Optimization method for automatic driving track prediction and automatic driving control system

Abstract

The invention relates to the technical field of intelligent traffic, and discloses an automatic driving track prediction optimization method and an automatic driving control system, and the method comprises the steps: obtaining a prediction track of a vehicle in a first prediction time domain at a first positioning position; determining a first state based on the reference trajectory and the predicted trajectory; and updating the first positioning or the first prediction time domain based on the first state, and returning to the step of obtaining the prediction trajectory until the prediction trajectory meets a preset condition, thereby obtaining a final prediction trajectory. According to the scheme of the invention, prediction deviation caused by sensing noise, environment sudden change or model extreme scene blind areas can be sensed and corrected in time, and the trajectory prediction capability, safety and reliability of an automatic driving control system in an extreme scene are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation technology, and in particular to an optimization method for predicting autonomous driving trajectories and an autonomous driving control system. Background Technology

[0002] To ensure the driving safety of autonomous vehicles, trajectory prediction, as a core supporting technology, needs to accurately predict the future behavior of the vehicle and surrounding traffic participants, providing a basis for decision-making by the planning and control modules. Currently, most mainstream trajectory prediction technologies are based on deep learning models, trained with a large amount of real road data, and can achieve high-precision complete trajectory prediction in common scenarios such as straight driving and regular turns.

[0003] However, the aforementioned methods heavily rely on large-scale real-world road data, while extreme scenario samples are naturally scarce, leading to inherent blind spots in the models and becoming a core bottleneck restricting the safety of autonomous driving. Furthermore, for complete trajectories of fixed duration or distance output by the model, if deviations occur during trajectory prediction due to perceived noise, sudden environmental changes, or limitations of the model's extreme scenario samples, the methods lack a real-time assessment and feedback mechanism for the uncertainty of the prediction results and its potential risks. They cannot promptly detect and correct these deviations, causing errors to accumulate and amplify in subsequent planning and control stages, potentially leading to serious consequences such as collisions. For example, in autonomous driving trajectory prediction scenarios, changes in road surface adhesion are common interference factors, such as a sudden switch from a high-adhesion state on a dry road to a low-adhesion state in rain. The aforementioned methods' one-time prediction of the complete path cannot detect and promptly correct the potential deviation trend caused by insufficient road surface adhesion when facing such dynamic disturbances. This results in deviations when driving according to the predicted path, accumulating over prediction time and ultimately leading to serious safety accidents such as crossing the line or colliding with vehicles in adjacent lanes.

[0004] Therefore, there is an urgent need for an optimized method for predicting autonomous driving trajectories and an autonomous driving control system to address the above shortcomings. Summary of the Invention

[0005] In view of this, in order to overcome at least one aspect of the above problems, embodiments of the present invention propose an optimization method for autonomous driving trajectory prediction, specifically including the following steps: Obtain the predicted trajectory of the vehicle in the first prediction time domain at the first location; Based on the reference trajectory and the predicted trajectory, a first state is determined; Update the first positioning or the first prediction time domain based on the first state, and return to the step of obtaining the prediction trajectory until the prediction trajectory meets the preset conditions to obtain the final prediction trajectory.

[0006] In some embodiments, the step of determining the first state based on the reference trajectory and the predicted trajectory includes: The deviation distance is determined based on the first positioning and the reference trajectory; Risk data is determined based on the predicted trajectory; The first state is determined based on both the risk data and the deviation distance.

[0007] In some embodiments, updating the first positioning or first prediction time domain based on the first state includes: In response to the first state being normal, the first predicted time domain is updated according to the preset time domain increment; In response to the first state anomaly, target driving data and steering commands are determined based on the vehicle's current driving data, the reference trajectory, and the predicted trajectory. The steering commands and target driving data are executed, the first positioning is updated based on the position information after execution, and it is determined whether to update the first prediction time domain based on the target driving data.

[0008] In some embodiments, after the step of updating the first location or the first prediction time domain based on the first state, the method further includes: Based on the vehicle's historical trajectory and the predicted trajectory, a preset number of short-term trajectories are generated; A first risk quantity is determined based on all the short-term trajectories, and the updated first prediction time domain is adjusted based on the first risk quantity.

[0009] In some embodiments, determining the first risk quantity based on all of the short-term trajectories includes: Based on all the aforementioned short-term trajectories, determine the collision probability matrix for all traffic participants in the real-time environment; Based on each of the short-term trajectories, a corresponding risk loss value is determined, and a tail risk indicator is determined based on all of the risk loss values. The collision probability matrix and the tail risk index are used as the first risk quantity.

[0010] In some embodiments, obtaining the predicted trajectory of the vehicle within a first prediction time domain at the first location includes: Acquire image information of the vehicle at the first location, and extract several first-view features from the image information; Based on all the first viewpoint features, determine the first mixed feature; The first prediction time domain and the first mixed feature are input into the pre-trained prediction model to obtain the predicted trajectory.

[0011] In some embodiments, determining the first hybrid feature based on all the first viewpoint features includes: Calculate the first hardness index for each first viewpoint feature, and determine the target viewpoint feature based on the first hardness index that meets the preset index. The first hybrid feature is obtained based on the target viewpoint features and the preset reference viewpoint features.

[0012] In some embodiments, determining the first state based on both risk data and the deviation distance includes: The dynamic radius is determined based on the deviation distance and the preset radius function; In response to the deviation distance not being greater than the dynamic radius and the risk data not being less than the safety threshold, the first state is determined to be normal; In response to the deviation distance being greater than the dynamic radius or the risk data being less than the safety threshold, the first state is determined to be abnormal.

[0013] In some embodiments, after the step of updating the first location or the first prediction time domain based on the first state, the method further includes: Based on the vehicle's historical trajectory and the predicted trajectory, a preset number of short-term trajectories are generated; A first risk quantity is determined based on all the short-term trajectories, and the safety threshold, the preset radius function, and the updated first prediction time domain are adjusted based on the first risk quantity.

[0014] Based on the same inventive concept, according to another aspect of the present invention, embodiments of the present invention also provide an autonomous driving control system, comprising: At least one sensor, said sensor being used to acquire the initial position of the vehicle; At least one processor, which acquires a final predicted trajectory based on the first positioning using the method described in any of the above embodiments.

[0015] The present invention has at least the following beneficial technical effects: This invention provides an optimization method for autonomous driving trajectory prediction and an autonomous driving control system. A first state is determined by comparing a reference trajectory and a predicted trajectory, and the first positioning or first prediction time domain is iteratively updated until the predicted trajectory meets preset conditions, thereby achieving real-time evaluation and feedback optimization of the predicted trajectory. This iterative mechanism can promptly detect and correct prediction deviations caused by perceived noise, sudden environmental changes, or blind spots in extreme scenarios. While ensuring safety, it gradually expands the prediction range, avoiding the accumulation and amplification of errors in subsequent planning and control stages, effectively reducing collision risks. Simultaneously, this method reduces reliance on large-scale real-world road data and enhances trajectory prediction capabilities in extreme scenarios through online optimization, thereby improving the safety and reliability of the autonomous driving control system. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0017] Figure 1 A block diagram illustrating an embodiment of the optimization method for autonomous driving trajectory prediction provided by the present invention; Figure 2 This is a schematic diagram of an embodiment of the trajectory sequence JSON structure provided by the present invention; Figure 3 This is a schematic diagram of an embodiment of the BEV generation method provided by the present invention; Figure 4 This is a schematic diagram of an embodiment of the FV generation method provided by the present invention; Figure 5 A flowchart of an embodiment of the optimization method for autonomous driving trajectory prediction provided by the present invention; Figure 6 This is a schematic diagram of an embodiment of the automatic driving control system provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to specific examples and the accompanying drawings.

[0019] It should be noted that all uses of "first" and "second" in the embodiments of the present invention are for the purpose of distinguishing two entities or parameters with the same name but different names. It is clear that "first" and "second" are only for the convenience of expression and should not be construed as limiting the embodiments of the present invention. Subsequent embodiments will not explain this in detail.

[0020] The present invention will now be described in detail with reference to the embodiments and accompanying drawings.

[0021] Vehicles can refer to road vehicles such as passenger cars and commercial vehicles equipped with autonomous driving functions. The autonomous driving trajectory prediction optimization method proposed in this invention is applied to the real-time dynamic optimization of the predicted trajectory of autonomous vehicles during driving, and specifically addresses the online correction of prediction deviations and risks caused by perception uncertainties, model limitations, or sudden environmental changes.

[0022] The first aspect of this invention provides an optimization method for predicting autonomous driving trajectories, such as... Figure 1 As shown, the detection method for racing drones may include steps S10 to S30.

[0023] Step S10: Obtain the predicted trajectory of the vehicle in the first prediction time domain at the first positioning location.

[0024] In this embodiment of the invention, the first positioning is obtained by sensors such as the Global Positioning System, Inertial Measurement Unit, and Wheel Speed ​​Meter through a fusion algorithm. The first positioning includes real-time position and attitude, such as position coordinates (x, y), heading angle (yaw), velocity (v), acceleration (ax, ay), and yaw rate.

[0025] The first prediction time domain refers to the duration of trajectory prediction, and its initial value is adaptively set based on the vehicle's current speed. Whenever the vehicle reaches a new first position, trajectory prediction begins from the initial value of the first prediction time domain. That is, trajectory prediction is initially performed within a relatively short initial prediction time domain. Subsequently, based on the first state of the predicted trajectory within each first prediction time domain, it is dynamically determined whether to extend the duration of the first prediction time domain or correct the vehicle's first position. The formula for calculating the initial value of the first prediction time domain is as follows: ; in, These are the initial values ​​for the first prediction time domain. The preferred value is the minimum time-domain value. , The optimal value is the maximum time domain value. . Forward-looking time domain, it represents the increase in speed for every 1 m / s increase in vehicle speed. The duration of a second, meaning the faster the vehicle speed, the longer the initial forward-looking time domain is required, especially in urban road scenarios. Typically less than or equal to 4 seconds, in highway driving scenarios. It takes 6-8 seconds. This represents the vehicle's current speed.

[0026] The vehicle's multi-view perception information at the first positioning point is obtained by the vehicle's sensors, and multi-view features are extracted from it. The multi-view features and the first preset time domain are then input into the pre-trained prediction model. Based on the output of the prediction model, the predicted trajectory of the vehicle in the first prediction time domain at the first positioning point is obtained.

[0027] It should be noted that, in this embodiment of the invention, the pre-trained prediction model refers to a trajectory predictor that has been trained on multiple iterations of data and is capable of outputting future trajectories based on the input viewpoint features and the predicted time domain. The trajectory predictor can be implemented using architectures such as GRU (Gated Recurrent Unit), LSTM (Long Short-Term Memory), Transformer (a deep learning architecture), CVAE (Conditional Variational Autoencoder), multimodal heads, or GMM (Gaussian Mixture Model). The training process of this prediction model focuses on learning motion patterns under extreme scenarios to achieve stronger generalization ability.

[0028] Step S20: Determine the first state based on the reference trajectory and the predicted trajectory.

[0029] In this embodiment of the invention, the reference trajectory refers to the path the vehicle is expected to travel, which can be given by the lane centerline or the path point sequence output by the local planner. The first state is a binary decision signal used to characterize the potential risks of the predicted trajectory in the future, including normal and abnormal states.

[0030] Based on some embodiments, the deviation distance can be calculated by obtaining a path point in the reference trajectory that has the same timestamp as the current first positioning point and calculating the distance between the first positioning and that path point. Based on the deviation distance, a dynamic radius is calculated using a preset radius function. Simultaneously, risk data is obtained based on the predicted trajectory, and a safety threshold is obtained based on the current road scenario. Then, the trajectory deviation is compared with the dynamic radius, and the risk data is compared with the preset safety threshold, thereby determining a first state based on the comparison result. By obtaining the first state, the future deviation trend and potential risks of vehicles in the predicted trajectory can be perceived in advance. When a vehicle in the predicted trajectory shows a future deviation trend or potential risk, the first state is determined to be an abnormal state; otherwise, it is determined to be a normal state.

[0031] It should be noted that the dynamic radius is a safety tolerance boundary that dynamically decreases as the deviation distance increases. Risk data is a key indicator used to assess collision risk, including minimum time distance and minimum clearance. For example, the expected collision time and spatial clearance between the vehicle and each traffic participant in the real-time environment are calculated time-by-time, and the risk data is determined based on the expected minimum collision time and spatial clearance. Safety thresholds include a time safety threshold corresponding to the minimum time distance and a spatial safety threshold corresponding to the minimum clearance. In some application scenarios, different road types correspond to different safety thresholds. For example, the spatial safety threshold between vehicles in urban road scenarios is 2-5 meters, while the spatial safety threshold between vehicles in highway scenarios is 20-50 meters.

[0032] Step S30: Update the first positioning or first prediction time domain based on the first state, and return to step S10 until the predicted trajectory meets the preset conditions to obtain the final predicted trajectory.

[0033] In this embodiment of the invention, the predicted trajectory meeting preset conditions includes the first prediction time domain corresponding to the predicted trajectory reaching or exceeding a time domain threshold, or the predicted trajectory reaching a convergence state. The first prediction time domain reaching or exceeding the time domain threshold indicates that during the trajectory prediction iteration process, the current prediction time domain has covered the maximum effective prediction dimension under the current scenario; further extending the time domain would increase the risk of prediction distortion. The time domain threshold is an upper limit of the prediction duration set for long-term planning, and it is adaptively adjusted according to the road scenario. For example, the value range of the time domain threshold is 5-8s in highway scenarios and 3-5s in urban road scenarios. The predicted trajectory reaching a convergence state means that during the trajectory prediction iteration process, the change amplitude of the current predicted trajectory has narrowed to a preset accuracy range, and the core features (such as tail risk values) tend to stabilize, indicating that the current prediction has sufficient accuracy and reliability. The predicted trajectory reaching a convergence state includes trajectory accuracy meeting standards and risk control meeting standards. Trajectory accuracy meeting standards means that the average displacement error (ADE) and final displacement error (FDE) of the predicted trajectory are less than the prediction accuracy threshold corresponding to the current road scenario. Risk control compliance means that the tail risk value of the predicted trajectory is less than or equal to the preset risk baseline.

[0034] When the first state is abnormal, the first prediction time domain is temporarily extended, and the correction operation is performed on the vehicle's current first position first. For example, the vehicle's current speed, current distance to other road users, and heading angle are determined based on the current first position. Based on these, the target speed and target distance are calculated using a preset correction strategy, and a steering command is generated using the same strategy. The target speed, target distance, and steering command are executed to control the vehicle's movement. During movement, in response to the vehicle reaching the target speed and target distance, the current position is updated based on the vehicle's position, and the initial value of the first prediction time domain is adaptively updated based on the vehicle's target speed. Subsequently, the process returns to step S10, where trajectory prediction begins from the new initial value of the new first prediction time domain at the new first position, thus achieving trajectory correction and re-prediction. This method allows for setting a new, safer benchmark for subsequent trajectory prediction when there are potential future risks in the predicted trajectory. Trajectory prediction continues under the new, safer first position and first prediction time domain, avoiding the accumulation of prediction trajectory errors and achieving the purpose of correction.

[0035] When the first state is normal, the time length of the first prediction time domain is extended, and predictions of longer trajectories are continued. This time domain extension is achieved by adding a preset time domain increment to the first prediction time domain of the current prediction iteration and using it as the first prediction time domain for the next iteration. Subsequently, the process returns to step S10, where predictions continue based on the same first location and the new first prediction time domain to obtain a new prediction trajectory. In each iteration, a decision is made based on the first state whether to continue extending the prediction time domain. While ensuring prediction accuracy, the time coverage of the prediction is gradually expanded to obtain more forward-looking trajectory information, thus providing a more comprehensive decision-making basis for vehicle planning and execution. This phased extension of the prediction time domain avoids the long-term accuracy degradation caused by a single prediction that is too long.

[0036] In summary, the autonomous driving trajectory prediction optimization method provided by this invention combines real-time state assessment and dynamic correction to improve the robustness of trajectory prediction and driving safety of autonomous vehicles in extreme scenarios. This method does not perform a one-time trajectory prediction, but continuously monitors and evaluates the predicted trajectory, proactively intervening when potential risks are anticipated. It re-predicts a safer trajectory by adjusting the vehicle's positioning reference and prediction time domain, thus forming an online optimization closed loop capable of adaptively responding to complex environments. Furthermore, once the vehicle's driving state is normal or the target trajectory is obtained, the prediction range is gradually expanded iteratively until the target prediction level is reached. This ensures that the look-ahead distance is gradually extended while maintaining driving safety, achieving a natural transition from short-term emergency handling to long-term stable planning.

[0037] In some implementations, the step of determining a first state based on a reference trajectory and a predicted trajectory includes: determining a deviation distance based on a first positioning and a reference trajectory; determining risk data based on the predicted trajectory; and determining the first state based on both the risk data and the deviation distance.

[0038] In this embodiment of the invention, DRHG (Dynamic-Radius & Horizon-Gating) technology is used to determine the first state based on the reference trajectory and the predicted trajectory. During DRHG execution, the reference trajectory is used as the target for trajectory alignment and correction, including parameters such as the position, heading angle, and curvature of each point on the path. The predicted trajectory includes predicted values ​​such as the position, velocity, and acceleration of the trajectory at each moment within the first prediction time domain. First, the reference trajectory, the predicted trajectory, and the vehicle's first positioning are unified to the same time coordinate system. Then, path points in the reference trajectory with the same timestamp as the current first positioning are obtained, and the distance between them is calculated to obtain the deviation distance. Simultaneously, the expected collision time and spatial clearance between the vehicle and each traffic participant at the first positioning point are calculated based on the predicted trajectory. The minimum collision time and minimum spatial clearance among all expected collision times and spatial clearances are obtained as risk data. The risk data and the deviation distance are used simultaneously as the basis for determining the first state.

[0039] The larger the deviation distance, the greater the potential deviation between the vehicle's real-time position and attitude and the safety baseline, due to sudden environmental changes, accumulated perception errors, or insufficient adaptation to extreme scenarios. The formula for calculating the deviation distance is as follows: ; in, The deviation distance, For the trajectory point at time t in the reference trajectory, Let t be the first location.

[0040] Based on both risk data and deviation distance, the method for determining the first state may include: determining the dynamic radius according to the deviation distance and a preset radius function; determining the first state as normal in response to the deviation distance not being greater than the dynamic radius and the risk data not being less than the safety threshold; and determining the first state as abnormal in response to the deviation distance being greater than the dynamic radius or the risk data being less than the safety threshold.

[0041] In this embodiment of the invention, the corresponding dynamic radius is calculated using a preset radius function in DRHG based on the deviation distance. When the deviation distance is greater than the dynamic radius, it indicates that the error of the current predicted trajectory has exceeded the deviation threshold in the current scenario, and the vehicle is facing a deviation trend; at this time, the first state is abnormal. Alternatively, when the risk data is less than the safety threshold, it indicates that the safety hazard of the current predicted trajectory has exceeded the controllable range, and the vehicle is facing an immediate or potential dangerous situation; at this time, the first state is abnormal. When the deviation distance is not greater than the dynamic radius and the risk data is not less than the safety threshold, ...

[0042] It should be noted that the preset radius function is an arbitrary function that monotonically decreases with deviation distance, based on lane width and vehicle dynamics constraints. Vehicle dynamics constraints include physical constraints limiting vehicle motion and vehicle parameters such as road adhesion coefficient, upper limit of lateral acceleration, maximum steering angle, maximum acceleration / deceleration, upper speed limit, and gradient, provided by the simulation kernel configuration or vehicle specification file. The expression for the preset radius function is as follows: ; Where k is the contraction coefficient, which determines the sensitivity of the dynamic radius to changes in the deviation distance. Its specific value is set according to the "half-life error". For example, if we want the radius to be halved when the error increases by 1m, then k=ln2≈0.693. Using the radius as a reference, it is usually taken as min(0.5 × lane width). / ), The vehicle's current speed. This is the upper limit of lateral acceleration. . For the minimum radius, in urban road scenarios Highway scenario . The maximum radius is defined as follows: its value does not exceed 0.6 times the lane width in the current road scenario, and does not exceed the upper limit of the lane turning radius.

[0043] The method of updating the first positioning or the first prediction time domain based on the first state may include: in response to the first state being normal, updating the first prediction time domain according to a preset time domain increment; in response to the first state being abnormal, determining the target driving data and steering command according to the vehicle's current driving data, reference trajectory and predicted trajectory, executing the steering command and target driving data, updating the first positioning according to the position information after execution, and determining whether to update the first prediction time domain according to the target driving data.

[0044] In this embodiment of the invention, the preset time-domain increment is a fixed time step, which is the time increment for each expansion of the prediction time domain. Its value is determined according to the actual control frequency requirements and computing power, and is usually set to a range of 0.3-0.5s. The current driving data includes the current speed, current distance to the target vehicle, and heading angle, which are determined based on the first positioning. The target driving data includes the target speed, target distance to the target vehicle, and heading angle, which are obtained by performing correction processing on the current driving data.

[0045] When the first state is normal, the first prediction time domain is updated by adding a preset time domain increment to the first prediction time domain in each prediction iteration, and using this increment as the first prediction time domain for the next iteration. When the first state is abnormal, the target driving data is calculated based on the current driving data corresponding to each prediction iteration using a preset correction strategy, and a steering command is generated. The vehicle position reached by executing the above target driving data and steering command is used as the first positioning for the next iteration to update the first positioning. At the same time, the initial value of the first prediction time domain is adaptively adjusted and used as the first prediction time domain for the next iteration to update the first prediction time domain.

[0046] It should be noted that during the process of updating the first positioning or first prediction time domain, the DRHG controls the first prediction time domain to perform expansion or pause expansion operations. When the first state is normal, the DRHG controls the first prediction time domain to increment to perform expansion. When the first state is abnormal, the DRHG controls the first prediction time domain to pause expansion, triggering its internal preset correction strategy for correction operations. The preset correction strategy includes a longitudinal control strategy and a lateral control strategy. The longitudinal control strategy employs longitudinal controllers such as IDM (Intelligent Driver Model) or MPC (Model Predictive Control) algorithms. Specifically, the longitudinal control strategy calculates the downward adjustment of the speed based on the vehicle's current speed to obtain the target speed, and calculates the upward adjustment of the following distance based on the current distance to obtain the target following distance, thereby reducing the risk of longitudinal collisions. Based on reducing the risk of rear-end collisions through the dual measures of speed limiting and distance increase, the longitudinal control strategy calculates a new safety threshold to optimize longitudinal safety redundancy and improve the ability to cope with sudden risks. The lateral control strategy employs lateral controllers such as the Stanley algorithm (a path tracking algorithm), Pure-Pursuit algorithm, or MPC (Model Predictive Control) algorithm. Specifically, it forms a discrete pathpoint sequence using a reference trajectory and a predicted trajectory, quantizes this sequence, and generates steering commands to correct the vehicle's lateral deviation. The lateral control strategy includes the following three steps: Step 1: Based on the reference trajectory as the foundation, and building upon the short-term vehicle motion trend reflected by the predicted trajectory, add a lateral offset to the side furthest from traffic participants. This lateral offset is typically set to a value... m forms a future discrete path point sequence, such as {(xi,yi,ti)} (i=1..K), where (xi,yi) is the coordinate of the i-th discrete path point, and ti is the timestamp of the i-th discrete path point. The setting of the lateral offset must ensure that the path does not cross the line and stays away from the risk after the offset. The time length of the discrete path point sequence is set according to the requirements. Step 2: Calculate the curvature and aiming point of each path point in the discrete path point sequence. The aiming distance is usually dynamically adjusted according to the vehicle speed. The higher the vehicle speed, the farther the aiming distance. Step 3: Based on the extracted aiming point or curvature data, the lateral controller calculates the steering command through its internal control law. This steering command is usually output in the form of front wheel angle, in radians or degrees. In some scenarios, the steering command can also be converted into curvature or yaw rate, depending on the underlying actuator interface.

[0047] After receiving the steering command, the vehicle executes the steering operation to gradually correct the lateral error. During this process, strict adherence to vehicle dynamics constraints is required to avoid instability caused by oversteering or sharp turns. After the correction operation is completed, the complete time-series data of the deviation distance, dynamic radius, first prediction time domain, safety threshold, lateral offset, and steering command, as well as the sequence of driving trajectory points after the vehicle executes the steering command and target driving data, are recorded in the log. This provides a basis for subsequent traceability and analysis. DRHG can use this log to make minor optimizations to its parameters.

[0048] In summary, by combining real-time risk assessment with corrective control decisions, the length of the first prediction time domain and the control strategy can be adjusted according to the first state, effectively suppressing error accumulation and providing reliable assurance for the safe operation of autonomous driving systems in complex environments.

[0049] In some implementations, after updating the first positioning or the first prediction time domain based on the first state, the method further includes: generating a preset number of short-term trajectories based on the vehicle's historical trajectory and predicted trajectory; determining a first risk quantity based on all short-term trajectories; and adjusting the updated first prediction time domain based on the first risk quantity.

[0050] In this embodiment of the invention, the historical trajectory includes a sequence of trajectory points of the vehicle over a recent period (e.g., 5 seconds), and may also include a sequence of driving trajectory points after the vehicle performs DRHG correction operations. The preset number is determined based on computing resources and accuracy requirements, and is typically an integer value between 64 and 200.

[0051] Based on historical and predicted trajectories, unobserved influencing factors in the real-time prediction scenario are reconstructed using a pre-defined encoder. A predetermined number of possible trajectories, i.e., short-term trajectories, are then inferred based on these influencing factors. Tail risk and collision probability are quantified by calculating risk data in each short-term trajectory, yielding a first risk quantity. The first prediction time domain is further optimized based on this first risk quantity, such as... Figure 5 As shown, through hypothetical extrapolation of the aforementioned multi-dimensional influencing factors, the potential risks of a single predicted path are exposed in advance, the robustness of the current decision is verified, and a more comprehensive decision-making basis is provided for dynamic adjustments. This avoids the problem of decreased reliability of predicted trajectories in extreme scenarios due to the predictive model's reliance on large-scale real road data.

[0052] The preset encoder is CAVE-IV (Conditional Variational Autoencoder with Instrumental Variables). CAVE-IV is a generative model capable of learning conditional probability distributions. Its encoder part is used to infer the distribution characteristics of latent variables from known conditions, and can be represented as follows: In this context, X represents the known input conditions, and U represents the latent variables. A latent variable is a multidimensional numerical vector used to characterize factors that cannot be directly observed but affect future risks, such as changes in road surface adhesion coefficient, sensor measurement bias, actuator response delay, and the behavioral preferences of other traffic participants. In addition to the historical and predicted trajectories mentioned above, the known conditions for CAVE-IV can also include past road events, high-precision map geometry, the most recent DRHG control results, and instrumental variables (IVs). IVs include vehicle wiper switches, ABS / ESP (Anti-lock Braking System / Electronic Stability Program) triggering, wheel slippage indicators, light levels, camera exposure, and positioning quality.

[0053] CAVE-IV calculates the posterior distribution parameters of latent variables, such as mean and variance, and samples specific instances of the latent variables, such as sampling K (i.e., a predetermined number) latent variables U(k). Each sampled value is then input into the decoder along with the known condition X. Generate K possible trajectories During the decoding and generation process, reasonable perturbations can be added to the behavior of other traffic participants, such as speed fluctuations, changes in reaction delays, or different yielding strategies, to make the generated short-term trajectories more diverse and closer to reality.

[0054] In some implementations, determining a first risk quantity based on all short-term trajectories includes: determining a collision probability matrix for all traffic participants in the real-time environment based on all short-term trajectories; determining the corresponding risk loss value for each short-term trajectory; determining a tail risk index based on all risk loss values; and using the collision probability matrix and the tail risk index as the first risk quantity.

[0055] In this embodiment of the invention, during the calculation of the risk loss value corresponding to each short-term trajectory, the risk data for the vehicle and each short-term trajectory is calculated. The calculation process is the same as the calculation process for the risk data of the predicted trajectory, and will not be repeated here. Subsequently, the risk loss value corresponding to each short-term trajectory is calculated based on the risk data. The formula for calculating the risk loss value of a short-term trajectory based on the risk data is as follows: ; in, Let the risk loss value be the k-th shortest path. Let be the minimum gap of the k-th shortest path. Let be the minimum distance of the k-th shortest path. This indicates the preset time safety threshold. Preset time-sensitive danger threshold. This is a preset space safety threshold. and This is a weighting coefficient, typically starting at 0.5, but can be adjusted according to the needs of the scenario; for example, it can be increased if time urgency is emphasized. .

[0056] In this embodiment of the invention, during the process of determining the tail risk index based on all risk loss values, the risk loss values ​​of all short-term trajectories are sorted from smallest to largest, and the risk loss values ​​located at a preset quantile in the sort are taken as the risk value. The average of all risk loss values ​​exceeding the preset quantile in the sort is then taken as the tail risk index. The tail risk index is a statistical value used to quantify extreme risk conditions, reflecting the average performance of the prediction model under worst-case scenarios.

[0057] In this embodiment of the invention, during the process of determining the collision probability matrix of all traffic participants in a real-time environment based on all short-term trajectories, the expected collision time and spatial gap between each pair of traffic participants in each short-term path are calculated time-by-time. The collision probability of each traffic participant pair is calculated based on these expected collision times and spatial gaps, and the collision probability matrix is ​​obtained from all traffic participant pairs and their collision probabilities. For each traffic participant pair (i,j), the proportion of collisions occurring in all short-term trajectories is statistically analyzed to obtain the collision probability matrix. The formula for the collision probability matrix is ​​as follows: [Intersection / collision occurs within window T at (i,j)]; in, Let be the collision probability matrix. As an indicator function, we statistically analyze the collision probability of each traffic participant pair (i,j) in each short-term trajectory. A probability of intersection or collision is recorded as 1, otherwise as 0. Then, we calculate the average of all short-term trajectories, resulting in a probability value between 0 and 1, which intuitively reflects the collision risk between any two traffic participants.

[0058] It should be noted that, as Figure 5 As shown in A1, A2, B1, and B2, DRHG and CVAE-IV are two parallel modules. DRHG and trajectory prediction operate at a high frequency (e.g., 10Hz), while CVAE-IV operates at a low frequency (e.g., 1–2Hz). This high-low frequency combination is a reasonable architecture based on the functional positioning, computational complexity, and real-time requirements of autonomous vehicles. The first risk quantity output by CVAE-IV in each evaluation is continuously supplied to DRHG until a new first risk quantity is generated. This ensures that DRHG always has a stable risk quantity to support its decision-making and control during the interval between two CVAE-IV evaluations, preventing interruptions in risk assessment. This achieves a seamless connection between accurate assessment and real-time control, ensuring both safety redundancy and meeting real-time requirements.

[0059] In summary, the above methods enable the exploration of future possibilities and risk assessment, providing important safety monitoring and early warning capabilities for autonomous driving systems. They can continuously assess the system's risk status during online operation and provide precise parameter tuning guidance for each module, effectively improving the system's safety performance and robustness in complex environments.

[0060] In some implementations, obtaining the predicted trajectory of the vehicle in a first prediction time domain at the first location includes: obtaining image information of the vehicle at the first location; extracting several first viewpoint features from the image information; determining a first mixed feature based on all the first viewpoint features; and inputting the first prediction time domain and the first mixed feature into a pre-trained prediction model to obtain the predicted trajectory.

[0061] In this embodiment of the invention, the image information is multi-view perception information acquired by sensors such as cameras and LiDAR point clouds. The first view feature is a scene representation extracted from the image information from multiple viewpoints, including but not limited to bird's-eye view (BEV) features, forward-looking view (FV) features, and driver's view (DV) features. A hardness index is calculated based on the first view feature, and the multi-view features are weighted and fused according to the hardness index to generate a first mixed feature dominated by the viewpoint corresponding to the high hardness index. Finally, the first mixed feature is input into a pre-trained prediction model to infer the future trajectory distribution of the vehicle within the first prediction time domain, i.e., the predicted trajectory.

[0062] The extraction of several first-person perspective features includes, but is not limited to, the following methods: After transforming the states of all traffic participants in each frame of image information to a coordinate system with the vehicle as the origin, the bird's-eye view is generated through rasterization, such as... Figure 3 As shown. For example, a 50m×50m grid map is constructed centered on the vehicle, with a grid resolution of 0.25m. The vehicle and pedestrians are projected onto the grid, and their positions, speeds, and heading angles are marked by channels, forming a 3-channel feature map. The forward-looking view is generated by projecting the target onto a fan-shaped grid centered on the vehicle, as shown. Figure 4 As shown. For example, a fan-shaped area within ±60° and 0-60m in front of the vehicle is taken and divided into grids with "angle intervals of 1° and distance intervals of 0.5m". The pedestrian's position is converted to "distance r=22.5m and azimuth angle θ=5°" and filled into the "occupancy status" and "relative speed" channels of the grid to form a 2-channel feature map. The driver's perspective is generated based on the forward-looking perspective, with the viewpoint shifted to the driver's eye point and the vehicle body occlusion effect taken into account. For example, with the driver's eye point as the viewpoint, the area is divided into left, front, and right sectors according to the window pillar. Ray detection is performed on each sector. If the pedestrian is occluded by the right-side pillar at t=3.0s and exits the occlusion at t=3.5s, the pedestrian's state during the visible period is projected onto the fan-shaped grid, and the "occluded area" and "visible target" are marked to form a 3-channel feature map.

[0063] In some implementations, determining the first hybrid feature based on all first viewpoint features includes: calculating a first hardness index for each first viewpoint feature; determining a target viewpoint feature based on the first hardness index that meets a preset index; and obtaining the first hybrid feature based on the target viewpoint feature and a preset reference viewpoint feature.

[0064] In this embodiment of the invention, the preset index can be the maximum value among all calculated first hardness indices. The baseline viewpoint feature is generally a bird's-eye view feature, which is a 256-dimensional feature vector. Four hardness components—occlusion rate, contrast loss, information gain loss, and risk term—are calculated for each first viewpoint feature, and the sum of these four hardness components is taken as the first hardness index. The first viewpoint feature corresponding to the first hardness index with the largest value is taken as the target viewpoint feature. The target viewpoint feature is then weighted and fused with the preset baseline viewpoint feature to obtain the first mixed feature.

[0065] The occlusion rate is calculated as the proportion of occluded grid cells to the total number of field cells in each viewpoint. Contrast loss assesses the sharpness of the image or geometric boundaries, calculated using grid boundary density; lower edge density indicates more severe contrast loss. Information gain loss quantifies the proportion of effective visibility for critical risk objects. The risk term R(v) assesses the urgency of critical objects within the viewpoint, taking the reciprocal of the minimum time distance or the normalized value of the maximum collision probability for critical objects within the viewpoint. After normalizing each component, a weighted summation is performed to obtain the first hardness index H(v), as shown in the following formula: ; Where Oc(v) is the occlusion rate. Due to lack of contrast, Ī Due to missing information gain, This is a risk item. , , and These are the weights for occlusion rate, contrast loss, information gain loss, and risk items, respectively, with a weight sum of 1. Specific values ​​are set based on experience, such as... .

[0066] The target viewpoint features and the baseline viewpoint features are weighted and fused to obtain the first mixed feature. This first mixed feature vector is a 256-dimensional feature vector that encapsulates key scene information and is used as input for the subsequent trajectory prediction model. The first mixed feature z is shown in the following formula: ; in, The visual feature with the highest first hardness index. The viewpoint features are based on the baseline viewpoint. is the weighting factor, and its value range is [0.3, 0.8].

[0067] In some implementations, after updating the first positioning or the first prediction time domain based on the first state, the method further includes: generating a preset number of short-term trajectories based on the vehicle's historical trajectory and predicted trajectory; determining a first risk quantity based on all short-term trajectories; and adjusting a safety threshold, a preset radius function, and the updated first prediction time domain based on the first risk quantity.

[0068] In this embodiment of the invention, the calculated tail risk index is normalized to obtain... ,according to The system dynamically scales the first prediction time domain, adjusts the contraction coefficient, adjusts the safety threshold, and adjusts the downward adjustment of the target speed and the upward adjustment of the target following distance in the longitudinal control strategy. Simultaneously, the collision probability matrix is ​​used to guide the selection of the lateral avoidance direction, such as guiding the direction and magnitude of the lateral offset in the lateral control strategy.

[0069] During the process of adjusting the first prediction time domain based on the first risk quantity, when As the value increases, indicating a rise in tail risk, the length of the first prediction time domain is shortened accordingly, focusing the prediction trajectory on near-term deterministic predictions. When When the value decreases, indicating a reduction in tail risk, the length of the first prediction time domain is extended accordingly, enabling longer-term forward planning.

[0070] During the process of adjusting the preset radius function based on the first risk amount, A positive correlation is established with the contraction coefficient in the dynamic radius calculation function. When When the value of the shrinkage coefficient is increased, the dynamic radius becomes more sensitive to changes in the trajectory deviation distance, thereby tightening the tolerance boundary of the deviation distance.

[0071] During the process of adjusting the safety threshold based on the first risk level, the following will be... Establish a positive correlation with spatial security thresholds and temporal security thresholds. As the threshold increases, the spatial safety threshold and temporal safety threshold requirements are correspondingly raised, making the first state judgment criteria for the predicted trajectory more stringent.

[0072] based on Adjusting the vertical control strategy, by A positive correlation is established with the magnitude of the target speed reduction and the magnitude of the target following distance increase. When When the target speed is increased, the downward adjustment of the target speed is increased accordingly, while the upward adjustment of the target following distance is increased. By reducing the vehicle speed and increasing the following distance, the safety margin is improved.

[0073] In using a collision probability matrix to guide lateral avoidance, the collision probability distribution of the vehicle and other road users in the matrix is ​​analyzed to identify the direction with the highest collision probability. Based on this, the direction of the lateral offset is determined, directing it away from high-risk areas. Simultaneously, the magnitude of the lateral offset is determined according to the value of the highest collision probability; the higher the probability, the larger the offset.

[0074] In summary, through normalized risk values... By adaptively adjusting relevant prediction parameters and using the collision probability matrix to guide the avoidance direction, autonomous vehicles can intelligently adjust their behavior characteristics according to the real-time risk level, ensuring efficiency in low-risk situations and prioritizing safety in high-risk situations, significantly improving the overall robustness of the system in complex scenarios.

[0075] In this embodiment of the invention, a pre-training method for the aforementioned prediction model is also provided, comprising: Based on preset condition parameters and simulation environment, a trajectory sequence is generated; several second-view features are determined based on the trajectory sequence, and a second mixed feature is determined based on all second-view features; the second mixed feature and the first prediction time domain are input into the prediction model to obtain the first trajectory; a second risk quantity is determined based on the first trajectory and it is determined whether it has reached a convergence state; in response to not reaching a convergence state, the preset condition parameters and the model parameters of the prediction model are adjusted based on the second risk quantity, and the step of generating the trajectory sequence is returned until the second risk quantity reaches a convergence state, resulting in a pre-trained prediction model. Further, determining the second risk quantity based on the first trajectory includes: determining a second state based on the trajectory sequence and the first trajectory; in response to an abnormal second state, updating the trajectory sequence based on the first trajectory, and returning to the step of determining several second-view features based on the trajectory sequence; in response to a normal second state, updating the first prediction time domain according to a preset time domain increment, and returning to the step of obtaining the first trajectory, until the first prediction time domain reaches a time domain threshold or the first trajectory reaches a convergence state, resulting in an updated first trajectory and trajectory sequence; based on the updated first trajectory and trajectory sequence, a preset number of second trajectories are generated, and a second risk quantity is determined based on all second trajectories.

[0076] This pre-training method constructs a closed-loop iterative training framework in a simulation environment with tail risk as the optimization objective. It automatically generates extreme scenario data and optimizes model parameters, thereby obtaining a predictive model capable of efficiently handling high-risk scenarios. The core of this pre-training method lies in dynamically adjusting the data generation and model training strategies using risk assessment results, continuously strengthening the model in its weakest areas.

[0077] The preset condition parameters include at least event intensity parameters controlling the injection frequency of extreme events, and physical disturbance parameters defining uncertainties such as road surface friction coefficient and braking delay. The physical disturbance parameters include at least the road surface adhesion coefficient, drag coefficient, road slope, and crosswind speed and direction, and are set through the physics engine's application programming interface. The event intensity parameter controls the expected number of rare events occurring per unit time. The simulation environment refers to a virtual testing environment with consistent vehicle dynamics, capable of simulating vehicle motion under various physical conditions. Trajectory sequence generation is achieved by running the simulation environment and injecting events defined by the preset condition parameters. This trajectory sequence contains time-series state data of the vehicle and traffic participants, along with corresponding multi-dimensional labels. The multi-dimensional labels include event labels, physical labels, risk labels, and reference point labels.

[0078] Initialize the simulation environment, set the simulation iteration step size, and load the high-precision map and initial traffic flow. The simulation iteration step size is a discrete time interval used for the stability of numerical integration and aligned with the data label time. Based on the event intensity parameter, generate a set of event trigger times {t} on the simulation time axis according to the Poisson process.e}, where t e This represents the time when the e-th event occurs. A Poisson process can effectively simulate the random occurrence of rare and independent events; by adjusting the event intensity parameter, the sparsity of events in the scene can be uniformly controlled. At each event trigger time t... e The system randomly selects event types from a pre-defined event library and samples their specific parameters. Event types include, but are not limited to, sudden obstacle appearance, pedestrian peeking out from behind, and emergency braking of the vehicle in front. Event parameters include the location of key objects, their speed, acceleration magnitude, and minimum safety clearance threshold. The minimum safety clearance threshold is the minimum safe distance between pedestrians and vehicles or between vehicles. The sampling process is constrained to be conducted within the reachable area of ​​the environment to ensure that event generation conforms to real-world traffic rules; for example, pedestrian peeking out from behind events only occurs on sidewalks, crosswalks, or in lane areas. The extracted events undergo geometric validity and dynamic reachability checks. Geometric validity checks ensure that the event generation point does not overlap with static obstacles and that its orientation is reasonable; dynamic reachability checks ensure that the event's speed or acceleration magnitude does not exceed limits. Simultaneously, short-time window predictions are performed to check whether an unavoidable serious collision will occur in the near future. Events that fail the checks are resampled or discarded to avoid generating invalid scenarios.

[0079] The validated events are injected into the simulation environment, and integration is performed according to the simulation iteration step size. The simulation process fully considers physical factors such as tire sideslip characteristics, wind resistance, slope, and crosswinds, and outputs accurate vehicle kinematics, i.e., state data, including position coordinates, heading angle, velocity, acceleration, and yaw rate. If the same event triggers at time t... e Multiple mutually exclusive events are triggered, and conflicts are resolved according to a pre-defined event priority table. The event priority table is based on common sense and safety considerations and is maintained through a configuration file. For events with similar priorities, time-shifting or random selection is used to ensure scenario diversity. For example, the event priority table could be: pedestrian crossing > sudden appearance > sudden braking by the vehicle in front > static obstacle. Time-series state data of the vehicle and key traffic participants are recorded, and multi-dimensional labels are automatically generated. Event labels record event type, occurrence time, and severity; physical labels record physical constraint violations such as maximum lateral acceleration, collision occurrence, and lane departure; risk labels record risk indicators such as minimum time distance, minimum gap, and collision probability; and reference point labels mark key interaction moments and occlusion change moments. The time-series state data and multi-dimensional labels are output in a unified JSON format, with layered organization of agent, label, and physics information, such as... Figure 2As shown, the resulting trajectory sequence ensures that training, evaluation, and auditing do not interfere with each other. Explicit timestamps and multi-dimensional labels ensure that the model output is accurately aligned with and reproducible from the simulation. The exported trajectory sequence can be used to generate features from different perspectives, as well as for subsequent bias correction and adaptive optimization. The application of the data in the trajectory sequence in subsequent bias correction and adaptive optimization is shown in Table 1 below.

[0080] Table 1

[0081] Second-view features are scene representations extracted from the time-series state data of the trajectory sequence from multiple perspectives, including bird's-eye view features, forward-looking view features, and driver's view features. All traffic participants in each frame of the trajectory sequence are transformed to a coordinate system with the vehicle as the origin, generating feature representations for three perspectives: BEV (Battery Escape), FV (Front-View), and DV (Driver's View). The second hybrid feature is obtained by calculating the hardness index of each second-view feature. The method for determining the second-view features is the same as that for the first-view features, and the method for determining the second hybrid feature is the same as that for the first hybrid feature; these will not be repeated here. The calculation process for the second risk quantity is the same as that for the first risk quantity; these will not be repeated here. The tail risk index is a statistical value used to quantify extreme risk conditions, reflecting the average performance of the prediction model under worst-case conditions. The second risk quantity reaches convergence when the tail risk index is less than or equal to the index threshold. If the second risk quantity does not reach convergence, for example, if the tail risk index shows that the prediction model exceeds the index threshold in a pedestrian crossing scenario, the preset condition parameters are adjusted, increasing the event intensity parameter of the pedestrian crossing event, and simultaneously increasing the weight of safety loss in model training. Then, based on the new preset condition parameters, a trajectory sequence containing more pedestrian crossing scenarios is regenerated, and the model is trained again. This process is repeated until the tail risk exhibited by the model in extreme scenarios of all event types drops below an acceptable metric threshold.

[0082] In summary, through the sequential and iterative steps described above, an offline pre-training closed loop of data generation, model training, risk assessment, and parameter tuning is achieved. This automatically focuses on hazardous scenarios where the prediction model performs poorly, dynamically generates targeted training samples, and adjusts the optimization direction, thereby efficiently training a prediction model with high safety awareness and robustness, laying a solid foundation for online real-time applications.

[0083] The second state is a warning state regarding the future driving condition of the vehicle, derived from risk assessment in the simulation environment. Its determination process and the updating of the first trajectory are the same as those for the first state and the updating of the predicted trajectory, and will not be repeated here. In response to an anomaly in the second state, DRHG is executed to obtain the new state data of the vehicle in the simulation environment after corrective control, and the trajectory sequence is updated accordingly. Based on the updated trajectory sequence, the updated first trajectory is obtained. The first trajectory reaching convergence includes achieving the required trajectory accuracy, risk control, and physical consistency. Physical consistency means that the proportion of physical violation samples is less than a preset proportion, and the event intensity is stable.

[0084] The updated first trajectory and trajectory sequence are input into a preset encoder, and a preset number of second trajectories are determined based on the output. The method for determining the second trajectory and the second risk quantity is the same as that for the short-term trajectory and the first risk quantity, and will not be repeated here.

[0085] In some implementations, the content of adaptive parameter tuning during the model training phase is described.

[0086] The event intensity update process is as follows: ; in, This represents the update step size, also known as the learning rate, and its value typically ranges from 0.05 to 0.2. This parameter determines the event intensity after each new risk assessment result is received. The magnitude of the adjustment. A larger value indicates a more aggressive adjustment. A smaller value indicates a more stable adjustment. This represents the target risk baseline, typically ranging from 0.2 to 0.4. When... At that time, the risk was deemed too high, requiring increased training difficulty; when If the risk is deemed low, the training difficulty can be appropriately reduced. and This represents the upper and lower safety limits, typically ranging from 0.5 to 5.0. By limiting the event intensity λ within these preset safety limits, this formula enables the linkage adjustment between event intensity and risk assessment results, ensuring that the training difficulty matches the actual risk level.

[0087] Event intensity updates can be further refined into category-based updates, as shown below: ; Where j represents a certain event type, such as a pedestrian crossing from the right or a vehicle cutting in from the left. This represents the event intensity corresponding to this event type. The risk weight for event type j is calculated by summing the collision probabilities of all traffic participant pairs related to this type of event and then normalizing the results, as shown below: ; in, Collision probability matrix The probability of collision between traffic participant p and traffic participant q. Let be the set of traffic participants corresponding to event type j. The denominator is the sum of collision probabilities across the entire scenario. To prevent the denominator from being 0, a fixed minimum value can be set, such as... .

[0088] Physical disturbance parameters The update process is as follows: ; in, The mapping coefficients represent the normalized risk values. The coefficient that maps to the disturbance amplitude. for The value after normalization. The maximum perturbation amplitude of each physical perturbation parameter is limited to prevent excessive distortion. Physical perturbation parameters It can be represented as a vector, such as including components such as changes in road surface friction coefficient, braking delay, and drag coefficient.

[0089] The training loss weights are updated in the following way: ; ; in, and As base value weight, and This is the sensitivity coefficient. , and and This defines the range of weight values. In this way, the weights of safety-related loss items are dynamically linked to the current risk level; the higher the risk, the greater the safety weight.

[0090] The training loss function is constructed in the following way: ; Where L is the total loss function, To normalize the bias distance, the prediction model parameters and CVAE-IV parameters are optimized through backpropagation. This is the clearance violation rate term, which represents the degree to which the minimum safety clearance is violated, and its value ranges from 0 to 1.

[0091] In summary, adaptive parameter tuning effectively maps risk assessment results to model parameters, dynamically adjusting training difficulty and optimization direction based on the model's actual performance. This ensures that data generation and model training consistently focus on the weakest and most dangerous scenarios, thereby efficiently improving the overall model performance and safety. This risk feedback-based closed-loop optimization mechanism is a crucial technical approach to addressing the long-tail problem in autonomous driving.

[0092] The technical solution of the present invention will be described in detail below with reference to the embodiments.

[0093] I. Regarding the model training phase: (1) Environmental configuration The scenario is a two-lane, two-way urban road. The initial road condition is sunny turning to light rain, with a road adhesion coefficient of μ=0.8. There is a light crosswind with a speed of 3.0 m / s, directed perpendicular to the lane to the right. The vehicle is a compact sedan with an initial speed of 8 m / s, traveling at a constant speed along the right lane. The surrounding traffic flow includes one vehicle in front, 20 m away from the vehicle, with a speed of 7.5 m / s.

[0094] (2) Extreme events At simulation time t=3.2s, the "pedestrian suddenly crossing the right blind spot" event is triggered. The pedestrian enters the zebra crossing from the right sidewalk blind spot with a walking speed of 1.2m / s, a severity of 0.7, and an initial position of 1.5m to the right of the vehicle. This is an extreme scenario with "short reaction time and high collision risk".

[0095] (3) Initial parameter configuration Event generation parameters: Initial event strength is The simulation step size is dt = 0.1 s; DRHG core parameters: Initial dynamic radius is The gain is k=0.5, and the initial value of the first prediction time domain is... The visual mixing coefficient is α=0.7; Physical and control parameters: The physical disturbance parameter θ includes road friction μ=0.8 and braking delay 0.10s; the adaptive parameter tuning step size is η1=0.1, and the target risk baseline is β=0.20; Model and evaluation parameters: training loss weights are μ1=0.5 and μ2=0.5; the number of CVAE-IV samples is K=32 and the tail risk quantile is α=0.95.

[0096] (4) Iteration 1 By injecting extreme events with λ0=2 and advancing the simulation, the key risk indicators for the scenario were calculated as follows: minimum safe time distance = 1.1s, minimum safe clearance = 0.7m, both of which were determined to be below the safety threshold for urban roads. Multi-view hardness assessment showed that the driver's view had the highest hardness; therefore, the driver's view and the BEV baseline view characteristics were combined with α=0.7.

[0097] The hybrid feature input trajectory predictor is combined with DRHG execution control, and the deviation distance is observed multiple times. In such cases, time-domain freeze and correction operations are triggered.

[0098] The CVAE-IV module generates 32 counterfactual trajectories and calculates the vehicle-pedestrian collision probability P. (ego,ped) =0.26, tail risk CVaR 0.95 =0.27, significantly higher than the target baseline β=0.20.

[0099] Based on the risk exceeding the limit, the event intensity is updated to λ1≈2.007. The weight μ1 of the safety term in the training loss is increased to 0.535, the physical disturbance parameter θ is slightly increased, and the road friction is increased by 0.02.

[0100] (5) Iteration 2 The model is trained using encrypted extreme samples generated by λ1 in iteration 1. The increased number of boundary samples makes the model more sensitive to the feature recognition of "sudden pedestrians", and can capture the signs of pedestrians entering the field of vision 0.3 seconds in advance.

[0101] Reassess the scenario risks, collision probability P (ego,ped) Decreased to 0.21, CVaR 0.95 =0.23, the risk indicator is approaching the target baseline.

[0102] The event intensity λ2 is updated to approximately 2.014 according to the feedback logic, the loss weight μ1 is slightly adjusted to 0.517, and the physical parameter θ continues to be adjusted slightly. The magnitude of parameter change is narrower than that of iteration 1, showing a trend towards stabilization.

[0103] (6) Iteration 5 After the first four rounds of iteration and optimization, the fifth iteration reached the following convergence criteria: The trajectory accuracy meets the standard: the average displacement error is 0.38m and the final displacement error is 0.80m, both of which are determined to be below the accuracy threshold for urban road trajectory prediction. Risk control meets standards: CVaR 0.95 =0.20, precisely matching the target risk baseline β; vehicle-pedestrian collision probability P (ego,ped) =0.14, a decrease of 46% compared to iteration 1; Physical consistency met: the proportion of physical violation samples was <0.3%, and the event intensity was stable at λ≈2.02.

[0104] At this point, the model training is complete, and it can be deployed to the online inference stage.

[0105] II. Regarding the real-time prediction stage: The prediction model obtained in iteration 5 and DRHG are deployed to the online system. When a similar "pedestrian crossing the right blind spot" scenario is triggered, the following real-time control process is executed: (1) Triggering conditions and control decisions When the positioning module detects the deviation between the vehicle's actual trajectory and the reference trajectory... Immediately activate the security control strategy.

[0106] (2) Temporal freeze and longitudinal control The prediction time domain is frozen, and the target speed is quickly reduced to v=min (current speed, 6.0 m / s) using DRHG. Simultaneously, the following distance is strategically increased by 2.0 m to reduce the risk of longitudinal collision. Furthermore, the lateral offset to the left is rapidly applied based on the reference trajectory using DRHG. =0.4m, generate a short-window safe path, and calculate the turning command. =0.03rad, guide the vehicle to make a slight adjustment to the left safe area, and pull back the lateral error.

[0107] (3) Counterfactual assessment The risk level is assessed, and based on whether the risk level converges, the DRHG module's online safety actions are triggered, such as freezing the prediction time domain, reducing the target speed, performing lateral correction, or even triggering a takeover warning to prevent the danger from escalating in the first instance. In addition, the entire real-time control process is recorded and marked, and then fed back to the offline training end to optimize the prediction model. Ultimately, this achieves continuous improvement in prediction accuracy and control stability in similar scenarios, resulting in an optimization effect that becomes safer with use.

[0108] Based on the same inventive concept, according to another aspect of the present invention, embodiments of the present invention also provide an automatic driving control system 10, such as... Figure 6 As shown, the autonomous driving control system includes at least one sensor 101 for acquiring the first positioning of the vehicle; and at least one processor 102 for acquiring the final predicted trajectory based on the first positioning using an optimization method for autonomous driving trajectory prediction as described in any of the above embodiments.

[0109] Finally, it should be noted that those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium for the program can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. The above computer program embodiments can achieve the same or similar effects as any of the corresponding foregoing method embodiments.

[0110] Those skilled in the art will also understand that the various exemplary logic blocks, modules, circuits, and algorithm steps described in conjunction with the disclosure herein can be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the functionality of various illustrative components, blocks, modules, circuits, and steps has been generally described. Whether this functionality is implemented as software or as hardware depends on the specific application and the design constraints imposed on the system as a whole. Those skilled in the art can implement the functionality in various ways for each specific application, but such implementation decisions should not be construed as departing from the scope of the embodiments disclosed herein.

[0111] The above are exemplary embodiments disclosed in this invention. However, it should be noted that various changes and modifications can be made without departing from the scope of the embodiments of this invention as defined by the claims. The functions, steps, and / or actions of the methods according to the disclosed embodiments described herein do not need to be performed in any particular order. The sequence numbers of the disclosed embodiments of this invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. Furthermore, although the elements disclosed in the embodiments of this invention may be described or claimed individually, they may be understood as multiple unless explicitly limited to a singular number.

[0112] It should be understood that, as used herein, the singular form “a” is intended to include the plural form as well, unless the context clearly supports an exception. It should also be understood that, as used herein, “and / or” refers to any and all possible combinations of one or more of the associated listed items.

[0113] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples. Within the framework of the invention, technical features of the above embodiments or different embodiments can be combined, and many other variations of different aspects of the invention exist, which are not provided in the details for the sake of brevity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

Claims

1. An optimization method for predicting the trajectory of autonomous driving, characterized in that, include: Obtain the predicted trajectory of the vehicle in the first prediction time domain at the first location; Based on the reference trajectory and the predicted trajectory, a first state is determined; Update the first positioning or the first prediction time domain based on the first state, and return to the step of obtaining the prediction trajectory until the prediction trajectory meets the preset conditions to obtain the final prediction trajectory.

2. The optimization method for autonomous driving trajectory prediction according to claim 1, characterized in that, The step of determining the first state based on the reference trajectory and the predicted trajectory includes: The deviation distance is determined based on the first positioning and the reference trajectory; Risk data is determined based on the predicted trajectory; The first state is determined based on both the risk data and the deviation distance.

3. The optimization method for autonomous driving trajectory prediction according to claim 1, characterized in that, The step of updating the first positioning or first prediction time domain based on the first state includes: In response to the first state being normal, the first predicted time domain is updated according to the preset time domain increment; In response to the first state anomaly, target driving data and steering commands are determined based on the vehicle's current driving data, the reference trajectory, and the predicted trajectory. The steering commands and target driving data are executed, the first positioning is updated based on the position information after execution, and it is determined whether to update the first prediction time domain based on the target driving data.

4. The optimization method for autonomous driving trajectory prediction according to claim 1, characterized in that, After the step of updating the first location or the first prediction time domain based on the first state, the method further includes: Based on the vehicle's historical trajectory and the predicted trajectory, a preset number of short-term trajectories are generated; A first risk quantity is determined based on all the short-term trajectories, and the updated first prediction time domain is adjusted based on the first risk quantity.

5. The optimization method for autonomous driving trajectory prediction according to claim 4, characterized in that, Determining the first risk quantity based on all the short-term trajectories includes: Based on all the aforementioned short-term trajectories, determine the collision probability matrix for all traffic participants in the real-time environment; Based on each of the short-term trajectories, a corresponding risk loss value is determined, and a tail risk indicator is determined based on all of the risk loss values. The collision probability matrix and the tail risk index are used as the first risk quantity.

6. The optimization method for autonomous driving trajectory prediction according to claim 1, characterized in that, The acquisition of the predicted trajectory of the vehicle in the first prediction time domain at the first location includes: Acquire image information of the vehicle at the first location, and extract several first-view features from the image information; Based on all the first viewpoint features, determine the first mixed feature; The first prediction time domain and the first mixed feature are input into the pre-trained prediction model to obtain the predicted trajectory.

7. The optimization method for autonomous driving trajectory prediction according to claim 6, characterized in that, The step of determining the first mixed feature based on all the first viewpoint features includes: Calculate the first hardness index for each first viewpoint feature, and determine the target viewpoint feature based on the first hardness index that meets the preset index. The first hybrid feature is obtained based on the target viewpoint features and the preset reference viewpoint features.

8. The optimization method for autonomous driving trajectory prediction according to claim 2, characterized in that, Determining the first state based on both risk data and the deviation distance includes: The dynamic radius is determined based on the deviation distance and the preset radius function; In response to the deviation distance not being greater than the dynamic radius and the risk data not being less than the safety threshold, the first state is determined to be normal; In response to the deviation distance being greater than the dynamic radius or the risk data being less than the safety threshold, the first state is determined to be abnormal.

9. The optimization method for autonomous driving trajectory prediction according to claim 8, characterized in that, After the step of updating the first location or the first prediction time domain based on the first state, the method further includes: Based on the vehicle's historical trajectory and the predicted trajectory, a preset number of short-term trajectories are generated; A first risk quantity is determined based on all the short-term trajectories, and the safety threshold, the preset radius function, and the updated first prediction time domain are adjusted based on the first risk quantity.

10. An automatic driving control system, characterized in that, include: At least one sensor, said sensor being used to acquire the initial position of the vehicle; At least one processor, the processor acquiring a final predicted trajectory based on the first positioning using the method as described in any one of claims 1-9.

Images