Obstacle trajectory prediction method, vehicle, storage medium and program product

Abstract

This application provides an obstacle trajectory prediction method, a vehicle, a storage medium, and a program product. The method includes: acquiring current perception data of an obstacle, including its location and motion state information, at the current moment; based on the current perception data of the obstacle, acquiring predicted obstacle trajectories for a preset future time period using multiple motion models; selecting the optimal motion model from among the multiple motion models based on the cumulative cost of each motion model, wherein the optimal motion model is the one with the lowest cumulative cost; if the currently used motion model is not the optimal motion model, then determining the first final obstacle prediction trajectory using the obstacle prediction trajectory of the optimal motion model. This improves the reliability of obstacle trajectory prediction, provides the vehicle with accurate and timely safety decision-making basis, and enables accurate obstacle avoidance operations to be taken in advance, reducing the risk of vehicle collisions.

Description

Technical Field

[0001] This application relates to the field of vehicle technology, specifically to an obstacle trajectory prediction method, a vehicle, a storage medium, and a program product. Background Technology

[0002] When a vehicle is in automatic parking or low-speed autonomous driving scenarios, the movement trajectories of surrounding dynamic obstacles (such as vehicles, two-wheeled vehicles, pedestrians, etc.) are flexible and unpredictable, increasing the difficulty of the vehicle's environmental perception and active obstacle avoidance. Failure to perceive the movement trajectories of surrounding obstacles in advance may lead to potential vehicle collision hazards. Therefore, accurately predicting the future movement trajectories of surrounding dynamic obstacles is a prerequisite for ensuring the vehicle's safe decision-making and control.

[0003] In related technologies, single motion models such as uniform linear model or constant turning rate model are usually used to extrapolate and predict the motion trajectory of obstacles.

[0004] However, in real-world traffic scenarios, obstacles exhibit complex and varied motion patterns. A single motion model struggles to simultaneously adapt to multiple motion modes, such as straight-line cruising, stable cornering, serpentine maneuvers, and sudden turns. This can lead to significant discrepancies between the predicted and actual motion trajectories, reducing the reliability of obstacle trajectory predictions. Consequently, this can interfere with the vehicle's subsequent safety decisions and obstacle avoidance control, creating potential collision hazards.

[0005] It should be noted that the information disclosed in the background section of this application is intended only to enhance the understanding of the general background of this application, and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] This application provides an obstacle trajectory prediction method, a vehicle, a storage medium, and a program product to address the problem in the prior art where the prediction of an obstacle's trajectory using a single motion model results in a large deviation between the predicted trajectory and the actual trajectory, thus creating a potential vehicle collision hazard.

[0007] In a first aspect, embodiments of this application provide an obstacle trajectory prediction method, including: At the current moment, acquire the current perception data of the obstacle, wherein the perception data includes location information and motion state information; Based on the current perception data of the obstacle, multiple motion models are used to predict the motion trajectory of the obstacle within a preset time period in the future, and the obstacle prediction trajectory of each motion model is obtained. Based on the cumulative cost of each of the aforementioned motion models, the optimal motion model is selected from among the various motion models. The optimal motion model is the motion model with the lowest cumulative cost, whereby the cumulative cost is used to characterize the degree of deviation between the predicted trajectory of the obstacle and the actual trajectory of the obstacle. If the currently used motion model is not the optimal motion model, then the obstacle prediction trajectory of the optimal motion model is used to determine the first final obstacle prediction trajectory.

[0008] In one possible implementation, determining the first final obstacle prediction trajectory using the obstacle prediction trajectory of the optimal motion model includes: The obstacle prediction trajectory of the currently used motion model and the obstacle prediction trajectory of the optimal motion model are fused within the first time interval to obtain the obstacle prediction trajectory within the first time interval. Using the obstacle motion trajectory predicted by the optimal motion model within the second time interval, the predicted obstacle trajectory within the second time interval is determined; The obstacle prediction trajectories in the first time interval and the obstacle prediction trajectories in the second time interval are combined to obtain the first final obstacle prediction trajectory; The first time interval is located before the second time interval.

[0009] In one possible implementation, determining the first final obstacle prediction trajectory using the obstacle prediction trajectory of the optimal motion model includes: The initial obstacle prediction trajectory is determined using the obstacle prediction trajectory of the optimal motion model; The effective duration of the initial obstacle prediction trajectory is determined based on the confidence level of the initial obstacle prediction trajectory. The first final obstacle prediction trajectory is obtained by intercepting the obstacle prediction trajectory within the effective time range from the initial obstacle prediction trajectory.

[0010] In one possible implementation, selecting the optimal motion model from among the multiple motion models based on the cumulative cost of each of the motion models includes: Based on the multidimensional kinematic characteristics of the obstacle, multiple motion models are screened to obtain at least one candidate motion model that matches the motion state represented by the multidimensional kinematic characteristics. Based on the cumulative cost of each of the candidate motion models, the optimal motion model is selected from at least one of the candidate motion models.

[0011] In one possible implementation, prior to filtering multiple motion models based on the multidimensional kinematic characteristics of the obstacle, the method further includes: Obtain the historical trajectory sequence of the obstacle, the historical trajectory sequence including multiple historical perception data of the obstacle; The multidimensional kinematic features of the obstacle are extracted from its historical trajectory sequence.

[0012] In one possible implementation, the step of determining the first final obstacle prediction trajectory using the obstacle prediction trajectory of the optimal motion model if the currently used motion model is not the optimal motion model includes: If the currently used motion model is not the optimal motion model, and the difference between the cumulative cost of the currently used motion model and the cumulative cost of the optimal motion model is greater than or equal to a preset cumulative cost difference threshold, then the obstacle prediction trajectory of the optimal motion model is used to determine the first final obstacle prediction trajectory.

[0013] One possible implementation also includes: If the obstacle is a target type obstacle, the Gaussian process regression algorithm is used to determine the second final obstacle prediction trajectory.

[0014] In one possible implementation, the step of using a Gaussian process regression algorithm to determine the second final obstacle prediction trajectory includes: The Gaussian process regression algorithm is used to determine the second obstacle prediction trajectory to be optimized; physical feasibility constraints are applied to the second obstacle prediction trajectory to obtain the second final obstacle prediction trajectory.

[0015] In one possible implementation, determining the first final obstacle prediction trajectory using the obstacle prediction trajectory of the optimal motion model includes: The first obstacle prediction trajectory to be optimized is determined by using the obstacle prediction trajectory of the optimal motion model; physical feasibility constraints are applied to the first obstacle prediction trajectory to be optimized to obtain the first final obstacle prediction trajectory.

[0016] One possible implementation also includes: Determine whether the identification mark of the obstacle is an identification mark that has been used by other obstacles; If the identifier of the obstacle is one that has been used by other obstacles, then the historical data corresponding to that identifier is deleted.

[0017] Secondly, embodiments of this application provide a vehicle, including: A controller configured to perform the method described in any one of the first aspects.

[0018] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in any one of the first aspects.

[0019] Fourthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method described in any one of the first aspects.

[0020] In this embodiment, multiple motion models are used to predict the obstacle's trajectory within a preset time period based on the obstacle's current perception data, obtaining the obstacle prediction trajectory for each motion model. Based on the cumulative cost of each motion model, the optimal motion model is selected. If the currently used motion model is not the optimal model, the optimal motion model is used to predict the obstacle's trajectory, and a first final obstacle prediction trajectory is determined. This improves the reliability of obstacle trajectory prediction, providing the vehicle with accurate and timely safety decision-making basis, thereby enabling accurate obstacle avoidance operations in advance and reducing the risk of vehicle collisions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of an application scenario provided by an embodiment of this application.

[0023] Figure 2 This is a flowchart illustrating an obstacle trajectory prediction method provided in an embodiment of this application.

[0024] Figure 3 This is a structural schematic diagram of a vehicle provided in an embodiment of this application. Detailed Implementation

[0025] To better understand the technical solution of this application, the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0026] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0027] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0028] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0029] See Figure 1 This is a schematic diagram illustrating an application scenario provided by an embodiment of this application. For example... Figure 1 As shown, this application scenario includes: a vehicle 101, a pedestrian obstacle 102, and an obstacle prediction trajectory 103. When the vehicle 101 is traveling on the road, the pedestrian obstacle 102 appears, and the predicted trajectory of the pedestrian obstacle 102 is the obstacle prediction trajectory 103. Based on the obstacle prediction trajectory 103, the vehicle 101 initiates the corresponding obstacle avoidance strategy in advance to avoid a vehicle collision.

[0030] In practical applications, when a vehicle is in automatic parking or low-speed autonomous driving scenarios, the movement trajectories of surrounding dynamic obstacles (such as vehicles, two-wheeled vehicles, pedestrians, etc.) are flexible and unpredictable, increasing the difficulty of the vehicle's environmental perception and active obstacle avoidance. Failure to perceive the movement trajectories of surrounding obstacles in advance may lead to potential vehicle collision hazards. Therefore, accurately predicting the future movement trajectories of surrounding dynamic obstacles is a prerequisite for ensuring the vehicle's safe decision-making and control.

[0031] In related technologies, single motion models such as uniform linear model or constant turning rate model are usually used to extrapolate and predict the motion trajectory of obstacles.

[0032] However, in real-world traffic scenarios, obstacles exhibit complex and varied motion patterns. A single motion model struggles to simultaneously adapt to multiple motion modes, such as straight-line cruising, stable cornering, serpentine maneuvers, and sudden turns. This can lead to significant discrepancies between the predicted and actual motion trajectories, reducing the reliability of obstacle trajectory predictions. Consequently, this can interfere with the vehicle's subsequent safety decisions and obstacle avoidance control, creating potential collision hazards.

[0033] To address the aforementioned issues, in this embodiment, at the current moment, current perception data of the obstacle, including its location and motion state information, is acquired. Based on this current perception data, multiple motion models are used to obtain predicted trajectories for the obstacle within a preset future timeframe. Based on the cumulative cost of each motion model, the optimal motion model is selected from among the multiple models, with the optimal model being the one with the lowest cumulative cost. If the currently used motion model is not the optimal model, the predicted obstacle trajectory of the optimal motion model is used to determine the first final predicted obstacle trajectory. This improves the reliability of obstacle trajectory prediction, providing the vehicle with accurate and timely safety decision-making support, thereby enabling accurate obstacle avoidance maneuvers to be taken in advance and reducing the risk of vehicle collisions.

[0034] Specifically, the following detailed description is provided in conjunction with the accompanying drawings and specific embodiments.

[0035] See Figure 2 This is a flowchart illustrating an obstacle trajectory prediction method provided in an embodiment of this application. This method can be applied to... Figure 1 In the application scenarios shown, such as Figure 2 As shown, it mainly includes the following steps.

[0036] Step S201: At the current moment, acquire the current perception data of the obstacle.

[0037] In this embodiment, at the current moment, the current perception data of the obstacle is acquired through the perception fusion module. The current perception data refers to the perception data collected at the current moment. This perception data includes position information and motion state information. Specifically, the position information may include the obstacle's position coordinates (x, y), and the motion state information may include the obstacle's velocity v and heading angle θ. Motion state identifiers (e.g., stationary / moving) and tracking stability metrics used to characterize the number of consecutive frames an obstacle has been tracked. In addition to collecting location and motion status information, the collected perception data also includes the obstacle's identification (ID) and the obstacle's type (e.g., vehicle / bicycle / pedestrian, etc.).

[0038] It should be noted that, in addition to collecting the aforementioned sensing data, other data related to obstacle movement can also be collected, but this application embodiment does not impose specific limitations on this.

[0039] In this embodiment, a historical trajectory sequence is independently maintained for each obstacle based on its identification. It is used to store the trajectory points of obstacles within the most recent preset time interval (e.g., 1.5 seconds). The sequence and corresponding timestamp , This represents the number of trajectory points in the historical trajectory sequence. Each trajectory point... Each corresponds to a specific timestamp. The collected perception data includes multiple historical perception data points for obstacles in the historical trajectory sequence.

[0040] In practice, after receiving the current perception data of an obstacle each time, the historical trajectory sequence corresponding to the obstacle needs to be updated. Specifically, the historical trajectory sequence of the obstacle can be updated in the following three ways: 1) If the obstacle's identification already has a corresponding historical trajectory sequence, then the location information of the current sensing data will be appended to the end of the corresponding historical trajectory sequence; 2) If the obstacle's identification is a new identification that appears for the first time, then maintain a new historical trajectory sequence for the obstacle and initialize its historical trajectory sequence with the location information of the current sensing data; 3) If the number of trajectory points in the historical trajectory sequence exceeds the preset capacity limit, the oldest data point at the head of the queue will be automatically removed to keep the length of the historical trajectory sequence constant.

[0041] In practical applications, when a newly appearing obstacle is assigned the identity of a previously disappeared obstacle, the perceived data of the historical trajectory sequence of the old obstacle can affect the prediction of the new obstacle's trajectory, causing a significant discrepancy between the predicted trajectory and the actual trajectory of the new obstacle. Therefore, identity reuse detection needs to be performed on each obstacle.

[0042] In one possible implementation, it is determined whether the obstacle's identifier is an identifier that has been used by other obstacles; if the obstacle's identifier is an identifier that has been used by other obstacles, the historical data corresponding to the identifier is deleted.

[0043] In practice, the tracking stability index of each obstacle that has appeared is saved from the previous moment. Value, denoted as Read the obstacle at the current moment. The value is evaluated using the following criteria: That is, during normal tracking, the same obstacle The value should increase frame by frame. If the judgment condition... If the event is successful, it indicates that the identity of a previously missing obstacle has been reassigned to a newly appearing obstacle, and an identity reuse event is considered to have occurred.

[0044] When an identity reuse event occurs, the historical data corresponding to that identity needs to be deleted. The specific deletion operation is as follows: 1) Clear the historical trajectory sequence corresponding to this identity identifier; 2) Clear the cumulative cost of each motion model corresponding to this identity. and the currently used motion model Record; 3) Clear the transition state (e.g., transition coefficient) of the motion model corresponding to this identity. Transition counter wait).

[0045] After the above deletion operation is completed, the current perception data of the newly appeared obstacle is used as the initial state data to reinitialize the historical trajectory sequence corresponding to the identity, and the current... The value is updated.

[0046] Among them, the cumulative cost The current motion model Records, transition coefficients and transition counter This will be explained in detail below, and will not be elaborated on here.

[0047] In this embodiment of the application, by performing identity reuse detection on each obstacle, the trajectory prediction process of each newly appearing obstacle is completely unaffected by the historical trajectory sequence of the disappeared old obstacles, and all relevant state data about the trajectory prediction of newly appearing obstacles are accumulated from zero.

[0048] Step S202: Based on the current perception data of the obstacle, use multiple motion models to predict the motion trajectory of the obstacle within a preset time period in the future, and obtain the obstacle prediction trajectory of each motion model.

[0049] In this embodiment of the application, based on the current perception data of the obstacle, multiple motion models are run in parallel to predict the motion trajectory of the obstacle within a preset time period in the future, and the obstacle prediction trajectory of each motion model is obtained.

[0050] The motion model refers to a mathematical operation model that predicts the trajectory of an obstacle based on its current perception data and a preset fixed motion law. In this application's embodiments, the main motion models used are the uniform velocity linear model, the uniform acceleration linear model, the constant turning rate model, and the curve fitting model. The following uses these four motion models to predict the obstacle's trajectory within a preset time period based on its current perception data: 1) Uniform linear model: Assuming the magnitude and direction of the velocity remain constant. Let the current perceived data of the obstacle be (x0, y0, θ0, v0), and the predicted position coordinates at a future preset time interval Δt be: x pred =x0+v0·Δt·cos(θ0), y pred=y0+v0·Δt·sin(θ0), θ pred =θ0. Where (x0, y0) are the current coordinates of the obstacle, v0 is the current velocity, and θ0 is the current heading angle. Since it is uniform linear motion, the predicted heading angle θ... pred Equal to the current heading angle θ 0。

[0051] 2) Uniformly Accelerated Linear Model: Assuming constant acceleration and linear velocity. Let the current perceived data of the obstacle be (x0, y0, θ0, v0, α0), where α0 is the current tangential acceleration. The predicted position coordinates at a future preset time Δt are: x pred =x0+s(Δt)·cos(θ0), y pred =y0+s(Δt)·sin(θ0),s(Δt)=v0·Δt+½·α0·Δt 2 v(Δt) = v0 + α0·Δt, θ pred =θ0. Similarly, since it is uniformly accelerated linear motion, the predicted heading angle θ pred Equal to the current heading angle θ 0。

[0052] 3) Constant Turning Rate Model: This model assumes that both the magnitude of the velocity and the yaw rate remain constant. Let the current state of the obstacle be (x0, y0, θ0, v0, ω0), where ω0 is the yaw rate. The predicted position coordinates at a future preset time Δt are calculated as follows: Scenario 1 (|ω0|>ε, significant turning state): x pred =x0+(v0 / ω0)·[sin(θ0+ω0·Δt)-sin(θ0)], y pred =y0+(v0 / ω0)·[cos(θ0)-cos(θ0+ω0·Δt)], θ pred =θ0+ω0·Δt.

[0053] Case 2 (|ω0|≤ε, approximately a straight line, degenerates into a uniform linear model): x pred =x0+v0·Δt·cos(θ0), y pred =y0+v0·Δt·sin(θ0), θ pred =θ0. Where ε is the preset angular velocity threshold, typically 0.01 rad / s, which can be calibrated according to the sensor noise level and the scene.

[0054] 4) Curve fitting model: Based on historical trajectory points, a cubic polynomial is used to fit the position. The function of coordinate change over time. Let the position coordinates and time series in the historical trajectory sequence be {(t...} i ,xi ,y i Let |i=1..N} be the number of coordinates, x and y are fitted with cubic polynomials with time t as the independent variable: x(t) = a0 + a1·t1 + a2·t2 + a3·t3, y(t) = b0 + b1·t1 + b2·t2 + b3·t3. Here, a0-a3 and b0-b3 are the fitting coefficients, which are obtained by the least squares method. Then, the predicted coordinates at the future time of the preset duration Δt are: x pred =x(t0+Δt), y pred =y(t0+Δt).

[0055] It should be noted that, in addition to the four motion models mentioned above, other motion models can also be used to predict the motion trajectory of obstacles. This application does not impose specific restrictions on this.

[0056] Step S203: Select the optimal motion model from among multiple motion models based on the cumulative cost of each motion model.

[0057] In practical applications, since each motion model has a different degree of deviation in predicting the trajectory of obstacles, it is necessary to judge the degree of deviation of each motion model in predicting the trajectory of obstacles, so as to select the motion model with the smallest degree of deviation in predicting the trajectory of obstacles to accurately predict the trajectory of obstacles.

[0058] In this embodiment, the optimal motion model is selected from multiple motion models based on the calculated cumulative cost of each motion model at the current moment. The optimal motion model is the one with the lowest cumulative cost, where the cumulative cost characterizes the trend change in the degree of deviation between the predicted obstacle trajectory and the actual obstacle trajectory in the recent period.

[0059] In the specific implementation, for each motion model m, the predicted position coordinates corresponding to the current time are taken from the obstacle prediction trajectory generated at the previous time. Simultaneously, the obstacle's position at the current moment is obtained from the perception fusion module. actual position coordinates Then, according to formula (1): To calculate the instantaneous prediction error of each motion model m at the current time t. Instantaneous prediction error The unit of measurement is square meters ( ), used to reflect the motion model The degree of deviation in the prediction of the current trajectory of the obstacle.

[0060] Meanwhile, in order to smooth out single-frame noise interference and highlight the trend changes in recent motion trajectory prediction accuracy, formula (2) can be used: The exponentially weighted moving average method is used to calculate the cumulative cost of each motion model m. An update will be performed. Among other things, For motion model At any moment Updated cumulative cost, initial value It can be set to zero or preset based on prior knowledge (e.g., 0.5 for vehicle obstacles and 0.9 for pedestrian obstacles). The attenuation factor has a value range of 1. . The closer it is to 1, the higher the proportion of historical cumulative costs retained, and the lower the sensitivity to instantaneous prediction errors; The closer λ is to 0, the more it emphasizes the immediate impact of recent prediction errors. In practical applications, a typical recommended value for λ is 0.8, which can achieve a good balance between historical smoothness and instantaneous response speed.

[0061] Understandably, by using the above-mentioned exponentially weighted moving average method to continuously update the cumulative cost of each motion model, it is possible to quickly respond to real changes in the motion state of obstacles, while also having a certain robustness to occasional perceived noise.

[0062] It should be noted that, in addition to using the exponentially weighted moving average method to update the cumulative cost of each motion model, a sliding window average or a Kalman filter residual method can also be used instead. 1) Sliding window averaging: Save the instantaneous prediction error sequence of the most recent W frames, and take its arithmetic mean as the model cost: Formula (3): Where W is the sliding window length, with a recommended value of 5-10 frames. This method does not require maintaining the cumulative state of historical costs, making it simpler to implement, but its response speed to recent abrupt changes in prediction errors is slightly slower than the exponential moving average method.

[0063] 2) Kalman filter residual method: Maintain a Kalman filter for each motion model, and use the normalized innovation squared (NIS) of the filter residuals as a measure of model fit: Formula (4): , where z t Let H be the observed values, H be the observation matrix, and S be the observation value. t The residual covariance matrix is... Let be the prior obstacle state estimate at time t. This method can utilize both prediction and observation information simultaneously, making it more robust to sensor noise, but its computational complexity is slightly higher than the previous two methods.

[0064] Of course, other methods can also be used to update the cumulative cost of each motion model, and this application does not impose specific limitations on this.

[0065] In practical applications, there may be motion models that are completely inconsistent with the current motion state of the obstacle. If the cumulative cost of each motion model is directly compared to select the optimal motion model, the selection process may be inefficient. Therefore, based on the cumulative cost of each motion model, a preliminary screening of multiple motion models is performed in each prediction cycle before selecting the optimal motion model from multiple motion models, in order to improve the selection efficiency and stability.

[0066] In one possible implementation, multiple motion models are screened based on the multidimensional kinematic characteristics of the obstacle to obtain at least one candidate motion model that matches the motion state represented by the multidimensional kinematic characteristics; and the optimal motion model is selected from at least one candidate motion model based on the cumulative cost of each candidate motion model.

[0067] Among them, multidimensional kinematic features are multidimensional motion parameters used to describe the motion state of obstacles. Specifically, in the embodiments of this application, kinematic features of three dimensions, namely curvature features, angular velocity features, and velocity features, are mainly extracted to perform preliminary screening of multiple motion models.

[0068] In one possible implementation, the historical trajectory sequence of the obstacle is obtained, and the multidimensional kinematic features of the obstacle are extracted from the historical trajectory sequence.

[0069] In this embodiment of the application, the historical trajectory sequence of the obstacle is obtained. and corresponding timestamps In the historical trajectory sequence Extract the following multidimensional kinematic features: First, assume the trajectory points The position coordinates are The heading angle is The speed is Then, two adjacent trajectory points can be defined. and driving distance between For: By formula (5): The time interval between two adjacent trajectory points is .

[0070] 1) Curvature characteristics: Through formula (6): Calculate the local curvature sequence of the line connecting adjacent trajectory points. , defined as the ratio of the change in heading angle to the distance traveled. Based on curvature sequences According to formula (7): Calculate the mean curvature Through formula (8): Calculate the rate of change of curvature , defined as the rate of change of the difference between adjacent curvature values ​​over time. Finally, it is expressed by formula (9): Take the mean of the rate of change of curvature As a curvature feature output.

[0071] 2) Angular velocity characteristics: According to formula (10): Calculate the rate of change of heading angle (angular velocity) sequence between adjacent time intervals. Based on angular velocity sequence According to formula (11): Calculate the average angular velocity Ultimately based on angular velocity sequences and mean angular velocity , Calculate the variance of angular velocity As an output of angular velocity characteristics.

[0072] 3) Velocity characteristics: Based on the velocity corresponding to each trajectory point in the historical trajectory sequence. Through formula (13): Calculate the velocity variance As a velocity feature output, The mean of the velocity sequence: .

[0073] It should be noted that, in addition to extracting the multidimensional kinematic features mentioned above, other kinematic features can also be extracted for screening multiple motion models. This application does not impose specific limitations on this.

[0074] In this embodiment of the application, by extracting multidimensional kinematic features from the historical trajectory sequence corresponding to each obstacle, the recent motion state of the obstacle can be more realistically reflected, thereby achieving a more accurate preliminary screening of multiple motion models.

[0075] In this embodiment, after extracting multidimensional kinematic features, multiple motion models are screened based on the multidimensional kinematic features of the obstacle to obtain at least one candidate motion model that matches the motion state represented by the multidimensional kinematic features. The candidate motion model is the motion model that matches the motion state of the obstacle and has the potential to become the optimal motion model.

[0076] In the specific implementation, the rules for filtering various motion models using the extracted multidimensional kinematic features are as follows: 1) If the average angular velocity (Typical value) And the mean rate of change of curvature (Typical value) This indicates that the obstacle is moving in approximately a straight line. In this case, only the uniform velocity linear model and the uniform acceleration linear model are retained for subsequent cumulative cost comparison; the constant turning rate model and the curve fitting model are directly excluded. The average angular velocity threshold is preset based on the actual scenario; This is the preset threshold for the average rate of change of curvature.

[0077] 2) If the average angular velocity (typical value) And angular velocity variance (typical value) This indicates that the obstacle is in a stable turning state. In this case, the priority of the uniform velocity linear model and the uniform acceleration linear model is reduced (by artificially adding a penalty term to their cumulative cost) to highlight the advantages of the constant turning rate model and the curve fitting model. This is the preset angular velocity variance threshold.

[0078] 3) If the velocity variance (typical value) This indicates a drastic change in velocity, at which point the priority of the uniformly accelerated linear model is increased (its cumulative cost is artificially reduced by one reward term). Among these, This is the preset velocity variance threshold.

[0079] It should be noted that, in addition to the above-mentioned screening rules, other screening rules can also be used to screen multiple motion models, and this application embodiment does not impose specific restrictions on this.

[0080] After the initial screening using the above selection rules, at least one candidate motion model will be obtained. Based on the cumulative cost calculated for each motion model at the current moment, the motion model with the minimum cumulative cost is selected as the optimal motion model from the candidate models. And recorded as: If multiple candidate motion models have the same cumulative cost and are all at their minimum, the optimal motion model is selected according to a preset priority. For example: uniform linear model > uniformly accelerated linear model > constant turning rate model > curve fitting model, thus ensuring the determinism of the output.

[0081] In this embodiment, at least one candidate motion model is obtained by first screening multiple motion models using multidimensional kinematic feature representation. Then, the optimal motion model is selected based on the cumulative cost calculated for each motion model at the current moment. This two-level screening method can more realistically and accurately predict the trajectory of obstacles, thus providing a more accurate strategy basis for the vehicle to take collision avoidance actions.

[0082] Step S204: If the current motion model is not the optimal motion model, then the obstacle prediction trajectory of the optimal motion model is used to determine the first final obstacle prediction trajectory.

[0083] In this embodiment of the application, after selecting the optimal motion model, it is necessary to first determine whether the motion model currently used by the obstacle is already the optimal motion model. If it is not the optimal motion model, then the optimal motion model is adopted, and the obstacle prediction trajectory of the optimal motion model is used to determine the first final obstacle prediction trajectory.

[0084] In specific implementation, if the current motion model is =Optimal Motion Model If so, the current motion model remains unchanged; If the current motion model is Optimal Motion Model In this case, it is necessary to consider the cumulative cost of each motion model to determine whether the currently used motion model needs to be changed. Replace with optimal motion model .

[0085] In one possible implementation, if the currently used motion model is not the optimal motion model, and the difference between the cumulative cost of the currently used motion model and the cumulative cost of the optimal motion model is greater than or equal to a preset cumulative cost difference threshold, then the obstacle prediction trajectory of the optimal motion model is used to determine the first final obstacle prediction trajectory.

[0086] In practical applications, to avoid high-frequency oscillation switching due to small perturbations when the cumulative costs of various motion models are close, a hysteresis comparison mechanism is introduced to determine whether the currently used motion model needs to be switched. Switch to optimal motion model .

[0087] Specifically, this can be achieved through formula (14): To determine whether a switch is needed. When the condition of formula (14) is met, it is determined that the currently used motion model is inferior to the optimal motion model, and at this time, the currently used motion model needs to be switched. Replace with optimal motion model The first final obstacle prediction trajectory is determined using the obstacle prediction trajectory of the optimal motion model. The current motion model The cumulative cost at the current time t; For the optimal motion model The cumulative cost at the current time t The preset cumulative cost difference threshold can be set according to the actual scenario; a typical value is [value to be filled in]. If the condition of formula (14) is not met, the current motion model is maintained to avoid unnecessary jitter.

[0088] In this embodiment, a hysteresis comparison mechanism is introduced to determine whether the currently used motion model needs to be changed. Replace with optimal motion model This greatly reduces unnecessary high-frequency oscillation switching caused by minor disturbances.

[0089] In practical applications, to avoid discontinuous jumps in the predicted obstacle trajectories output between adjacent prediction moments during motion model switching, which could lead to false alarms or missed alarms in the collision detection module and thus affect the stability and ride comfort of the autonomous driving system, a smooth transition method is adopted to complete the motion model switching.

[0090] In one possible implementation, the obstacle prediction trajectory of the currently used motion model and the obstacle prediction trajectory of the optimal motion model within a first time interval are fused to obtain the obstacle prediction trajectory within the first time interval; the obstacle motion trajectory predicted by the optimal motion model within a second time interval is used to determine the obstacle prediction trajectory within the second time interval; the obstacle prediction trajectory within the first time interval and the obstacle prediction trajectory within the second time interval are combined to obtain the first final obstacle prediction trajectory. The first time interval is preceding the second time interval.

[0091] In this application example, when model switching is required, in order to achieve a smooth transition between different motion models, a weighted transition period of a preset time length is used as the first time interval, with the moment that triggers the motion model switching as the starting moment. Within the first time interval, the obstacle prediction trajectory of the currently used motion model and the obstacle prediction trajectory of the optimal motion model are fused to obtain the obstacle prediction trajectory within the first time interval.

[0092] In the specific implementation, the time length is set to The weighted transition period of the frame is used as the first time interval. Assume the obstacle prediction trajectory of the current motion model is as follows: The obstacle prediction trajectory of the optimal motion model is ,in, The number of trajectory points predicted for obstacles (determined by the prediction time domain and time step). Each trajectory point Includes position coordinates and heading angle .

[0093] At the same time, a transition frame counter k=0 is set, and in the weighted transition period... frame( ), through formula (15): Calculate the transition coefficients for the current frame. , from It began to grow linearly, until Frame time reached .

[0094] For each index in the obstacle prediction trajectory ( The fused trajectory points are generated using the following weighted interpolation formula and used as the obstacle prediction trajectory within the first time interval. :

[0095] For heading angle When interpolating, attention must be paid to the periodicity of the angle. At this time, the difference in heading angle should first be normalized to The weighting is then applied after the interval is calculated. Among them, ( (j) represents the position coordinates of the fused trajectory point. Let be the heading angle of the fused trajectory point with index j.

[0096] go through After the frame, achieve , This indicates that the motion model has overcompleted. In this case, the obstacle trajectory predicted by the optimal motion model will be used entirely in the second time interval as the obstacle prediction trajectory in the second time interval.

[0097] During the weighted transition period, the system maintains a status flag. This indicates that a motion model switch is in progress. Reset This indicates that the motion model has been over-completed.

[0098] Finally, the obstacle prediction trajectories in the first time interval and the obstacle prediction trajectories in the second time interval are combined to obtain the first final obstacle prediction trajectory.

[0099] In this embodiment, the switching of motion models is completed by setting a weighted transition period, which effectively avoids discontinuous jumps in the obstacle prediction trajectory output between adjacent prediction times when switching motion models, reduces false alarms or missed alarms in the collision detection module, and improves the stability and ride comfort of the autonomous driving system.

[0100] In practical applications, obstacles such as two-wheeled vehicles or pedestrians often exhibit unconventional motion states when traveling at low speeds, such as tilting, turning on the spot, or lateral swerving. The accuracy of traditional rigid body kinematic models (uniform linear motion model / uniformly accelerated linear motion model / constant turning rate model, etc.) in predicting the motion trajectory will significantly decrease. Therefore, for such obstacles, it is necessary to use Gaussian process regression algorithms to determine the predicted obstacle trajectory when necessary.

[0101] In one possible implementation, if the obstacle is a target type obstacle, a Gaussian process regression algorithm is used to determine the second final obstacle prediction trajectory.

[0102] Among them, target type obstacles refer to obstacles such as two-wheeled vehicles or pedestrians that often exhibit unconventional movement states such as tilting to the side, turning on the spot, or lateral swerving when traveling at low speeds.

[0103] In practice, the system first determines whether the obstacle is in an unconventional motion state by detecting whether the slip angle is mismatched, and then decides whether to use the Gaussian process regression algorithm to determine the second final obstacle prediction trajectory.

[0104] Slip angle Used to characterize the velocity vector direction of an obstacle. Heading angle output by the perception fusion module The absolute difference between them is used to reflect whether the above-mentioned unconventional motion state occurs when there is an obstacle. This is expressed by formula (16): Calculate the direction of the obstacle's velocity vector ,in The velocity component of the obstacle in the global coordinate system is provided directly by the perception fusion module or estimated by position difference. This is expressed by formula (17): Calculate the slip angle If the angle difference exceeds It needs to be done through formula (18): Normalize to the modulo operation Interval.

[0105] Preset slip angle mismatch threshold Typical value (about (radians). If This indicates that the obstacle's velocity direction is basically consistent with its orientation, and the obstacle is in a normal motion state. We will continue to use the traditional rigid body kinematics model to predict the obstacle's trajectory.

[0106] like If the obstacle is determined to be in an unconventional motion state such as sideslipping or stationary maneuvering, the traditional rigid body kinematics model fails, and the Gaussian process regression algorithm needs to be used to determine the second final obstacle prediction trajectory.

[0107] Specifically, when a slip angle mismatch is detected, if the currently used motion model is a uniform linear model or a constant turning rate model, the prediction weights of the uniform linear model and the constant turning rate model in the current frame are forcibly reset to zero. That is, these two motion models do not participate in the generation of obstacle prediction trajectories in the current frame, and the obstacle prediction trajectory in the current frame is marked as "to be replaced". If the current motion model is a uniformly accelerated linear model or a curve fitting model, it is retained for later use.

[0108] To compensate for the gaps in the predicted trajectory caused by the failure of the uniform linear model or the constant turning rate model, the Gaussian process regression algorithm is used to generate a brand-new obstacle prediction trajectory as the second final obstacle prediction trajectory.

[0109] First, extract the most recent M trajectory points from the historical trajectory sequence corresponding to the obstacle (e.g., Using the tangent direction in the historical trajectory sequence of the obstacle at the current moment as the reference line, a Frenet (Frenet Coordinate System) coordinate system is constructed. The specific method for constructing the coordinate system is as follows: take the last two trajectory points in the historical trajectory sequence. and Calculate the direction of the line connecting them as the longitudinal reference direction. :Formula (19): Using the current coordinates of the obstacle With the origin as the point, Direction is Axis (longitudinal arc length), perpendicular to The direction to the left is Establish a local right-handed coordinate system by offsetting the axis (horizontal offset).

[0110] For each historical trajectory point in the historical trajectory sequence corresponding to the obstacle ( ), calculate its lateral offset in the Frenet coordinate system mentioned above. This value is the trajectory point. The vertical distance to the reference line represents the lateral deviation of the obstacle relative to its current direction of motion at the corresponding moment, and its sign is determined by whether the trajectory point is to the left or right of the reference line. Simultaneously, with the current moment as the time zero point, the normalized timestamp corresponding to that trajectory point is recorded. (Historical moments are negative). Based on horizontal offset and normalized timestamps According to formula (20): To construct the training dataset.

[0111] Meanwhile, in order to make the historical trajectory sequence with similar slip angle characteristics have a greater impact on the obstacle prediction trajectory of the current frame, the smooth motion trend and lateral maneuvering abrupt change are captured simultaneously, based on formula (21): To design a combined kernel function Among them, through formula (22): To calculate the squared exponent kernel To capture smooth motion trends, signal variance σ f 2 Used to control the output amplitude of the Gaussian process, with a recommended range of 1.0~5.0; length scale l f This value controls the decay rate of the temporal correlation of historical trajectory points. A larger value indicates a longer-lasting influence of the historical trajectory sequence data on the obstacle prediction trajectory. The recommended range is 0.5~3.0s. This is achieved through formula (23): To calculate the lateral maneuvering core With the slip angle As an additional dimension for similarity measurement, among which , Inputs and The slip angle calculated at the corresponding time. Signal variance σ lat 2 The recommended range for controlling the amplitude of slip angle compensation is 0.1 to 1.0; the slip angle length scale is l. lat For controlling the decay rate of slip angle similarity, a recommended range is 5°~20° (0.087~0.349 rad); time scale l t For controlling the duration of compensation, a range of 0.5 to 2.0 seconds is recommended. In practical applications, typical values ​​for the above parameters are suggested as follows: The above parameters can be optimized and calibrated offline by maximizing the marginal likelihood.

[0112] Finally, the standard Gaussian process regression algorithm formula (24) is used: Calculate the future Time points Predicted mean vector .in, To train the relationship between trajectory points in the dataset kernel matrix, elements ; To compare the trajectory points in the training dataset with the predicted trajectory points kernel matrix, elements ; The lateral offset of each trajectory point in the training dataset The resulting lateral offset vector; To observe the noise variance, the recommended range for the noise level of the corresponding perception fusion module is 0.001 to 0.1, with a typical value being [value missing]. =0.01, Offline optimization and calibration can also be performed by maximizing the marginal likelihood; It is an identity matrix. The predicted mean is taken. As the corrected lateral offset sequence, combined with the longitudinal arc length of each trajectory point The position coordinates are converted back to the global coordinate system, forming a complete Gaussian process obstacle prediction trajectory. Finally, the Gaussian process obstacle prediction trajectory replaces the original obstacle prediction trajectory generated by the uniform linear model or constant turning rate model at the current moment and is determined as the second final obstacle prediction trajectory; if the current motion model is a uniformly accelerated linear model or a curve fitting model, its obstacle prediction trajectory is retained unchanged and not merged with the Gaussian process obstacle prediction trajectory.

[0113] It should be noted that, in addition to using the Gaussian process regression algorithm to predict the trajectory of target type obstacles, other methods can also be used for trajectory prediction, and this application does not impose specific restrictions on this.

[0114] In this embodiment of the application, for obstacles such as two-wheeled vehicles or pedestrians that often exhibit unconventional motion states such as tilting, turning on the spot, and lateral avoidance when traveling at low speeds, the Gaussian process regression algorithm is used to determine the second final obstacle prediction trajectory, which effectively avoids the problem of low obstacle motion trajectory prediction accuracy caused by using traditional rigid body kinematics models.

[0115] In practical applications, to ensure that the predicted obstacle trajectory is reachable from the obstacle in the lateral acceleration dimension and to avoid outputting obstacle prediction trajectories that violate dynamics, physical feasibility constraints need to be applied to both the first obstacle prediction trajectory to be optimized (determined by the optimal motion model) and the second obstacle prediction trajectory to be optimized (determined by the Gaussian process regression algorithm) to correct the obstacle prediction trajectories.

[0116] In one possible implementation, the obstacle prediction trajectory of the optimal motion model is used to determine the first obstacle prediction trajectory to be optimized; the first obstacle prediction trajectory to be optimized is subject to physical feasibility constraints to obtain the first final obstacle prediction trajectory; the Gaussian process regression algorithm is used to determine the second obstacle prediction trajectory to be optimized; the second obstacle prediction trajectory to be optimized is subject to physical feasibility constraints to obtain the second final obstacle prediction trajectory.

[0117] In the embodiments of this application, the physical feasibility constraint means that for any trajectory point on the predicted trajectory of the obstacle, its lateral acceleration must not exceed the physical limit of that type of obstacle.

[0118] In the specific implementation, the sequence of predicted obstacle trajectory points to be optimized is... Projecting onto the Frenet coordinate system yields the lateral offset sequence. and the corresponding longitudinal arc length sequence ,in , Index of each trajectory point, To predict the time step (e.g.) Second).

[0119] Horizontal offset Considered as longitudinal arc length function The relationship is fitted using a cubic polynomial: Formula (25): , The coefficients are those of a cubic polynomial; the objective function for the original least squares fitting is: ,in As the weight, recent trajectory points have a higher weight.

[0120] Therefore, the physical constraint can be set such that for any point on the predicted trajectory of the obstacle, its lateral acceleration must not exceed the physical limit of lateral acceleration for that type of obstacle. .

[0121] Refer to Table 1 for the physical limits of lateral acceleration for different types of obstacles provided in the embodiments of this application. Example table: Table 1:

[0122] As shown in Table 1, the physical limits of lateral acceleration for vehicle-type obstacles. It is 3.0. The physical limit of lateral acceleration for two-wheeled vehicle type obstacles. It is 2.5 The physical limit of lateral acceleration for pedestrian-type obstacles. It is 1.5 .

[0123] In low-speed scenarios, the lateral acceleration can be approximated as: Equation (26): Then the physical constraints can be expressed by formula (27): This transforms into a restriction on the second derivative of the polynomial. For a cubic polynomial, ,for The linear function. Therefore, the above physical constraints are equivalent to the following holding at the endpoints of the interval: Equation (28): Among them, the predicted total longitudinal arc length , This represents the total forecast duration.

[0124] Solving the above least squares problem with linear inequality constraints involves handling the following cases based on the solution results: Case 1 (solution exists and constraints are satisfied): Output the polynomial coefficients after physical constraint correction to generate a new obstacle prediction trajectory, which replaces the original obstacle prediction trajectory and serves as the first or second final obstacle prediction trajectory.

[0125] Scenario 2 (No feasible solution): This indicates that the maneuver cannot be described by a single polynomial within the physical limits. In this case, polynomial correction is abandoned, and the model degenerates into a combination of circular arc and straight line. The method for generating this combined model is as follows: ① Radius of curvature estimation: Take the last two trajectory points in the historical trajectory buffer. and Calculate the radius of curvature of the line connecting them. If the distance between two points is too small or the change in heading angle is approximately zero, then it is determined to be linear motion. .

[0126] ② Circular arc generation: Starting from the current position of the obstacle, along the current heading angle, with a constant radius of curvature... Generate a circular arc trajectory. The central angle of the arc... Extrapolation distance from the preset arc Decision: Formula (29): ,in Can be taken as , The preset duration of the arc segment (e.g.) (seconds). Position coordinates of each point on the arc segment. and heading angle Calculated using the parametric equation of the circular arc.

[0127] ③ Line segment generation: After the circular arc segment ends, a uniformly accelerated straight extension segment is generated, using the heading angle at the end of the circular arc as the direction, until the total prediction time reaches [the desired value]. Length of a straight line segment: Formula (30): .

[0128] ④ Trajectory splicing: Splice the trajectory points of the circular arc segment and the trajectory points of the straight line segment according to the time sequence. This constitutes a complete obstacle prediction trajectory, which serves as either the first final obstacle prediction trajectory or the second final obstacle prediction trajectory.

[0129] It should be noted that, in addition to using the physical limit of lateral acceleration not exceeding that of the obstacle as the physical constraint, the physical constraint can also be replaced by the following friction circle constraint or vehicle dynamics envelope constraint: (1) Friction circle constraint: The longitudinal acceleration and the transverse acceleration are constrained together within the friction circle, i.e.: Formula (31) ,in lat For lateral acceleration, lon For longitudinal acceleration, this method is suitable for scenarios where acceleration / deceleration and steering are coupled simultaneously, providing more comprehensive constraints.

[0130] (2) Vehicle dynamics envelope constraint: The steady-state response constraint of the two-degree-of-freedom single-track vehicle model is introduced, and the yaw rate is constrained within the range of |ω|≤μg / v, where μ is the road adhesion coefficient, g is the gravitational acceleration (taken as 9.8m / s2), and v is the current vehicle speed. This method can directly limit the distortion of obstacle prediction trajectory caused by excessive angular velocity and is suitable for the correction of obstacle prediction trajectory in medium and high speed scenarios.

[0131] In this embodiment of the application, by applying physical feasibility constraints to all the obstacle prediction trajectories to be optimized, it is effectively ensured that the obstacle prediction trajectory is reachable from the obstacle in the lateral acceleration dimension, thus avoiding the output of obstacle prediction trajectories that violate dynamics.

[0132] In practical applications, when the motion model is in a weighted transition period switching motion model, and when the cumulative cost of the motion model for obstacles is constantly increasing (i.e., the historical weighted moving average error is rapidly diverging), the accuracy of obstacle prediction trajectory will decrease. Therefore, before officially outputting the final obstacle prediction trajectory, it is necessary to quantitatively evaluate the confidence level of the initial obstacle prediction trajectory at the current moment, and accordingly shorten the time length of the output obstacle prediction trajectory.

[0133] In one possible implementation, the obstacle prediction trajectory is determined by using the obstacle prediction trajectory of the optimal motion model; the effective duration of the initial obstacle prediction trajectory is determined based on the confidence level of the initial obstacle prediction trajectory; and the obstacle prediction trajectory within the effective duration range is extracted from the initial obstacle prediction trajectory to obtain the first final obstacle prediction trajectory.

[0134] The initial obstacle prediction trajectory refers to the obstacle prediction trajectory that will be output after being corrected by physical constraints.

[0135] In this application embodiment, confidence level refers to the influencing factors affecting the accuracy of obstacle prediction trajectory, such as the transition coefficient during the weighted transition period when switching motion models. and the divergence rate of the weighted moving average error. The effective duration of the initial obstacle prediction trajectory refers to the duration of the initial obstacle prediction trajectory when the accuracy of the obstacle prediction trajectory is high.

[0136] In the specific implementation, the transition coefficients within the weighted transition period during motion model switching are first extracted. value range When the transition is complete or no switchover occurs, . The closer This indicates that the closer the weights of the old and new motion models are, the higher the uncertainty of the prediction results. Secondly, the divergence rate of the weighted moving average error is extracted. divergence rate Defined as the cumulative cost of the current motion model at the current moment. The changing trend.

[0137] Save recent Frame (e.g.) The cumulative cost sequence , For each frame, the index is where A linear fit is performed on the cumulative cost sequence using formula (32): .in, The index mean, This represents the cumulative cost average. Since we only consider the uncertainty caused by the increase in weighted moving average error (a decrease in weighted moving average error indicates improved prediction), the divergence rate is taken as a non-negative value. like This indicates that the weighted moving average error has not diverged or is converging, and does not trigger the obstacle prediction trajectory truncation penalty.

[0138] Define transition instability , used to measure whether the current period is within the weighted transition period for motion model switching and fusion, is a "hesitation period": Formula (33): .

[0139] when hour, This indicates that the old and new motion models each account for half, making it the most unstable; when or hour, This indicates that a single motion model is used, which is the most stable. A threshold for the unstable interval is set. (correspond ).like It is determined that the current state is in a high-uncertainty transition state.

[0140] Based on transition instability and the divergence rate of the weighted moving average error Two factors affecting the accuracy of obstacle prediction trajectory are defined by formula (34): Calculate the effective duration of the initial obstacle prediction trajectory ,in The base duration for predicting the trajectory of the initial obstacle, typical value Second; This is the transition penalty coefficient, which controls the attenuation intensity during the transition period; typical value. ; This is the weighted moving average error divergence penalty coefficient, which controls the attenuation intensity as the error increases; a typical value is... ; Only when Transitional penalties only occur at certain times, and The smaller the punishment, the heavier it is.

[0141] Meanwhile, to prevent excessive truncation of the duration from rendering obstacle prediction trajectories unavailable in downstream modules, a lower limit is set for the effective duration. (like (seconds). Therefore, the effective duration At the same time, ensure that at least [amount] is retained. Predicted trajectory points of obstacles (e.g.) (Number of obstacles) to meet the requirement of minimum obstacle prediction trajectory length for collision detection.

[0142] After determining the effective duration, the obstacle prediction trajectory within the effective duration is extracted from the initial obstacle prediction trajectory determined by the optimal motion model. Specifically, let the time series of the initial obstacle prediction trajectory be... ,in , To predict the time step (e.g.) Second),( ) is the position coordinate of each trajectory point. For the first The heading angle of each trajectory point at the corresponding timestamp.

[0143] The obstacle prediction trajectory truncation operation is as follows: traverse all trajectory points in the initial obstacle prediction trajectory determined by the optimal motion model, and retain those that satisfy the following conditions. All trajectory points; if the number of trajectory points to retain is less than Then the previous one will be forcibly retained. Each point, and accordingly Revised to The truncated obstacle prediction trajectory sequence is used as the first final obstacle prediction trajectory.

[0144] Meanwhile, in order for downstream modules to differentiate between obstacle prediction trajectories with different confidence levels, an enumerated confidence level flag is output. Similar to the confidence levels of the predicted trajectories for differentiating obstacles, the specific classification criteria are as follows: 1) HIGH: The prediction results are reliable and have a high confidence level. 2) MEDIUM: The prediction results have a certain degree of uncertainty, with a medium confidence level. 3) LOW: The prediction results have high uncertainty and low confidence level.

[0145] The downstream collision risk assessment module can dynamically adjust the safety boundary based on this confidence level. For example, for obstacle prediction trajectories with a LOW confidence level, its equivalent expansion radius is increased. Alternatively, the collision time threshold can be lowered to compensate for the risks associated with inaccurate predictions.

[0146] Similarly, for the initial obstacle prediction trajectory generated by using the Gaussian process regression algorithm to predict the motion trajectory of the target type obstacle, the uncertainty assessment of the prediction result of the modified prediction trajectory is also performed. The obstacle prediction trajectory truncation and confidence level classification operations are performed in the same way, and finally the second final obstacle prediction trajectory is obtained.

[0147] In this embodiment, by evaluating the uncertainty of the generated initial obstacle prediction trajectory, truncating the initial obstacle prediction trajectory by calculating the effective duration, and calculating the confidence level of the obstacle prediction trajectory, the long-term output of erroneous obstacle movement trajectories is effectively avoided. This allows collision avoidance measures to be taken in advance based on the confidence level, greatly reducing the probability of collision and improving driving safety.

[0148] Finally, the final obstacle prediction trajectory, after correction for physical constraints and evaluation of the uncertainty of the initial obstacle prediction trajectory, is output. The output information mainly includes: obstacle identification, the final obstacle prediction trajectory point sequence (each trajectory point includes a timestamp, position coordinates, and heading angle), the type of motion model used for prediction, and a confidence level flag. The downstream collision risk assessment and decision planning module will execute subsequent safety logic based on this output information.

[0149] In this embodiment, at the current moment, current perception data of the obstacle, including its location and motion state information, is acquired. Based on the current perception data, multiple motion models are used to obtain predicted trajectories of the obstacle within a preset future time period. Based on the cumulative cost of each motion model, the optimal motion model is selected from among the multiple motion models; the optimal motion model is the one with the lowest cumulative cost. If the currently used motion model is not the optimal motion model, the predicted obstacle trajectory of the optimal motion model is used to determine the first final predicted obstacle trajectory. This improves the reliability of obstacle trajectory prediction, provides the vehicle with accurate and timely safety decision-making basis, and enables accurate obstacle avoidance operations to be taken in advance, reducing the risk of vehicle collisions.

[0150] Corresponding to the above embodiments, this application also provides a vehicle.

[0151] See Figure 3 This is a structural schematic diagram of a vehicle provided in an embodiment of this application. Figure 3 As shown, vehicle 300 includes controller 301, which is configured to perform some or all of the steps in the above method embodiments.

[0152] Corresponding to the above embodiments, this application also provides a computer-readable storage medium, wherein the computer-readable storage medium may store a program, wherein when the program runs, it can control the device where the computer-readable storage medium is located to execute some or all of the steps in the above method embodiments. In specific implementation, the computer-readable storage medium may be a magnetic disk, an optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0153] Corresponding to the above embodiments, this application also provides a computer program product containing executable instructions that, when executed on a computer, cause the computer to perform some or all of the steps in the above method embodiments.

[0154] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0155] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0156] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0157] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0158] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for predicting obstacle trajectories, characterized in that, include: At the current moment, acquire the current perception data of the obstacle, wherein the perception data includes location information and motion state information; Based on the current perception data of the obstacle, multiple motion models are used to predict the motion trajectory of the obstacle within a preset time period in the future, and the obstacle prediction trajectory of each motion model is obtained. Based on the cumulative cost of each of the aforementioned motion models, the optimal motion model is selected from among the various motion models. The optimal motion model is the motion model with the lowest cumulative cost, whereby the cumulative cost is used to characterize the trend change of the degree of deviation between the predicted trajectory of the obstacle and the actual trajectory of the obstacle in the near future. If the currently used motion model is not the optimal motion model, then the obstacle prediction trajectory of the optimal motion model is used to determine the first final obstacle prediction trajectory.

2. The method according to claim 1, characterized in that, The step of determining the first final obstacle prediction trajectory using the obstacle prediction trajectory of the optimal motion model includes: The obstacle prediction trajectory of the currently used motion model and the obstacle prediction trajectory of the optimal motion model are fused within the first time interval to obtain the obstacle prediction trajectory within the first time interval. Using the obstacle motion trajectory predicted by the optimal motion model within the second time interval, the predicted obstacle trajectory within the second time interval is determined; The obstacle prediction trajectories in the first time interval and the obstacle prediction trajectories in the second time interval are combined to obtain the first final obstacle prediction trajectory; The first time interval is located before the second time interval.

3. The method according to claim 1, characterized in that, The step of determining the first final obstacle prediction trajectory using the obstacle prediction trajectory of the optimal motion model includes: The initial obstacle prediction trajectory is determined using the obstacle prediction trajectory of the optimal motion model; The effective duration of the initial obstacle prediction trajectory is determined based on the confidence level of the initial obstacle prediction trajectory. The first final obstacle prediction trajectory is obtained by intercepting the obstacle prediction trajectory within the effective time range from the initial obstacle prediction trajectory.

4. The method according to claim 1, characterized in that, The step of selecting the optimal motion model from among the multiple motion models based on the cumulative cost of each of the motion models includes: Based on the multidimensional kinematic characteristics of the obstacle, multiple motion models are screened to obtain at least one candidate motion model that matches the motion state represented by the multidimensional kinematic characteristics. Based on the cumulative cost of each of the candidate motion models, the optimal motion model is selected from at least one of the candidate motion models.

5. The method according to claim 4, characterized in that, Before filtering the various motion models based on the multidimensional kinematic characteristics of the obstacle, the method further includes: Obtain the historical trajectory sequence of the obstacle, the historical trajectory sequence including multiple historical perception data of the obstacle; The multidimensional kinematic features of the obstacle are extracted from its historical trajectory sequence.

6. The method according to claim 1, characterized in that, If the currently used motion model is not the optimal motion model, then the obstacle prediction trajectory of the optimal motion model is used to determine the first final obstacle prediction trajectory, including: If the currently used motion model is not the optimal motion model, and the difference between the cumulative cost of the currently used motion model and the cumulative cost of the optimal motion model is greater than or equal to a preset cumulative cost difference threshold, then the obstacle prediction trajectory of the optimal motion model is used to determine the first final obstacle prediction trajectory.

7. The method according to claim 1, characterized in that, Also includes: If the obstacle is a target type obstacle, the Gaussian process regression algorithm is used to determine the second final obstacle prediction trajectory.

8. The method according to claim 7, characterized in that, The step of using Gaussian process regression algorithm to determine the second final obstacle prediction trajectory includes: The Gaussian process regression algorithm is used to determine the second obstacle prediction trajectory to be optimized; physical feasibility constraints are applied to the second obstacle prediction trajectory to obtain the second final obstacle prediction trajectory.

9. The method according to claim 1, characterized in that, The step of determining the first final obstacle prediction trajectory using the obstacle prediction trajectory of the optimal motion model includes: The first obstacle prediction trajectory to be optimized is determined by using the obstacle prediction trajectory of the optimal motion model; physical feasibility constraints are applied to the first obstacle prediction trajectory to be optimized to obtain the first final obstacle prediction trajectory.

10. The method according to claim 1, characterized in that, Also includes: Determine whether the identification mark of the obstacle is an identification mark that has been used by other obstacles; If the identifier of the obstacle is one that has been used by other obstacles, then the historical data corresponding to that identifier is deleted.

11. A vehicle, characterized in that, include: A controller configured to perform the method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1 to 10.

13. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1 to 10.

Images