trajectory state is determined as the second uncertainty data. The calculation formula is as follows: in, Let be the trace of the covariance matrix of the trajectory state. The element in the i-th row and i-th column of the covariance matrix of the trajectory state is used to quantify the uncertainty of the second trajectory prediction result at time t, and m is the dimension of the covariance matrix of the trajectory state. .

12. A vehicle trajectory prediction system, characterized in that, The system includes: The first prediction module is used to obtain the first motion state of the vehicle; the first motion state of the vehicle includes at least the vehicle speed, the distance from the vehicle's center of gravity to the front axle and the rear axle, the steering angle of the four wheels, the body yaw angle, the steering wheel angle, the throttle opening and the brake pedal travel; a first network model is constructed, and based on the first motion state of the vehicle, a first trajectory prediction result is obtained; The second prediction module is used to obtain the driver's first fusion feature; the driver's first fusion feature is obtained through the driver's intention signal, the driver's psychological and physiological signals, and the traffic scene; the driver's first fusion feature is processed to obtain the second trajectory prediction result; The gating module is used to perform uncertainty assessment on the first trajectory prediction result to obtain first uncertainty data, and to perform uncertainty assessment on the second trajectory prediction result to obtain second uncertainty data; to acquire the first vehicle speed, and to determine a prediction strategy based on the first uncertainty data, the second uncertainty data, and the first vehicle speed; the step of determining the prediction strategy based on the first uncertainty data, the second uncertainty data, and the first vehicle speed includes: A gating network model is constructed by inputting the first uncertainty data, the second uncertainty data, and the vehicle's first speed into the gating network model to obtain the gating coefficients. If the gating coefficient is higher than the first threshold, the first trajectory prediction result is determined as the prediction strategy; If the gating coefficient is lower than the second threshold, the second trajectory prediction result is determined as the prediction strategy; When the gating coefficient is higher than the second threshold and lower than the first threshold, the third trajectory prediction result is obtained by combining the first trajectory prediction result and the second trajectory prediction result, and the third trajectory prediction result is determined as the prediction strategy. The target prediction module is used to determine the target trajectory prediction result of the vehicle according to the prediction strategy.

13. A computing device, characterized in that, Including memory and processor; The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method as described in any one of claims 1 to 11.

Images