Automatic driving vehicle track prediction and uncertainty evaluation method
By combining the Singer maneuvering target state-space model and EM adaptive noise parameters, the problem of vehicle trajectory prediction accuracy and robustness in complex urban road scenarios is solved, achieving high-precision and stable trajectory prediction and providing uncertainty assessment support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN TECH UNIV
- Filing Date
- 2025-12-22
- Publication Date
- 2026-04-17
AI Technical Summary
Existing vehicle trajectory prediction technologies lack accuracy and robustness in complex urban road scenarios. Traditional methods rely on experience to set noise parameters, while deep learning methods are not robust enough to abnormal operating conditions and data gaps.
The Singer maneuvering target state-space model is combined with EM adaptive noise parameters. The noise parameters are adaptively estimated using the expectation-maximization algorithm. Combined with Kalman filtering and smoothing operations, vehicle trajectory prediction and uncertainty assessment are performed.
It improves the accuracy and robustness of trajectory prediction in complex urban road scenarios, reduces interference from abnormal data, achieves adaptive noise parameters, enhances stability under different working conditions and the stability of the filtering process under extreme working conditions, and provides uncertainty quantification support.
Smart Images
Figure CN121880847A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of vehicle trajectory prediction technology, and in particular to a method for predicting the trajectory and assessing the uncertainty of autonomous vehicles. Background Technology
[0002] With the development of artificial intelligence, sensor technology, and vehicle-road cooperative technology, the testing and application of autonomous vehicles in urban road environments are becoming increasingly widespread. Autonomous driving systems typically consist of modules such as environmental perception, localization and mapping, behavior prediction, decision planning, and execution control. Among these, accurate prediction of the surrounding traffic participants and the future trajectory of the autonomous vehicle is one of the fundamental tasks to ensure safe driving and smooth traffic flow.
[0003] Existing trajectory prediction methods can be broadly categorized into two types: one is based on physical models and state-space modeling, represented by Kalman filtering, such as the classic constant velocity / constant acceleration model, Extended Kalman Filter (EKF), and the Singer maneuvering target state-space model suitable for maneuvering targets; the other is based on data-driven deep learning methods, such as Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), and Transformer networks. Both types of methods have limitations under complex urban road conditions. On the one hand, traditional Kalman filtering typically uses fixed process noise covariance Q and observation noise covariance R. These noise parameters often rely on empirical settings or simple parameter tuning, making it difficult to accurately reflect changes in vehicle maneuverability and sensor noise characteristics under different operating conditions, easily leading to decreased prediction accuracy or distorted uncertainty estimation. On the other hand, while deep learning methods have achieved high prediction accuracy on certain datasets, they heavily rely on large-scale labeled data and complex network training processes, resulting in poor model interpretability and insufficient robustness to abnormal operating conditions and missing data.
[0004] In summary, the accuracy and robustness of existing vehicle trajectory prediction technologies for predicting the trajectories of autonomous vehicles in complex urban road scenarios need to be improved. Summary of the Invention
[0005] This application provides a method for predicting the trajectory and assessing the uncertainty of autonomous vehicles, which addresses the problem that the accuracy and robustness of existing vehicle trajectory prediction technologies in complex urban road scenarios need to be improved.
[0006] On the one hand, this application provides a method for predicting the trajectory and assessing the uncertainty of autonomous vehicles, including the following steps: Step 1: Obtain vehicle trajectory observation data in urban road scenarios to form a trajectory observation sequence.
[0007] Step 2: Based on the trajectory observation sequence, construct the Singer maneuvering target state space model with the vehicle's motion characteristics as the objective.
[0008] Step 3: Based on the Singer maneuvering target state-space model, after initializing the noise parameters, the expectation-maximization algorithm is introduced to adaptively estimate the noise parameters, and the EM adaptive noise parameters are obtained by convergence.
[0009] Step 4: Perform Kalman filtering and smoothing on the Singer maneuvering target state space model that integrates EM adaptive noise parameters, and output the vehicle's predicted trajectory.
[0010] Step 5: Perform an uncertainty assessment on the predicted vehicle trajectory to obtain the uncertainty assessment result.
[0011] In one possible implementation, in step one, after the trajectory observation sequence is formed, the trajectory observation sequence is preprocessed, and the preprocessing includes smoothing and interpolation.
[0012] In one possible implementation, step three, which involves introducing an expectation-maximization algorithm for adaptive noise parameter estimation and converging to obtain the EM adaptive noise parameters, includes: The expectation estimate of the state is calculated in the E-step of the expectation maximization algorithm, the model noise parameters are updated in the M-step, and the iteration is repeated until convergence to obtain the EM adaptive noise parameters.
[0013] In one possible implementation, the model noise parameters updated in the M-step include the update process noise covariance Q and the observation noise covariance R.
[0014] In one possible implementation, in step four, when there are missing or anomalous observations, the measurement update step of the Kalman filter is skipped for unavailable observation times, and only the time prediction step is performed.
[0015] In one possible implementation, the absence or anomaly of the observation is identified using an availability determination.
[0016] The rules for determining availability include: observations being empty / NaN, sampling time intervals exceeding a threshold, and innovation residuals exceeding a threshold.
[0017] In one possible implementation, in step five, the evaluation metrics for uncertainty assessment include: root mean square error, mean absolute error, and Q95 quantile error.
[0018] The method for predicting the trajectory and assessing the uncertainty of an autonomous vehicle in this application has the following advantages: By combining the Singer maneuvering target state-space model with EM adaptive noise parameters, the accuracy and robustness of trajectory prediction for autonomous vehicles in complex urban road scenarios are improved.
[0019] By smoothing and interpolating the trajectory observation sequence, abnormal fluctuations in the trajectory observation sequence are reduced, data gaps are filled, high-quality input is provided for subsequent model estimation, and interference from abnormal data is reduced.
[0020] By introducing the expectation-maximization algorithm for adaptive noise parameter estimation, the EM adaptive noise parameter is obtained by convergence, thus realizing noise parameter adaptation, reducing reliance on human experience, and improving stability under different working conditions.
[0021] By updating the process noise covariance Q and the observation noise covariance R, the vehicle's maneuvering characteristics and sensor noise changes are accurately matched, further optimizing the filtering accuracy.
[0022] By skipping the measurement update step of the Kalman filter at unavailable observation times and only performing the time prediction step, estimation bias caused by erroneous observations is avoided, ensuring the stability and continuity of the filtering process when observations are missing or abnormal, and improving robustness under extreme conditions.
[0023] By employing availability determination to identify whether there are missing or abnormal observations, a reliable basis is provided for filtering branch processing, ensuring the effective implementation of the robust mechanism.
[0024] Uncertainty is assessed by using root mean square error, mean absolute error, and Q95 quantile error, thereby quantifying uncertainty and providing data support for autonomous driving safety decisions. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 A flowchart illustrating an autonomous vehicle trajectory prediction and uncertainty assessment method provided in this application embodiment; Figure 2 This application provides a randomly selected real trajectory map for an embodiment of the present application. Figure 3 A visual overlay comparison of the vehicle prediction trajectories of an autonomous vehicle trajectory prediction and uncertainty assessment method provided in this application embodiment and six existing trajectory prediction methods. Figure 4Provided for the embodiments of this application Figure 3 The results of the local segment comparison are shown in the figure. Detailed Implementation The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0027] like Figure 1 As shown in the figure, this application provides a method for predicting the trajectory and assessing the uncertainty of an autonomous vehicle, including the following steps: Step 1: Obtain vehicle trajectory observation data in urban road scenarios to form a trajectory observation sequence.
[0028] Step 2: Based on the trajectory observation sequence, construct the Singer maneuvering target state space model with the vehicle's motion characteristics as the objective.
[0029] Step 3: Based on the Singer maneuvering target state-space model, after initializing the noise parameters, the expectation-maximization algorithm is introduced to adaptively estimate the noise parameters, and the EM adaptive noise parameters are obtained by convergence.
[0030] Step 4: Perform Kalman filtering and smoothing on the Singer maneuvering target state space model that integrates EM adaptive noise parameters, and output the vehicle's predicted trajectory.
[0031] Step 5: Perform an uncertainty assessment on the predicted vehicle trajectory to obtain the uncertainty assessment result.
[0032] For example, in step one, after the trajectory observation sequence is formed, the trajectory observation sequence is preprocessed, and the preprocessing includes smoothing and interpolation.
[0033] Specifically, smoothing and interpolation are used to reduce the impact of abnormal fluctuations on subsequent model estimation and to construct the trajectory observation sequence required for filtering.
[0034] Specifically, in this embodiment, in step two, constructing the Singer maneuvering target state space model can describe the vehicle's maneuvering behaviors such as turning, acceleration, and deceleration on urban roads, providing dynamic constraints for subsequent filtering estimation.
[0035] Specifically, the composition and construction process of the Singer maneuvering target state space model are as follows: The Singer maneuvering target state-space model is used to characterize the maneuvering behaviors of vehicles in urban roads, such as turning, acceleration, and deceleration. Its core idea is to model vehicle acceleration as a time-dependent stochastic process, thus describing the dynamic changes of maneuvering segments better than a constant-velocity (CV) model. This embodiment uses two-dimensional planar motion modeling, where the state vector consists of position, velocity, and acceleration. .
[0036] Among them, (p x ,p y ) represents the vehicle's position components in a planar coordinate system (converted from latitude and longitude to local planar coordinates, or approximated linearly over a small area), (v x ,v y ) represents the velocity component, (a x ,a y () represents the acceleration component.
[0037] Singer's maneuvering target state-space model treats acceleration as a first-order Gaussian-Markov process (correlated acceleration), and its continuous form can be expressed as: .
[0038] Where τ is the acceleration-related time constant used to describe maneuver sustainability; w a (t) represents zero-mean white noise, used to describe the intensity of unmodeled maneuvering disturbances.
[0039] Let the sampling period be Δt, and let α = exp(−Δt / τ). For a single direction (x or y), the discrete state transition form can be obtained: .
[0040] One commonly used discretization form is: .
[0041] The two-dimensional global state transition matrix has a block diagonal structure: .
[0042] The observations are vehicle position observations (e.g., transformed plane coordinates of GPS position), and the observation equation is: .
[0043] in, .
[0044] Process noise satisfies w k ~N(0,Q), the observation noise satisfies v k~N(0,R). Q and R represent the process noise covariance and observation noise covariance, respectively. The above F, H, Q, and R together constitute the Singer maneuvering target state-space model, providing dynamic constraints for subsequent filtering estimation.
[0045] For example, in step three, the introduction of the expectation-maximization algorithm for adaptive noise parameter estimation, converging to obtain the EM adaptive noise parameters, includes: The expectation estimate of the state is calculated in the E-step of the expectation maximization algorithm, the model noise parameters are updated in the M-step, and the iteration is repeated until convergence to obtain the EM adaptive noise parameters.
[0046] For example, the model noise parameters updated in the M-step include the update process noise covariance Q and the observation noise covariance R.
[0047] Specifically, in this embodiment, the introduction of the Expectation Maximization (EM) algorithm can reduce the dependence of noise parameters on human experience settings and improve stability under different operating conditions.
[0048] Specifically, the implementation process of the E-step and M-step of the expectation maximization algorithm is as follows: To reduce the reliance on human experience in setting noise parameters, this embodiment introduces the Expectation-Maximization (EM) algorithm to perform data-driven estimation of the noise covariance Q and R based on a given state-space model (F,H). Taking a linear Gaussian model as an example: , .
[0049] w k ~N(0,Q),v k ~N(0,R), the parameters to be estimated are θ={Q,R}, and the observation sequence is Z={z0,…,z N−1 The EM algorithm operates on the current parameter θ. (j) The following two steps are performed iteratively: In the j-th iteration, the current parameter θ... (j) ={Q (j) ,R (j) The EM iteration consists of the following two steps: E-step: Calculation of the posterior expectation of the state: Fixed Q (j) ,R (j) The forward Kalman filter and backward RTS smoother are applied to the entire observation sequence to obtain the smoothed state mean at each time step. k|N Cooperative variance P k|N and the cross covariance P between adjacent time points k,k−1|NBased on this, the second-order expectations of the process residual and the observation residual are calculated: , .
[0050] M-step: Noise covariance update: By maximizing the expected value of the log-likelihood of the complete data, the noise covariance is updated as follows: , .
[0051] The iteration continues until the changes in Q and R are less than the threshold or the log-likelihood gain is less than the threshold. Convergence is then considered achieved, and the final estimated Q* and R* are output.
[0052] For example, in step four, when there are missing or abnormal observations, the measurement update step of the Kalman filter is skipped for unavailable observation times, and only the time prediction step is performed.
[0053] For example, the absence or anomalies in the observations are identified using an availability determination.
[0054] The rules for determining availability include: observations being empty / NaN, sampling time intervals exceeding a threshold, and innovation residuals exceeding a threshold.
[0055] Specifically, in this embodiment, by skipping the measurement update step of the Kalman filter for unavailable observation times and only performing the time prediction step, the filter is kept stable and estimation offsets caused by erroneous observations are avoided.
[0056] Specifically, in this embodiment, Kalman filtering (abbreviated as Singer-KF(EM) filtering) is performed on the Singer maneuvering target state space model that incorporates EM adaptive noise parameters, and a smoothing operation is performed. The specific process is as follows: Specifically, the implementation process of Kalman filtering for the Singer maneuvering target state-space model with EM adaptive noise parameters (fusing EM's Q and R) is as follows: After obtaining the convergent noise parameters Q*, R*, they are fused with the Singer state-space model (F, H), and Singer-KF(EM) filtering and smoothing are performed. Let the initial state and covariance be... 0|0 ,P 0|0 For k=1,…,N, forward Kalman filtering and RTS smoothing are performed sequentially.
[0057] Forward Kalman filtering includes prediction (time update) and observation update.
[0058] Forecast (time updated): , .
[0059] Observation Update: , , , .
[0060] When observations are missing or anomalous, skip the observation update and retain only the prediction results: k|k= k|k-1, P k|k =P k|k-1 .
[0061] RTS smoothing: To obtain smoother trajectory estimation results, backward smoothing is performed on the filtered output. This is done according to k=N-1,…,0: , , .
[0062] Smoothed k|N Given the final trajectory estimate, P k|N It can be used for the calculation and output of uncertainty assessment indicators.
[0063] For example, in step five, the evaluation indicators for uncertainty assessment include: root mean square error, mean absolute error, and Q95 quantile error.
[0064] Specifically, in one possible embodiment, to verify the effectiveness and advantages of the autonomous vehicle trajectory prediction and uncertainty assessment method of this application in real urban scenarios, a trajectory prediction comparison experiment is designed. The experiment selects a vehicle trajectory with typical urban road maneuvering characteristics as the test object. This trajectory covers long straight-line travel, multiple intersection turns, and sections with local speed changes, representing the common "straight-turn-re-straight-re-turn" hybrid motion pattern of autonomous vehicles on urban roads. The spatial distribution of the trajectory on the map is as follows: Figure 2 As shown, Figure 2 The red route in the diagram reflects the vehicle's travel path in the urban road network, including obvious turns and maneuvering sections at intersections. Therefore, it is suitable for testing the improvement effect of maneuvering models and noise adaptive estimation on prediction performance.
[0065] To ensure fairness in the comparison, the experiment, under the same test trajectory, the same observation sequence, and the same evaluation index, compared the autonomous vehicle trajectory prediction and uncertainty assessment method of this application (denoted as the Singer-KF(EM) prediction method) with six representative trajectory prediction methods: LSTM prediction method, Transformer prediction method, CT-EKF prediction method, Baseline KF prediction method, Baseline KF(EM) prediction method, and Singer-KF(Basic) prediction method. LSTM and Transformer represent data-driven sequence prediction methods, CT-EKF represents a filtering method based on a nonlinear maneuvering model; Baseline KF and Baseline KF(EM) are used to evaluate the differences between traditional Kalman frameworks under empirical noise and adaptive noise estimation; Singer-KF(Basic) is used to reflect the performance boundary of "only having a maneuvering model but lacking noise adaptation"; the Singer-KF(EM) prediction method of this application simultaneously possesses two key improvements: maneuvering modeling and noise adaptive estimation. The evaluation indicators selected are root mean square error, mean absolute error, and Q95 quantile error. Among them, Q95 quantile error is used to characterize the upper bound of high confidence error, which can reflect the uncertainty control capability of the prediction results and has stronger engineering safety significance.
[0066] The experimental procedure is as follows: Firstly, based on Figure 2 The trajectory shown extracts the test trajectory sequence from the dataset and performs the necessary preprocessing operations to ensure that the trajectory observation sequence meets the input requirements of each method.
[0067] Subsequently, the Singer-KF(EM) prediction method of this application and six other representative trajectory prediction methods were run to obtain the corresponding vehicle prediction trajectories. For the filtering methods (Baseline KF prediction method, Baseline KF(EM) prediction method, Singer-KF(Basic) prediction method, and Singer-KF(EM) prediction method), a unified state estimation process was used to output the vehicle prediction trajectory.
[0068] The Singer-KF(EM) prediction method in this application performs an expectation-maximization algorithm to adaptively estimate noise parameters before filtering, converging to obtain adaptive EM noise parameters. This allows process noise and observation noise to be adaptively updated according to the characteristics of trajectory data, thereby maintaining a more reasonable noise characterization in different maneuvering phases. For deep learning methods (LSTM prediction method, Transformer prediction method), the predicted vehicle trajectory is output based on a predetermined network structure.
[0069] The predicted trajectories from each method are spatially overlaid with the actual trajectories to obtain a comprehensive visual overlay comparison result, such as... Figure 3 As shown. To further highlight the tracking differences in turning and maneuvering sections, local segments with significant directional changes are selected and magnified for comparison, as shown in the example. Figure 4 As shown, this is used to visually compare the deviation and trajectory continuity of each method during the maneuvering phase. Figure 3 and Figure 4 In the diagram, the horizontal axis represents longitude, the vertical axis represents latitude, Ground Truth represents the actual trajectory, and Trajectory: True vs Predicted represents the comparison between the actual and predicted values of the trajectory.
[0070] Based on the visual comparison, the prediction errors of each method were statistically analyzed, and three quantitative indicators—root mean square error, mean absolute error, and Q95 quantile error—were calculated and summarized in Table 1. Table 1 compares the differences in the overall error and the ability to control the upper bound of high-confidence error among the various methods, thereby verifying the effectiveness and advantages of the method proposed in this application.
[0071] Table 1 Comparison of performance metrics for different method models
[0072] As shown in Table 1, the Singer-KF(EM) prediction method of this application achieves the best results in all three quantitative indicators: root mean square error, mean absolute error, and Q95 quantile error, indicating its comprehensive advantages in overall prediction accuracy and high-confidence error upper bound control. Among these, root mean square error and mean absolute error reflect the overall error level. The Singer-KF(EM) prediction method is significantly lower than the traditional Baseline KF and its EM variants, indicating that introducing the Singer maneuver model can better characterize maneuvering behaviors such as turning, acceleration, and deceleration on urban roads. Simultaneously, the Singer-KF(EM) prediction method of this application has the lowest Q95 quantile error, indicating that the error upper bound under high confidence is smaller, resulting in more reliable prediction results and more controllable risks. Comparing with Singer-KF(Basic), it can be seen that using only the maneuver model still suffers from instability due to the reliance on empirically set noise parameters. EM noise adaptive estimation can further reduce the error and narrow the Q95 range, demonstrating the key role of the combined improvement of "maneuver modeling + noise adaptive estimation". Furthermore, the error metrics of deep learning methods (LSTM, Transformer) on this test trajectory are significantly higher, indicating that purely data-driven methods may suffer from insufficient generalization under conditions of limited training data or scene transfer. In contrast, the method in this application, based on an interpretable maneuver model and achieving noise adaptation through EM, has more stable engineering applicability. The method in this application includes a robust update branch for missing / anomaly observations. The experiments in this embodiment focus on demonstrating the accuracy and uncertainty assessment performance under typical maneuver trajectory scenarios.
[0073] This application's embodiments improve the accuracy and robustness of autonomous vehicle trajectory prediction in complex urban road scenarios by combining the Singer maneuvering target state space model with EM adaptive noise parameters.
[0074] By smoothing and interpolating the trajectory observation sequence, abnormal fluctuations in the trajectory observation sequence are reduced, data gaps are filled, high-quality input is provided for subsequent model estimation, and interference from abnormal data is reduced.
[0075] By introducing the expectation-maximization algorithm for adaptive noise parameter estimation, the EM adaptive noise parameter is obtained by convergence, thus realizing noise parameter adaptation, reducing reliance on human experience, and improving stability under different working conditions.
[0076] By updating the process noise covariance Q and the observation noise covariance R, the vehicle's maneuvering characteristics and sensor noise changes are accurately matched, further optimizing the filtering accuracy.
[0077] By skipping the measurement update step of the Kalman filter at unavailable observation times and only performing the time prediction step, estimation bias caused by erroneous observations is avoided, ensuring the stability and continuity of the filtering process when observations are missing or abnormal, and improving robustness under extreme conditions.
[0078] By employing availability determination to identify whether there are missing or abnormal observations, a reliable basis is provided for filtering branch processing, ensuring the effective implementation of the robust mechanism.
[0079] Uncertainty is assessed by using root mean square error, mean absolute error, and Q95 quantile error, thereby quantifying uncertainty and providing data support for autonomous driving safety decisions.
[0080] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0081] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A method for predicting the trajectory and assessing the uncertainty of an autonomous vehicle, characterized in that, Includes the following steps: Step 1: Acquire vehicle trajectory observation data in urban road scenarios to form a trajectory observation sequence; Step 2: Based on the trajectory observation sequence, and taking the vehicle's motion characteristics as the objective, construct the Singer maneuvering target state space model; Step 3: Based on the Singer maneuvering target state space model, after initializing the noise parameters, the expectation-maximization algorithm is introduced to adaptively estimate the noise parameters, and the EM adaptive noise parameters are obtained by convergence. Step 4: Perform Kalman filtering and smoothing on the Singer maneuvering target state space model that integrates EM adaptive noise parameters, and output the vehicle's predicted trajectory. Step 5: Perform an uncertainty assessment on the predicted vehicle trajectory to obtain the uncertainty assessment result.
2. The method for predicting the trajectory and assessing the uncertainty of an autonomous vehicle according to claim 1, characterized in that, In step one, after the trajectory observation sequence is formed, the trajectory observation sequence is preprocessed, including smoothing and interpolation.
3. The method for trajectory prediction and uncertainty assessment of autonomous vehicles according to claim 1, characterized in that, In step three, the introduction of the expectation-maximization algorithm for adaptive noise parameter estimation, converging to obtain the EM adaptive noise parameters, includes: The expectation estimate of the state is calculated in the E-step of the expectation maximization algorithm, the model noise parameters are updated in the M-step, and the iteration is repeated until convergence to obtain the EM adaptive noise parameters.
4. The method for trajectory prediction and uncertainty assessment of autonomous vehicles according to claim 3, characterized in that, The noise parameters of the model updated in the M-step include the update process noise covariance Q and the observation noise covariance R.
5. The method for trajectory prediction and uncertainty assessment of autonomous vehicles according to claim 1, characterized in that, In step four, when there are missing or abnormal observations, the measurement update step of the Kalman filter is skipped for unavailable observation times, and only the time prediction step is performed.
6. The method for trajectory prediction and uncertainty assessment of autonomous vehicles according to claim 5, characterized in that, The absence or anomalies in the observations are identified using availability determination. The rules for determining availability include: observations being empty / NaN, sampling time intervals exceeding a threshold, and innovation residuals exceeding a threshold.
7. The method for trajectory prediction and uncertainty assessment of autonomous vehicles according to claim 1, characterized in that, In step five, the evaluation indicators for uncertainty assessment include: root mean square error, mean absolute error, and Q95 quantile error.