An automatic driving trajectory planning method based on dynamic risk prediction

Abstract

The application is suitable for the field of automatic driving technology, and provides an automatic driving trajectory planning method based on dynamic risk prediction, which comprises the following steps: converting perception data to a local coordinate system with the ego vehicle as the center to generate time sequence feature vectors of traffic participants and map topology embedding feature vectors; performing node message passing through a heterogeneous graph attention network to extract road right game features and physical boundary constraints, outputting a multi-intention Gaussian mixture parameter set by a Gaussian mixture model density decoder, and correcting through a physical covariance matrix rigidity rotation remodeling mechanism to generate a spatiotemporal dynamic risk field; triggering cross-attention addressing after fusion, outputting candidate trajectories through a multi-modal double-head network, and training based on branch optimization and gradient truncation mechanisms of minimum error matching; and finally recursively generating trajectories conforming to kinematic constraints. The application realizes multi-agent interaction uncertainty modeling, solves the risk field and physical boundary mismatch problem, and improves the safety and executability of trajectory planning.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, and in particular relates to an autonomous driving trajectory planning method based on dynamic risk prediction. Background Technology

[0002] With the development of intelligent connected vehicle technology, autonomous driving systems are expanding from structured scenarios such as highways to open urban roads. Urban traffic contains a large number of long-tail scenarios, such as unprotected intersections, roundabouts, and mixed pedestrian and vehicle sections. The movement trajectories of traffic participants (vehicles, pedestrians, and non-motorized vehicles) are highly random and their intentions are divergent, posing significant challenges to the safety and traffic efficiency of autonomous driving systems.

[0003] In existing technologies, most risk identification and trajectory prediction methods can only output a single, deterministic prediction result, failing to effectively characterize the spatial probability distribution and dynamic risk evolution of the intentions of surrounding traffic participants at future moments, resulting in insufficient collision risk prediction by the system. Furthermore, when faced with multiple reasonable driving options (such as other vehicles potentially turning left, straight, or right at an intersection), existing deep learning-based decision-making and planning models often suffer from a "mean regression" problem due to the optimization mechanism of the loss function (such as mean squared error), outputting a physically infeasible compromise trajectory located between the reasonable trajectories, such as pointing towards the median strip, leading to serious safety accidents.

[0004] Furthermore, existing learning-based planning architectures often directly regress a sequence of two-dimensional spatial coordinate points, ignoring the inherent nonholonomic kinematic constraints of a vehicle as a rigid body, such as Ackermann steering geometry, maximum steering angle limits, and inability to laterally translate. This results in some of the planned trajectories being physically unexecutable, leading to problems such as high-frequency steering wheel vibration and lateral vehicle drift during tracking control, which seriously compromises driving safety and ride comfort. Summary of the Invention

[0005] The purpose of this invention is to provide an autonomous driving trajectory planning method based on dynamic risk prediction, which aims to solve the problems mentioned in the background art.

[0006] The present invention is implemented as follows: an autonomous driving trajectory planning method based on dynamic risk prediction includes the following steps:

[0007] Step 1: Acquire motion state data and static environment data of the vehicle and surrounding traffic participants, transform the data from the global absolute coordinate system to the local coordinate system centered on the vehicle, encode the temporal kinematic state of the surrounding traffic participants to generate temporal feature vectors of traffic participants, and encode the static environment to generate map topology embedding feature vectors.

[0008] Step 2: The temporal feature vectors of traffic participants are concatenated with the map topology embedding feature vectors, and then fed into a heterogeneous graph attention network for heterogeneous node message passing. The road right-of-way game features and physical boundary constraints of multi-vehicle interactions are extracted. The output features are input into the Gaussian mixture model density decoder, which outputs a Gaussian mixture parameter set of three potential intentions of surrounding traffic participants within the next sixteen time steps. A rigid rotation and reshaping mechanism of the physical covariance matrix is ​​constructed using the physical dimensions and heading angle of the target vehicle to correct the Gaussian mixture parameter set and generate a spatiotemporal dynamic risk field with physical covariance correction components.

[0009] Step 3: Fuse the spatiotemporal dynamic risk field with the temporal feature vectors of traffic participants to generate a traffic participant feature vector that incorporates future uncertainty risks. Concatenate this feature vector with the map topology embedding feature vector to form a context key-value pair. Use the vehicle temporal feature vector as the global query vector to trigger global explicit cross-attention addressing and output a comprehensive game context vector. Input the comprehensive game context vector into a multimodal dual-head network containing parallel action decoding heads and probability classification heads to output multiple sets of candidate trajectory control sequences and their corresponding intent confidence scores. Employ a branch optimization and gradient truncation mechanism based on minimum error matching to determine the champion trajectory that is closest to the real trajectory. Calculate the regression loss only for the champion trajectory and perform synchronous supervised training on the probability classification head.

[0010] Step 4: Input the longitudinal acceleration of the vehicle and the front wheel steering angle output by the motion decoder into the differentiable Ackerman vehicle kinematics model, and use the discrete Euler difference integral equation system to recursively deduce the vehicle pose and generate a planned trajectory that conforms to the vehicle kinematic constraints.

[0011] A further technical solution, in step one, involves transforming the data from a global absolute coordinate system to a local coordinate system centered on the vehicle, including:

[0012] Real-time extraction of the global absolute longitudinal coordinates of the vehicle's center of gravity at the current moment. Global absolute horizontal coordinates With the current global absolute heading angle By using the inverse rotation projection matrix, the global absolute longitudinal coordinates of all dynamic objects and target nodes in the static environment within the system's perception field of view are obtained. and global absolute horizontal coordinates The vertical relative coordinates are uniformly projected onto the local coordinate system. and horizontal relative coordinates For any target within the sensing range, the formula for inverse rotation projection from global absolute spatial coordinates to the local coordinate system is as follows:

[0013]

[0014]

[0015] At the same time, the longitudinal velocities of surrounding traffic participants in the global absolute coordinate system are also monitored. and lateral speed Perform the same inverse rotational transformation to convert it into its longitudinal relative velocity component in a local coordinate system centered on the vehicle. and lateral relative velocity components The velocity component mapping transformation formula is shown below:

[0016]

[0017] .

[0018] A further technical solution, in step one, involves encoding the temporal kinematic states of surrounding traffic participants to generate temporal feature vectors for those participants, including:

[0019] The local temporal kinematic states, physical external dimensions, and categories of continuous historical frames of surrounding traffic participants are integrated into a tensor and input into a long short-term memory network temporal encoder. The hidden layer activation values ​​of the final state of the sequence are extracted using a recurrent gating mechanism to generate a traffic participant temporal feature vector with the historical motion trend of the surrounding traffic participants.

[0020] A further technical solution, in step one, involves encoding the static environment to generate map topology embedding feature vectors, which includes:

[0021] The point set of the vector map is reconstructed by fixed-point downsampling, and the discrete states of online traffic lights are fused together. After normalization by a multilayer perceptron, a map topology embedding feature vector is generated.

[0022] In a further technical solution, in step two, the heterogeneous graph attention network performs heterogeneous node message passing as follows: using other vehicle nodes as query vectors, it performs cross-attention addressing interaction with other vehicle nodes and lane line feature nodes in the global graph topology network, and extracts the right-of-way game features and physical boundary constraints of multi-vehicle interaction.

[0023] A further technical solution involves, in step two, explicitly decomposing the original Gaussian distribution parameters output by the Gaussian mixture model density decoder into multimodal mixture weight reset confidence, predicted mean center location, basic standard deviation, and correlation coefficient intermediate variables. Using activation functions and hard numerical truncation functions, the numerical range of the basic standard deviation is nonlinearly mapped and bounded, as shown in the following formula:

[0024]

[0025]

[0026] in, and The longitudinal and lateral baseline standard deviations of the network predictions after truncation and activation processing; It is a positive activation function, and , ; and These are the raw regression values ​​for the longitudinal and transverse directions output by the decoder. and Boundary limits set to prevent infinite expansion or infinite collapse of variance; To find the minimum value function; The function is for finding the maximum value.

[0027] A further technical solution, in step two, includes the following rigid rotation reshaping mechanism for the physical covariance matrix:

[0028] Real-time extraction of the inherent physical length of the target vehicle body With width Based on this, the longitudinal physical variance benchmark for vehicles is defined respectively. Compared with the cross-sectional physical variance benchmark The corresponding calculation formula is shown below:

[0029]

[0030]

[0031] Then, extract the heading angle of the last local observation frame after the target is aligned with the vehicle at the current moment. Using this as a rigid rotation angle to construct a rotation matrix, a rigid similarity transformation is performed on the previous vehicle longitudinal physical variance benchmark and lateral physical variance benchmark to obtain physical covariance correction components that match the vehicle's rectangular envelope attitude, including the longitudinal physical variance correction component. Lateral physical variance correction component and physical covariance cross-correction components The specific mathematical expression is as follows:

[0032]

[0033]

[0034] .

[0035] A further technical solution involves, in step two, after obtaining the physical covariance correction component that conforms to the geometric orientation of the vehicle body, coupling it linearly with the baseline standard deviation predicted by the network itself to calculate the final core parameters of the Gaussian mixture risk field containing rigid body boundary constraints, including the longitudinal standard deviation. Horizontal standard deviation and correlation coefficient This forms a spatiotemporal dynamic risk field with a physical covariance correction component, which is output to step three. The synthesis formula is shown below:

[0036]

[0037]

[0038]

[0039] in, The hyperbolic tangent activation function is used. These are the raw correlation coefficient feature values ​​directly output by the decoder. A safety constant to prevent the denominator from being zero;

[0040] When training spatiotemporal dynamic risk prediction, a negative log-likelihood loss function based on logarithmic and exponential spillover prevention mechanisms is introduced. The loss function and the intermediate multivariate Gaussian-Mahalanx distance spatial term are calculated as follows:

[0041]

[0042]

[0043] in, The total number of pre-defined interactive intents for the Gaussian mixture model. For the currently calculated number of One intention, The first branch prediction of the network probability The confidence level of an intention occurring For expert demonstration data in the first The quadratic term of the two-dimensional Mahalanobis distance of an intention. and These are the actual longitudinal and lateral spatial coordinates of the vehicle in the future, based on the expert demonstration data. and They were respectively his car in the first The vertical and horizontal coordinates of the mean center location of the risk field prediction for each intention. For the first The correlation coefficient of an intention For the first The longitudinal standard deviation of each intention In the first The lateral standard deviation of each intention.

[0044] A further technical solution, in step three, includes a branch selection and gradient truncation mechanism based on minimum error matching, comprising:

[0045] Calculate multiple candidate trajectories Real driving trajectory of experts The average Euclidean distance is used to determine the unique champion trajectory closest to the true value by finding its minimum value. It also blocks the losing candidate branches and calculates the robust Huber regression loss for the winning mode only. The formula is as follows:

[0046]

[0047] in, The Hubble regression loss function;

[0048] Simultaneously, the index of the champion trajectory is defined as the true label of the classification supervision, and cross-entropy loss is used. The probabilistic classification head is subjected to synchronized supervision and penalty, and the loss function formula is as follows:

[0049]

[0050] in, The total number of candidates set for the planner. For the currently calculated number of One candidate trajectory, The first output of the probability classification head Confidence scores for each candidate trajectory The confidence score for the champion's trajectory.

[0051] A further technical solution, the specific steps of step four are as follows:

[0052] A differentiable Ackerman vehicle kinematics model based on pure tensor operations is embedded at the end of the motion decoder. The motion decoder outputs the vehicle's longitudinal acceleration and front wheel steering angle, which conform to chassis motion. Tensor values ​​are truncated before being input into the kinematics model, limiting the input values ​​to the vehicle's mechanical limits. The kinematics model extracts the vehicle's current instantaneous velocity. With instantaneous local heading angle Using these as initial values, the vehicle pose is recursively calculated along the time axis using the discrete Euler difference integral equations system for the acceleration and front wheel steering angle sequences. The corresponding recursive calculation formula is as follows:

[0053]

[0054]

[0055]

[0056]

[0057] in, and These are the vehicle's local longitudinal and lateral coordinates at the current moment, respectively. , , and These represent the longitudinal coordinates, lateral coordinates, local heading angle, and speed of the vehicle at the next time step after recursion. For a fixed time step, The inherent front and rear wheelbase geometry of the vehicle chassis. For longitudinal acceleration, Front wheel steering angle.

[0058] The autonomous driving trajectory planning method based on dynamic risk prediction provided in this embodiment of the invention has the following advantages:

[0059] (1) In terms of dynamic risk prediction, the cross-attention interaction between multi-vehicle nodes and lane line nodes is carried out through heterogeneous graph attention network, fully extracting right-of-way game and physical boundary features, and outputting multi-intention probability distribution by combining Gaussian mixture model, which significantly improves the accuracy of multi-subject interaction uncertainty modeling.

[0060] (2) A rigid rotation reshaping mechanism for the physical covariance matrix is ​​introduced. The target vehicle body length, width and heading angle are used to perform a rigid similarity transformation on the physical variance benchmark to obtain the covariance correction component that fits the rectangular envelope attitude of the vehicle. It is coupled with the network prediction standard deviation residual to solve the distortion problem caused by the mismatch between the traditional Gaussian risk field and the actual boundary of the vehicle.

[0061] (3) A branch selection and gradient cutoff mechanism based on minimum error matching is designed. The regression loss is calculated only for the champion trajectory and the gradient of the losing branch is blocked. At the same time, the cross-entropy loss is used to supervise the probability classification head, which effectively overcomes the failure problems such as compromise and wall crossing caused by multimodal trajectory mean regression and achieves complete decoupling of multiple intentions.

[0062] (4) Embed a differentiable Ackerman vehicle kinematics model, output acceleration and front wheel angle from the motion decoding head, generate trajectory through pure tensor recursion, support lossless propagation of reverse gradient, set anti-negative value defense for velocity recursion, and ensure that the planned trajectory strictly meets the chassis kinematic constraints. Attached Figure Description

[0063] Figure 1A flowchart illustrating an autonomous driving trajectory planning method based on dynamic risk prediction, provided in an embodiment of the present invention.

[0064] Figure 2 A detailed flowchart of step two in an autonomous driving trajectory planning method based on dynamic risk prediction provided in an embodiment of the present invention;

[0065] Figure 3 A detailed flowchart of step three in an autonomous driving trajectory planning method based on dynamic risk prediction provided in an embodiment of the present invention;

[0066] Figure 4 This is a detailed flowchart of step four in an autonomous driving trajectory planning method based on dynamic risk prediction, provided in an embodiment of the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0068] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0069] like Figure 1 As shown, an embodiment of the present invention provides an autonomous driving trajectory planning method based on dynamic risk prediction, comprising the following steps:

[0070] Step 1: Data Acquisition and Preprocessing;

[0071] Firstly, high-precision spatiotemporal pose alignment is achieved through online coordinate transformation and vectorized feature reconstruction, enabling the autonomous driving system to transform data from a global absolute coordinate system to a local coordinate system centered on the vehicle at every moment of operation. Specifically, the system extracts the global absolute longitudinal coordinates of the vehicle's centroid in real time. Global absolute horizontal coordinates With the current global absolute heading angle By using an inverse rotation projection matrix, the global absolute longitudinal coordinates of all dynamic objects and static environments within the system's perception field of view—namely, surrounding traffic participants and vector lane network topology segments—are determined. These coordinates include the target nodes (such as the current location of other vehicles and map topology control points). and global absolute horizontal coordinates The vertical relative coordinates are uniformly projected onto the local coordinate system. and horizontal relative coordinates For any target within the perception range, the formula for inverse rotation projection from global absolute spatial coordinates to the local coordinate system is as follows:

[0072]

[0073]

[0074] While aligning the spatial coordinate system, the system also monitors the longitudinal velocities of surrounding traffic participants in the global absolute coordinate system. and lateral speed Perform the same inverse rotational transformation to convert it into its longitudinal relative velocity component in a local coordinate system centered on the vehicle. and lateral relative velocity components This ensures that the physical derivation of the velocity direction and relative pose is consistent. The velocity component mapping transformation formula is shown below:

[0075]

[0076]

[0077] After the aforementioned data preprocessing, the system needs to encode each feature. First, the local temporal kinematic state, physical external dimensions, and category (car, pedestrian, and non-motorized vehicle) of continuous historical frames of surrounding traffic participants are integrated into a tensor and input into a Long Short-Term Memory (LSTM) temporal encoder. A recurrent gating mechanism is used to extract the hidden layer activation values ​​of the final sequence state, generating a temporal feature vector of traffic participants with their historical motion trends. Next, for the static environment, the system reconstructs the point set of the vector map through fixed-point downsampling, while organically integrating the discrete states of online traffic lights. After normalization processing using a multilayer perceptron, a map topology embedding feature vector is generated.

[0078] Step Two: Dynamic Risk Prediction;

[0079] like Figure 2As shown, the system constructs a spatiotemporal dynamic risk field prediction module based on a heterogeneous graph attention network and a Gaussian mixture model to characterize the uncertainty of dynamic interactions. The temporal feature vectors of traffic participants are concatenated with the map topology embedding feature vectors and then fed into the heterogeneous graph attention network. A multi-head attention mechanism facilitates message passing between heterogeneous nodes. The network uses other vehicle nodes as query vectors and engages in deep cross-attention addressing interactions with other vehicle nodes and lane line feature nodes in the global graph topology network, thereby extracting the right-of-way game features and physical boundary constraints of multi-vehicle interactions. Its output features are then input into the Gaussian mixture model density decoder, ultimately outputting a set of Gaussian mixture parameters predicting the three potential divergent intentions of surrounding traffic participants over the next sixteen time steps. Furthermore, to address the data divergence and instability issues easily caused by the Gaussian variance output of traditional networks, the original Gaussian distribution parameters output by this decoder are explicitly decomposed into multimodal mixture weighted confidence, predicted mean center location, base standard deviation, and correlation coefficient intermediate variables. Furthermore, by utilizing activation functions and hard numerical cutoff functions, the numerical range of the basic standard deviation is nonlinearly mapped and bounded. The calculation formula is as follows:

[0080]

[0081]

[0082] in, and The longitudinal and lateral baseline standard deviations of the network predictions after truncation and activation processing; It is a positive activation function, and , ; and These are the raw regression values ​​for the longitudinal and transverse directions output by the decoder. and Boundary limits set to prevent infinite expansion or infinite collapse of variance; To find the minimum value function; The function is for finding the maximum value.

[0083] Simultaneously, a rigid rotation and reshaping mechanism for the physical covariance matrix is ​​introduced. The specific operating mechanism is to extract the inherent physical length of the target vehicle body in real time. With width Based on this, the longitudinal physical variance benchmark for vehicles is defined respectively. Compared with the cross-sectional physical variance benchmark The corresponding calculation formula is shown below:

[0084]

[0085]

[0086] Then, extract the heading angle of the last local observation frame after the target is aligned with the vehicle at the current moment. This is used as a rigid rotation angle to construct a rotation matrix, and a rigid similarity transformation is performed on the previous vehicle longitudinal and lateral physical variance benchmarks. Through algebraic expansion and matrix multiplication derivation, the physical covariance correction components that match the vehicle's rectangular envelope attitude are finally obtained, including the longitudinal physical variance correction component. Lateral physical variance correction component and physical covariance cross-correction components The specific mathematical expression is as follows:

[0087]

[0088]

[0089]

[0090] After obtaining the physical covariance correction component that conforms to the vehicle body's geometric orientation, it is linearly coupled with the baseline standard deviation predicted by the network itself to calculate the core parameters of the final Gaussian mixture risk field containing rigid body boundary constraints, including the longitudinal standard deviation. Horizontal standard deviation and correlation coefficient (The range of values ​​is constrained by) (To ensure the matrix is ​​positive definite), a spatiotemporal dynamic risk field with physical covariance correction components is formed and output to step three. The synthesis formula is shown below:

[0091]

[0092]

[0093]

[0094] in, The hyperbolic tangent activation function is used. These are the raw correlation coefficient feature values ​​directly output by the decoder. A safety constant to prevent the denominator from being zero.

[0095] When training spatiotemporal dynamic risk prediction, a negative log-likelihood loss function based on logarithmic and exponential spillover prevention mechanisms is introduced. This loss function is used to solve the maximum likelihood estimation problem of multimodal probability density. At the computational graph level, it effectively avoids the numerical overflow caused by excessively large or small exponential terms in traditional Gaussian density functions, ensuring the smoothness of gradient backpropagation under complex long-tailed samples. The formulas for the loss function and the intermediate multivariate Gaussian Mahalanobis distance spatial term are as follows:

[0096]

[0097]

[0098] in, The total number of pre-defined interactive intents for the Gaussian mixture model. For the currently calculated number of One intention, The first branch prediction of the network probability The confidence level of an intention occurring For expert demonstration data in the first The quadratic term of the two-dimensional Mahalanobis distance of an intention. and These are the actual longitudinal and lateral spatial coordinates of the vehicle in the future, based on the expert demonstration data. and They were respectively his car in the first The vertical and horizontal coordinates of the mean center location of the risk field prediction for each intention. For the first The correlation coefficient of an intention For the first The longitudinal standard deviation of each intention In the first The lateral standard deviation of each intention.

[0099] Step 3: Multimodal trajectory planning;

[0100] After obtaining the spatiotemporal dynamic risk field with physical covariance correction components, the system feeds it into the multimodal trajectory planning module to completely decouple the multimodal intent, such as... Figure 3 As shown, after dimensionality reduction and flattening, it is added element-wise to the temporal feature vector of traffic participants to generate a traffic participant feature vector that incorporates future uncertainty risks. This vector is then concatenated with the map topology embedding feature vector to form a context key-value pair. Vehicle data is temporally encoded using a long short-term memory network to form a vehicle temporal feature vector, which is used as a global query vector to trigger global explicit cross-attention addressing.

[0101] During the computation, a mask matrix is ​​used to completely remove the attention weights of invalid zero-padded nodes at the exponential layer. The cross-attention mechanism outputs a comprehensive game context vector, which is then fed into a multimodal dual-head network containing parallel action decoding and probability classification heads. This network outputs multiple sets of parallel candidate trajectory control sequences and their corresponding intent confidence scores. To address system failures such as mean regression-induced wall-crossing, a branch optimization and gradient truncation mechanism based on minimum error matching is designed. When the network outputs multiple predicted candidate trajectories, this mechanism calculates the error between all candidate trajectories and the actual driver's trajectory, selects the single candidate trajectory with the smallest error as the optimal branch, and only allows the backpropagation of the loss gradient on this optimal branch. Simultaneously, the backpropagation gradients of other branches are truncated, preventing them from participating in network parameter updates. Specifically, the system calculates multiple sets of candidate trajectories. Real driving trajectory of experts The average Euclidean distance is used to determine the unique champion trajectory closest to the true value by finding its minimum value. It also blocks the losing candidate branches and calculates the robust Huber regression loss for the winning mode only. The formula is as follows:

[0102]

[0103] in, Let be the Hubble regression loss function.

[0104] Meanwhile, the system defines the index of the champion's trajectory as the true label for classification supervision, and utilizes cross-entropy loss. The probabilistic classification head is simultaneously supervised and penalized to ensure that the network assigns the highest confidence score to the winning trajectory in the current scene. The loss function formula is as follows:

[0105]

[0106] in, The total number of candidates set for the planner. For the currently calculated number of One candidate trajectory, The first output of the probability classification head Confidence scores for each candidate trajectory The confidence score for the champion's trajectory.

[0107] By combining Huber regression loss with cross-entropy loss, the system achieves complete decoupling of the network planning intent, enabling the system to decisively generate multiple trajectories and decide on the safest route that best matches the expert's driving when faced with multiple driving options such as intersections.

[0108] Step 4: Vehicle kinematics decoding;

[0109] Finally, to address the issue of some planned trajectories deviating from the vehicle chassis kinematics constraints, a differentiable Ackerman vehicle kinematics model based on pure tensor operations was embedded at the end of the motion decoding head, such as... Figure 4 As shown. The motion decoder does not directly output two-dimensional spatial coordinates, but instead outputs the vehicle's longitudinal acceleration and front wheel steering angle, which conform to chassis motion. Tensor values ​​are truncated before being input into the kinematic model, restricting the input values ​​to the vehicle's mechanical limits, where the longitudinal acceleration... Controlled within [-6.0, 4.0] m / s 2 Front wheel cornering The speed is controlled within [-0.5, 0.5] rad. The kinematic model is extracted from the vehicle's current instantaneous speed. With instantaneous local heading angle As initial values, the vehicle pose is recursively derived along the time axis using the discrete Euler difference integral equations for the acceleration and front wheel steering angle sequences. Simultaneously, to prevent the forward velocity integral from erroneously becoming negative due to numerical perturbations under extreme deceleration conditions, a safeguard is implemented in the velocity state recursive equations. The corresponding recursive calculation formulas are as follows:

[0110]

[0111]

[0112]

[0113]

[0114] in, and These are the vehicle's local longitudinal and lateral coordinates at the current moment, respectively. , , and These represent the longitudinal coordinates, lateral coordinates, local heading angle, and speed of the vehicle at the next time step after recursion. The time step is fixed (set to 0.2 s). Let be the inherent front and rear wheel geometry of the vehicle chassis (set to 2.9 m).

[0115] Because the kinematic equations use pure tensors, the entire integral recursion possesses fully continuous and differentiable mathematical properties, thus supporting lossless propagation of the backpropagation error gradient. Simultaneously, a branch optimization and gradient truncation mechanism based on minimum error matching is designed to cut off and select candidate trajectories based on the confidence scores output by the probability classification head, filtering out the optimal trajectory. Through step 4, the anthropomorphic safe planning trajectory finally output by the network is constrained within the kinematic boundaries.

[0116] To fully demonstrate the advancement of this method, the displacement error of this method is compared with that of some baseline methods in different prediction time domains (see Table 1 for details). This comparison covers traditional deep learning, classic graph neural networks, and various mainstream autonomous driving planning baseline models.

[0117] In the real-world performance of autonomous driving systems, the ability of a vehicle to accurately follow a safe and reasonable expected route is the most intuitive standard for evaluating the quality of a planning algorithm. Average Displacement Error (ADE) and Final Displacement Error (FDE) are key indicators for quantifying this core spatial capability. ADE reflects the average physical distance the vehicle deviates from the ideal human driving trajectory over the entire future time period, while FDE specifically measures the magnitude of the positional deviation at the exact moment of the predicted time.

[0118] Table 1. Comparison of displacement errors of the present invention and some baseline methods in different prediction time domains.

[0119]

[0120] (Baseline IL, FF, EO, and ST-P3 data are from relevant comparative literature; LSTM and VectorNet data are from the official Argoverse1 test.)

[0121] During the core game window of 3.0 seconds, the model's ADE was only 1.02 meters, and FDE was controlled at around 1.89 meters. In complex urban intersections that are often tens of meters wide, an average comprehensive error of 1.02 meters means that vehicles can steadily stay near the center line of their current lane, with very few instances of crossing the line or dangerous intrusion into adjacent lanes.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic risk prediction based automatic driving trajectory planning method, characterized in that, Includes the following steps: Step 1: Acquire motion state data and static environment data of the vehicle and surrounding traffic participants, transform the data from the global absolute coordinate system to the local coordinate system centered on the vehicle, encode the temporal kinematic state of the surrounding traffic participants to generate temporal feature vectors of traffic participants, and encode the static environment to generate map topology embedding feature vectors. Step 2: The temporal feature vectors of traffic participants are concatenated with the map topology embedding feature vectors, and then fed into a heterogeneous graph attention network for heterogeneous node message passing. The road right-of-way game features and physical boundary constraints of multi-vehicle interactions are extracted. The output features are input into the Gaussian mixture model density decoder, which outputs a Gaussian mixture parameter set of three potential intentions of surrounding traffic participants within the next sixteen time steps. A rigid rotation and reshaping mechanism of the physical covariance matrix is ​​constructed using the physical dimensions and heading angle of the target vehicle to correct the Gaussian mixture parameter set and generate a spatiotemporal dynamic risk field with physical covariance correction components. Step 3: Fuse the spatiotemporal dynamic risk field with the temporal feature vectors of traffic participants to generate a traffic participant feature vector that incorporates future uncertainty risks. Concatenate this feature vector with the map topology embedding feature vector to form a context key-value pair. Use the vehicle temporal feature vector as the global query vector to trigger global explicit cross-attention addressing and output a comprehensive game context vector. Input the comprehensive game context vector into a multimodal dual-head network containing parallel action decoding heads and probability classification heads to output multiple sets of candidate trajectory control sequences and their corresponding intent confidence scores. Employ a branch optimization and gradient truncation mechanism based on minimum error matching to determine the champion trajectory that is closest to the real trajectory. Calculate the regression loss only for the champion trajectory and perform synchronous supervised training on the probability classification head. Step 4: Input the longitudinal acceleration of the vehicle and the front wheel steering angle output by the motion decoder into the differentiable Ackerman vehicle kinematics model, and use the discrete Euler difference integral equation system to recursively deduce the vehicle pose and generate a planned trajectory that conforms to the vehicle kinematic constraints.

2. The autonomous driving trajectory planning method based on dynamic risk prediction according to claim 1, characterized in that, In step one, the process of transforming the data from the global absolute coordinate system to a local coordinate system centered on the vehicle includes: Real-time extraction of the global absolute longitudinal coordinates of the vehicle's center of gravity at the current moment. Global absolute horizontal coordinates With the current global absolute heading angle By using the inverse rotation projection matrix, the global absolute longitudinal coordinates of all dynamic objects and target nodes in the static environment within the system's perception field of view are obtained. and global absolute horizontal coordinates The longitudinal relative coordinates are uniformly projected onto a local coordinate system centered on the vehicle. and horizontal relative coordinates For any target within the sensing range, the formula for inverse rotation projection from global absolute spatial coordinates to the local coordinate system is as follows: At the same time, the longitudinal velocities of surrounding traffic participants in the global absolute coordinate system are also monitored. and lateral speed Perform the same inverse rotational transformation to convert it into its longitudinal relative velocity component in a local coordinate system centered on the vehicle. and lateral relative velocity components The velocity component mapping transformation formula is shown below: 。 3. The autonomous driving trajectory planning method based on dynamic risk prediction according to claim 2, characterized in that, In step one, the process of encoding the temporal kinematic states of surrounding traffic participants to generate temporal feature vectors for traffic participants includes: The local temporal kinematic state, physical external size, and category of continuous historical frames of surrounding traffic participants are integrated into a tensor and input into a long short-term memory network temporal encoder. The hidden layer activation values ​​of the final state of the sequence are extracted using a recurrent gating mechanism to generate a traffic participant temporal feature vector with the historical motion trend of the surrounding traffic participants. The process of encoding a static environment to generate map topology embedding feature vectors includes: The point set of the vector map is reconstructed by fixed-point downsampling, and the discrete states of online traffic lights are fused together. After normalization by a multilayer perceptron, a map topology embedding feature vector is generated.

4. The autonomous driving trajectory planning method based on dynamic risk prediction according to claim 3, characterized in that, In step two, the heterogeneous graph attention network performs heterogeneous node message passing as follows: using other vehicle nodes as query vectors, it performs cross-attention addressing interaction with other vehicle nodes and lane line feature nodes in the global graph topology network, and extracts the right-of-way game features and physical boundary constraints of multi-vehicle interaction.

5. The autonomous driving trajectory planning method based on dynamic risk prediction according to claim 4, characterized in that, In step two, the original Gaussian distribution parameters output by the Gaussian mixture model density decoder are explicitly decomposed into multimodal mixture weighted confidence, predicted mean center location, base standard deviation, and correlation coefficient intermediate variables. Using activation functions and hard numerical truncation functions, the numerical range of the base standard deviation is nonlinearly mapped and bounded, as shown in the following formula: in, and The longitudinal and lateral baseline standard deviations of the network predictions after truncation and activation processing; It is a positive activation function, and , ; and These are the raw regression values ​​for the longitudinal and transverse directions output by the decoder. and Boundary limits set to prevent infinite expansion or infinite collapse of variance; To find the minimum value function; The function is for finding the maximum value.

6. The autonomous driving trajectory planning method based on dynamic risk prediction according to claim 5, characterized in that, In step two, the rigid rotation reshaping mechanism of the physical covariance matrix includes: Real-time extraction of the inherent physical length of the target vehicle body With width Based on this, the longitudinal physical variance benchmark for vehicles is defined respectively. Compared with the cross-sectional physical variance benchmark The corresponding calculation formula is shown below: Then, extract the heading angle of the last local observation frame after the target is aligned with the vehicle at the current moment. Using this as a rigid rotation angle to construct a rotation matrix, a rigid similarity transformation is performed on the previous vehicle longitudinal physical variance benchmark and lateral physical variance benchmark to obtain physical covariance correction components that match the vehicle's rectangular envelope attitude, including the longitudinal physical variance correction component. Lateral physical variance correction component and physical covariance cross-correction components The specific mathematical expression is as follows: 。 7. The autonomous driving trajectory planning method based on dynamic risk prediction according to claim 6, characterized in that, In step two, after obtaining the physical covariance correction component that conforms to the geometric orientation of the vehicle body, it is linearly coupled with the baseline standard deviation predicted by the network itself to calculate the core parameters of the final Gaussian mixture risk field containing rigid body boundary constraints, including the longitudinal standard deviation. Horizontal standard deviation and correlation coefficient This forms a spatiotemporal dynamic risk field with a physical covariance correction component, which is output to step three. The synthesis formula is shown below: in, The hyperbolic tangent activation function is used. These are the raw correlation coefficient feature values ​​directly output by the decoder. A safety constant to prevent the denominator from being zero; When training spatiotemporal dynamic risk prediction, a negative log-likelihood loss function based on logarithmic and exponential spillover prevention mechanisms is introduced. The loss function and the intermediate multivariate Gaussian-Mahalanx distance spatial term are calculated as follows: in, The total number of pre-defined interactive intents for the Gaussian mixture model. For the currently calculated number of One intention, The first branch prediction of the network probability The confidence level of an intention occurring For expert demonstration data in the first The quadratic term of the two-dimensional Mahalanobis distance of an intention. and These are the actual longitudinal and lateral spatial coordinates of the vehicle in the future, based on the expert demonstration data. and They were respectively his car in the first The vertical and horizontal coordinates of the mean center location of the risk field prediction for each intention. For the first The correlation coefficient of an intention For the first The longitudinal standard deviation of each intention In the first The lateral standard deviation of each intention.

8. The autonomous driving trajectory planning method based on dynamic risk prediction according to claim 7, characterized in that, In step three, the branch selection and gradient truncation mechanism based on minimum error matching includes: Calculate multiple candidate trajectories Real driving trajectory of experts The average Euclidean distance is used to determine the unique champion trajectory closest to the true value by finding its minimum value. It also blocks the losing candidate branches and calculates the robust Huber regression loss for the winning mode only. The formula is as follows: in, The Hubble regression loss function; Simultaneously, the index of the champion trajectory is defined as the true label of the classification supervision, and cross-entropy loss is used. The probabilistic classification head is subjected to synchronized supervision and penalty, and the loss function formula is as follows: in, The total number of candidates set for the planner. For the currently calculated number of One candidate trajectory, The first output of the probability classification head Confidence scores for each candidate trajectory The confidence score for the champion's trajectory.

9. The autonomous driving trajectory planning method based on dynamic risk prediction according to claim 8, characterized in that, The specific steps of step four are as follows: A differentiable Ackerman vehicle kinematics model based on pure tensor operations is embedded at the end of the motion decoder. The motion decoder outputs the vehicle's longitudinal acceleration and front wheel steering angle, which conform to chassis motion. Tensor values ​​are truncated before being input into the kinematics model, limiting the input values ​​to the vehicle's mechanical limits. The kinematics model extracts the vehicle's current instantaneous velocity. With instantaneous local heading angle Using these as initial values, the vehicle pose is recursively calculated along the time axis using the discrete Euler difference integral equations system for the acceleration and front wheel steering angle sequences. The corresponding recursive calculation formula is as follows: in, and These are the vehicle's local longitudinal and lateral coordinates at the current moment, respectively. , , and These represent the longitudinal coordinates, lateral coordinates, local heading angle, and speed of the vehicle at the next time step after recursion. For a fixed time step, The inherent front and rear wheelbase geometry of the vehicle chassis. For longitudinal acceleration, Front wheel steering angle.

Images