Autonomous driving risk perception and trajectory planning method fusing vehicle state prediction

Abstract

The application belongs to the technical field of automatic driving, and particularly relates to an automatic driving risk perception and trajectory planning method fusing vehicle state prediction. The method comprises the following steps: collecting historical motion information of an obstacle vehicle and surrounding vehicles and road structural information, and constructing a space-time feature representation based on the historical motion information of the vehicle; combining a Transformer and a multi-task learning framework to construct a vehicle state prediction model, so as to realize prediction of a future motion state of the vehicle; subsequently, constructing a road risk field model and an obstacle vehicle risk field model based on an artificial potential field theory, and constructing a comprehensive risk field model through a weighted superposition manner; the future comprehensive risk field is jointly constituted by the obstacle vehicle risk field and the global road risk field of a future time period; a candidate trajectory cluster of an automatic driving vehicle is generated by using a quintic polynomial, a low-risk trajectory set satisfying a safety threshold is screened out, and a comfort evaluation function is introduced, so as to obtain an optimal planning trajectory.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and specifically to an autonomous driving risk perception and trajectory planning method that integrates vehicle state prediction. Background Technology

[0002] With the rapid development of autonomous driving technology, the operational safety of intelligent vehicles in complex traffic environments has become an increasingly important concern. While traditional path planning methods can generate feasible paths for vehicles, they often fail to adequately characterize the uncertain behaviors and potential risks of other traffic participants. Therefore, risk field-based path planning methods are gaining traction. These methods transform elements such as obstacles, road conditions, and vehicles into a risk field, enabling quantitative modeling of driving environment risks and providing safety constraints for vehicle decision-making. Existing research shows that risk field modeling has evolved from early static artificial potential fields to models that comprehensively consider vehicle size, speed, driving direction, and the dynamic motion characteristics of obstacles. Furthermore, some methods introduce data-driven prediction techniques to estimate the future movement trends of traffic participants and integrate the prediction results into risk field modeling to describe the spatiotemporal distribution of environmental risks. Based on this, safe, smooth, and efficient driving trajectories can be generated through sampling, optimization, or candidate trajectory generation methods, and these methods are gradually being extended to complex scenarios such as multi-vehicle interactions and mixed vehicle-pedestrian traffic, improving the versatility and robustness of the methods.

[0003] However, existing technologies still have significant limitations. On the one hand, most risk field models are insufficient in modeling obstacle behavior patterns and future situations, making it difficult to adapt to complex and ever-changing traffic environments. On the other hand, even with the introduction of trajectory prediction information, the computational cost of high-precision prediction remains significant, affecting real-time performance. Furthermore, local extrema problems and path feasibility in high-density scenarios remain challenges. Therefore, there is an urgent need to develop a future risk field modeling method that can integrate vehicle state prediction to achieve accurate risk identification in dynamic environments and guide safe and comfortable path planning. Summary of the Invention

[0004] In view of the above problems, the purpose of this invention is to provide an autonomous driving risk perception and trajectory planning method that integrates vehicle state prediction. This method is used to construct a vehicle state prediction model based on Transformer network and multi-task learning framework to predict the future motion state of vehicles with obstacles. Based on this, it combines artificial potential field theory to construct a future comprehensive risk field, thereby achieving forward-looking risk identification of the autonomous driving vehicle's driving environment. Furthermore, it generates candidate trajectory clusters through a fifth-order polynomial and combines the future risk field to perform safe and comfortable trajectory planning, thereby improving the safety, rationality, and forward-looking decision-making of the autonomous driving system in complex traffic environments, and overcoming the shortcomings of the prior art.

[0005] This invention provides a method for autonomous driving risk perception and trajectory planning that integrates vehicle state prediction, comprising the following steps:

[0006] Step 1: Collect historical motion information and static road structure information of the obstacle vehicle and its surrounding vehicles. Perform unified vectorization processing on the acquired historical motion information and static road structure information. Further standardize and preprocess the historical motion information to obtain the spatiotemporal feature representation of the obstacle vehicle and its surrounding vehicles.

[0007] Step 2: Combine Transformer (neural network model) and multi-task learning framework to build a vehicle state prediction model. Use the spatiotemporal features of the obstacle vehicle and its surrounding vehicles as the input features of the vehicle state prediction model to predict the future motion state of the obstacle vehicle.

[0008] Step 3: Based on static road structure information and historical motion information, and combined with artificial potential field theory, construct a risk field model, including a road risk field model and an obstacle vehicle risk field model, and construct a comprehensive risk field model by weighted superposition.

[0009] Step 4: Input the vehicle state prediction information output by the vehicle state prediction model in Step 2 into the obstacle vehicle risk field model to obtain the obstacle vehicle risk field for the future time period; input the static road structure information into the road risk field model to obtain the global road risk field; the obstacle vehicle risk field for the future time period and the global road risk field together constitute the future comprehensive risk field.

[0010] Step 5: Generate candidate trajectory clusters for autonomous vehicles using a fifth-order polynomial. Combine the spatiotemporal distribution of the future comprehensive risk field to evaluate the cumulative risk of each candidate trajectory in the candidate trajectory cluster over the entire future period, and select a set of low-risk trajectories that meet the safety threshold. Introduce a comfort evaluation function to further screen and evaluate the low-risk trajectory set to obtain the optimal planning trajectory that is both safe and comfortable.

[0011] As a preferred embodiment of the present invention, in step 1, the historical motion information includes the trajectory position, speed, acceleration, and heading angle of the obstacle vehicle and its surrounding vehicles, and the relative distance and relative speed between the obstacle vehicle and its surrounding vehicles are calculated to form vehicle interaction features; subsequently, the trajectory coordinates, speed, acceleration, heading angle, and vehicle interaction features are standardized to eliminate dimensional differences; the standardized historical motion information and vehicle interaction features constitute the spatiotemporal feature representation of the obstacle vehicle and its surrounding vehicles, and serve as the input features of the vehicle state prediction model; in addition, the obtained static road structure information includes the location of the road boundary line and the lane line, providing basic data for the subsequent construction of the road risk field model.

[0012] As a preferred embodiment of the present invention, in step 2, the input features of the vehicle state prediction model are derived from step 1. The prediction of the future motion state of the obstacle vehicle includes the lateral trajectory position, longitudinal trajectory position, longitudinal velocity, longitudinal acceleration, and heading angle motion state of the obstacle vehicle, which serves as the prediction task. In the input layer of the vehicle state prediction model, the input features are embedded and positionally encoded. The trajectory information and vehicle interaction features are modeled by the Transformer encoder of the shared layer of the vehicle state prediction model, and the feature matrix reflecting the temporal relationship and multi-vehicle interaction relationship is extracted as the shared feature serving the prediction task. The prediction results of the obstacle vehicle's trajectory position, motion velocity, acceleration, and heading angle in the future time period are generated by the task-specific layer of the vehicle state prediction model, respectively, to obtain multi-dimensional vehicle state prediction information that is time-continuous and state-aligned.

[0013] As a preferred embodiment of the present invention, in step 2, the vehicle state prediction model adopts a multi-task learning framework, including an input layer, a shared layer, and a task-specific layer. The input layer performs embedding processing and position encoding, the shared layer is composed of a Transformer encoder, and there are 5 task-specific layers, each corresponding to a prediction task. Each task-specific layer uses a multilayer perceptron (MLP). The vehicle state prediction model uses 4 seconds of historical motion information as input to predict the vehicle trajectory position, speed, and acceleration information for the next 3 seconds. The vehicle state prediction model obtains multidimensional continuous state prediction information of the obstacle vehicle in the future time period. The multidimensional continuous state prediction information is used for subsequent construction and evaluation of the future comprehensive risk field.

[0014] As a preferred embodiment of the present invention, in step 3, the road risk field model includes two components, namely the lane line risk field. and road boundary risk field The details are as follows:

[0015] ;

[0016] ;

[0017] ;

[0018] Road risk field for:

[0019] ;

[0020] In the formula, It is a lane line risk field. It is a risk field at the road boundary. It is the total number of lane dividing lines. It is the risk field coefficient of the lane dividing line. It is the safety distance coefficient for lane dividing lines. It is the vertical coordinate value of point A. It is the first Lane dividing line Axis position coordinates, It is from point A to the th Lane dividing line Distance of axis position, It is the risk field coefficient of the lane boundary line.

[0021] As a preferred embodiment of the present invention, in step 3, when constructing the obstacle vehicle risk field model, a corrected distance is introduced to characterize the changes in safety risk when surrounding vehicles approach the obstacle vehicle from different angles in a dynamic driving environment. By correcting the distance through coordinate rotation and velocity, a pseudo-distance of the vehicle is obtained, and an equivalent mass related to vehicle speed is introduced to characterize the risk intensity of the vehicle under dynamic driving conditions. In a two-dimensional Cartesian coordinate system, let... The axis represents the longitudinal direction of the road, corresponding to the horizontal direction in which vehicles travel along the road. The axis represents the transverse direction of the road, corresponding to the vertical direction of vehicle movement; this yields the obstacle vehicle risk field (described using the Yukawa potential function). Finally, the comprehensive risk field is obtained by weighted superposition of the road risk field and the obstacle vehicle risk field. ,

[0022] ;

[0023] ;

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] In the formula, It is the vehicle horizontal safety distance coefficient. It is the vehicle vertical safety distance coefficient. It is a distance correction, ( , ( ) represents the coordinates of the obstacle vehicle's center of mass in space. , ) is the original position ( , Corrected coordinates It is the current acceleration of the vehicle with the obstacle. It is the current speed of the vehicle blocking the obstacle. It is a safety factor related to speed. It is a safety factor used to adjust the intensity of risk. It is the heading angle of the obstacle vehicle in its current state. For equivalent quality, For actual quality, It is the clockwise angle formed by the line connecting any point around the obstacle vehicle to the vehicle's center of mass and the direction of the vehicle's motion. It is the first A risk area for vehicles with obstacles. It is the number of vehicles obstructing traffic. and These are the weights for the road risk field and the obstacle vehicle risk field, respectively. It is a road risk area. It is a natural constant.

[0029] As a preferred embodiment of the present invention, in step 4, the vehicle state prediction information includes trajectory coordinates, speed, acceleration, and heading angle; the obstacle vehicle risk field for the future time period is used to characterize the risk distribution of obstacle vehicles in the future time period. Since the lateral positions of the road boundary lines and lane lines remain unchanged in the future time period, the global road risk field distribution is constant in time and is used as global information to be shared by all vehicles. The obstacle vehicle risk field for the future time period is combined with the global road risk field to obtain the future comprehensive risk field.

[0030] As a preferred embodiment of the present invention, in step 5, trajectory planning is performed in a two-dimensional Cartesian coordinate system with the longitudinal direction of the road as the transverse axis and the direction perpendicular to the longitudinal direction of the road as the longitudinal axis. The vehicle's driving direction is denoted as... The longitudinal position of the vehicle, perpendicular to the direction of travel, is denoted as... Corresponding to the vehicle's lateral position, the 3D trajectory planning is decomposed into the process where the vehicle's lateral position changes as its longitudinal position changes. Curve, using Indicates, and vertical position As time goes by changing Thread, using It means that, with Discrete sampling of lateral positions at intervals generates candidate trajectory clusters covering various driving behaviors.

[0031] As a preferred embodiment of the present invention, in step 5, trajectory planning is performed with the road longitudinal direction as the reference point. Axis, perpendicular to the longitudinal direction of the road The operation is performed in a two-dimensional Cartesian coordinate system, where, Indicates the longitudinal position of the vehicle along the direction of travel. This represents the lateral position of the vehicle relative to the road centerline. Corresponding to the vehicle's lateral position, through coordinate transformation, the 3D trajectory planning problem is decomposed into two independent 1D planning problems: one where the lateral position changes with the longitudinal position... Curve, using The second is the change in vertical position over time. Curve, using It means that, with Discrete sampling of lateral position for step size generates a candidate trajectory cluster covering various driving behaviors;

[0032] ;

[0033] ;

[0034] In the formula, ~ and ~ These are the coefficients of a fifth-order polynomial, used to describe the lateral and longitudinal motion characteristics of the vehicle. The lower-order coefficients reflect the position and its changing trend, while the higher-order coefficients are used to ensure the smoothness of the trajectory within the planned range.

[0035] Based on the future comprehensive risk field, the candidate trajectory cluster is risk-assessed, and low-risk paths in the low-risk trajectory set that meet the safety threshold are selected. The paths are then evaluated through a comfort evaluation function to finally obtain the optimal planned trajectory for the autonomous vehicle.

[0036] Based on the spatiotemporal distribution of the future comprehensive risk field, risk assessment is performed on the generated candidate trajectory clusters. The comprehensive risk field value corresponding to each sampling point along the driving path of each candidate trajectory is statistically calculated and accumulated as the risk assessment index of the candidate trajectory.

[0037] Based on risk assessment indicators, a set of low-risk candidate trajectories whose risk assessment values ​​meet the safety threshold is selected. A comfort assessment function is then introduced for secondary screening to finally determine the safe and comfortable optimal planned trajectory for the autonomous vehicle. The comfort assessment function is defined as follows:

[0038] ;

[0039] In the formula, It's about the comfort of the trajectory. , , It is the influence coefficient. It is a trajectory segment. It is the curvature corresponding to the time trajectory segment. It is the rate of change of curvature corresponding to the trajectory segment. and These are the start and end positions of the trajectory planning cycle. It represents a small increment in arc length along the longitudinal direction of the road.

[0040] The beneficial effects of this invention are as follows:

[0041] 1. This invention addresses the problems of existing path planning methods in complex traffic environments, such as insufficient characterization of the uncertainties of dynamic traffic participants, inadequate forward-looking risk assessment, and the difficulty in balancing real-time performance and feasibility. It effectively improves upon existing technologies by introducing a technical solution that combines vehicle state prediction with risk field modeling. Specifically, the future motion state of obstacle vehicles is predicted before path planning, and the prediction results are dynamically integrated into the risk field modeling. This enables forward-looking identification of risks in the driving environment, overcoming the limitation of traditional risk field methods that rely solely on current state information.

[0042] 2. In the problem of vehicle state prediction, this invention uses a vehicle state prediction model based on Transformer and multi-task learning to comprehensively predict the motion state of the vehicle. While ensuring the effectiveness of the prediction, it reduces computational redundancy, which is beneficial to meeting the real-time requirements of autonomous driving systems.

[0043] 3. In the path planning problem, this invention combines a candidate trajectory generation mechanism with a hierarchical evaluation and screening mechanism based on the future comprehensive risk field and comfort function. While avoiding local extrema, it achieves layer-by-layer optimization of trajectory safety and ride comfort, thereby obtaining a safer, more stable and feasible autonomous driving planning trajectory in complex traffic environments. Attached Figure Description

[0044] Other objects and results of the invention will become more apparent and readily understood with reference to the following description taken in conjunction with the accompanying drawings. In the drawings:

[0045] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram of the vehicle state prediction model structure based on Transformer and multi-task learning framework according to an embodiment of the present invention.

[0047] Figure 3 This is a schematic diagram illustrating the road risk field effect according to an embodiment of the present invention;

[0048] Figure 4 This is a schematic diagram illustrating the vehicle risk field effect according to an embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram illustrating the overall risk field effect of an embodiment of the present invention;

[0050] Figure 6 This is a schematic diagram of a candidate trajectory cluster generated based on a fifth-order polynomial according to an embodiment of the present invention;

[0051] Figure 7 This is a schematic diagram of the trajectory planning method according to an embodiment of the present invention. Detailed Implementation

[0052] See Figure 1-7 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] This embodiment focuses on autonomous vehicles in a structured road environment. The motion information of the autonomous vehicle, the obstacle vehicle and the surrounding vehicles, as well as the static road structure information, can be obtained through on-board sensors, high-precision maps, vehicle-road cooperative systems or other existing technologies. This embodiment does not limit these methods.

[0054] This invention provides a method for autonomous driving risk perception and trajectory planning that integrates vehicle state prediction, comprising the following steps:

[0055] Step 1: Collect historical motion information and static road structure information of the obstacle vehicle and its surrounding vehicles. Perform unified vectorization processing on the acquired historical motion information and static road structure information. Further standardize and preprocess the historical motion information to obtain the spatiotemporal feature representation of the obstacle vehicle and its surrounding vehicles.

[0056] The historical motion information includes the trajectory position, speed, acceleration, and heading angle of the obstacle vehicle and its surrounding vehicles. The relative distance and relative speed between the obstacle vehicle and its surrounding vehicles are calculated to form vehicle interaction features. Subsequently, the trajectory coordinates, speed, acceleration, heading angle, and vehicle interaction features are standardized to eliminate dimensional differences. The standardized historical motion information and vehicle interaction features constitute the spatiotemporal feature representation of the obstacle vehicle and its surrounding vehicles, and serve as input features for the vehicle state prediction model. In addition, the acquired static road structure information includes the location of road boundary lines and lane lines, providing basic data for the subsequent construction of the road risk field model.

[0057] Step 2: Combine Transformer (neural network model) and multi-task learning framework to build a vehicle state prediction model. Use the spatiotemporal features of the obstacle vehicle and its surrounding vehicles as the input features of the vehicle state prediction model to predict the future motion state of the obstacle vehicle.

[0058] The input features of the vehicle state prediction model come from step 1. Predicting the future motion state of the obstacle vehicle includes its lateral trajectory position, longitudinal trajectory position, longitudinal velocity, longitudinal acceleration, and heading angle, which serves as the prediction task. In the input layer of the vehicle state prediction model, the input features are embedded and positionally encoded. The trajectory information and vehicle interaction features are modeled using the Transformer encoder in the shared layer of the vehicle state prediction model. A feature matrix reflecting temporal relationships and multi-vehicle interaction relationships is extracted as shared features for the prediction task. The task-specific layers of the vehicle state prediction model generate prediction results for the obstacle vehicle's trajectory position, velocity, acceleration, and heading angle in future time periods, obtaining temporally continuous and state-aligned multidimensional vehicle state prediction information. This provides forward-looking input for the subsequent construction of the future risk field.

[0059] The vehicle state prediction model adopts a multi-task learning framework, including an input layer, a shared layer, and a task-specific layer. The input layer performs embedding processing and position encoding, the shared layer is composed of a Transformer encoder, and there are 5 task-specific layers, each corresponding to a prediction task. Each task-specific layer uses a multilayer perceptron (MLP). The vehicle state prediction model uses 4 seconds of historical motion information as input to predict the vehicle's trajectory position, velocity, and acceleration information for the next 3 seconds. The model obtains multi-dimensional continuous state prediction information of the obstacle vehicle in the future time period, which is used for subsequent construction and evaluation of the future comprehensive risk field.

[0060] See Figure 2 First, in the input feature construction stage, multivariate temporal input features are extracted and constructed based on the historical motion information of obstacle vehicles (OV) and surrounding vehicles (SV) in the traffic scene. This stage comprehensively considers the motion state information of vehicles such as trajectory position, speed, acceleration, and heading angle, while introducing the interaction features of surrounding vehicles to characterize the dynamic evolution process of vehicles in complex traffic environments, thereby forming an input feature representation with spatiotemporal correlation, namely, spatiotemporal feature representation.

[0061] Secondly, in the input layer stage, the constructed spatiotemporal features are input into the vehicle state prediction model for preprocessing. Specifically, firstly, embedding processing is used to map the original features in different dimensions of spatiotemporal features to a unified high-dimensional feature space to enhance feature expressive power; then, positional encoding is introduced to provide temporal position information for the feature sequence, enabling the vehicle state prediction model to perceive the sequential relationship between each time step in the sequence, thereby laying the foundation for subsequent modeling of the comprehensive risk field.

[0062] Building upon this foundation, the model enters the shared layer stage, where the input features are uniformly modeled based on the Transformer encoder. This stage utilizes the Transformer encoder's multi-head self-attention mechanism to globally model the dependencies between different time steps, enabling features at each time step to adaptively aggregate historical information. Simultaneously, a feedforward neural network is combined to perform nonlinear transformations on the features, and residual connections and layer normalization are used to improve the stability and expressive power of the model training. Through multi-layer stacking, the model can progressively extract higher-level temporal features and interaction information between variables, obtaining a shared feature representation with strong discriminative capabilities.

[0063] Finally, at the task-specific layer stage, based on the high-level features output from the shared layer, multiple parallel task branches are introduced to predict different vehicle state variables. Each branch uses an independent multilayer perceptron (MLP) to analyze the trajectory's horizontal coordinate. , trajectory ordinate ,speed acceleration and heading angle The modeling and output process enables joint prediction of the vehicle's future trajectory and motion state. This stage achieves task decoupling based on shared features, effectively improving the model's expressive power and prediction accuracy across different prediction targets.

[0064] Step 3: Based on static road structure information and historical motion information, and combined with artificial potential field theory, construct a risk field model, including a road risk field model and an obstacle vehicle risk field model, and construct a comprehensive risk field model by weighted superposition.

[0065] The road risk field model contains two components: lane line risk field and lane line risk field. and road boundary risk field The details are as follows:

[0066] ;

[0067] ;

[0068] ;

[0069] Road risk field for:

[0070] ;

[0071] In the formula, It is a lane line risk field. It is a risk field at the road boundary. It is the total number of lane dividing lines. It is the risk field coefficient of the lane dividing line. It is the safety distance coefficient for lane dividing lines. These are the vertical coordinates of point A. It is the first Lane dividing line Axis position coordinates, It is from point A to the... Lane dividing line Distance of axis position, It is the risk field coefficient of the lane boundary line.

[0072] When constructing the obstacle vehicle risk field model, a corrected distance is introduced to characterize the changes in safety risk when surrounding vehicles approach the obstacle vehicle from different angles in a dynamic driving environment. By correcting the distance through coordinate rotation and velocity, a pseudo-distance for the vehicle is obtained. An equivalent mass related to vehicle speed is introduced to characterize the risk intensity of the vehicle under dynamic driving conditions. In a two-dimensional Cartesian coordinate system, let... The axis represents the longitudinal direction of the road, corresponding to the horizontal direction in which the vehicle travels along the road; that is, the horizontal coordinate of the trajectory. The axis represents the horizontal direction of the road, corresponding to the vertical direction of vehicle movement, i.e., the longitudinal coordinate of the trajectory; this yields the obstacle vehicle risk field (described based on the Yukawa potential function form). Finally, the comprehensive risk field is obtained by weighted superposition of the road risk field and the obstacle vehicle risk field. ,

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] In the formula, It is the vehicle horizontal safety distance coefficient. It is the vehicle vertical safety distance coefficient. It is a distance correction, ( , ( ) represents the coordinates of the obstacle vehicle's center of mass in space. , ) is the original position ( , Corrected coordinates of ) It is the current acceleration of the vehicle in the obstacle. It is the current speed of the vehicle in the obstacle. It is a safety factor related to speed. It is a safety factor used to adjust the intensity of risk. It is the heading angle of the obstacle vehicle in its current state. For equivalent quality, For actual quality, It is the clockwise angle formed by the line connecting any point around the obstacle vehicle to the vehicle's center of mass and the direction of the vehicle's motion. It is the first A risk area for vehicles with obstacles. It is the number of vehicles obstructing traffic. and These are the weights for the road risk field and the obstacle vehicle risk field, respectively. It is a road risk area. It is a natural constant.

[0080] Step 4: Input the vehicle state prediction information output by the vehicle state prediction model in Step 2 into the obstacle vehicle risk field model to obtain the obstacle vehicle risk field for the future time period; input the static road structure information into the road risk field model to obtain the global road risk field; the obstacle vehicle risk field for the future time period and the global road risk field together constitute the future comprehensive risk field.

[0081] The vehicle state prediction information includes trajectory coordinates, speed, acceleration, and heading angle; the obstacle vehicle risk field for the future time period is used to characterize the risk distribution of obstacle vehicles in the future time period. Since the lateral position of the road boundary line and lane line remains unchanged in the future time period, the global road risk field distribution is constant in time and is used as global information to be shared by all vehicles. The obstacle vehicle risk field for the future time period is combined with the global road risk field to obtain the future comprehensive risk field.

[0082] See Figure 3 From a spatial perspective, the horizontal axis represents the vehicle's lateral position, and the vertical axis represents its longitudinal position. The vertical direction and color gradient together characterize the risk value (gradually increasing from blue to red). It can be observed that the risk value increases significantly near the road boundaries, exhibiting a steep upward trend, reflecting the strong constraint of the road boundaries on vehicle movement; that is, the closer a vehicle is to the road edge, the higher its potential risk. Meanwhile, within the lane, the overall risk level is lower and the distribution is relatively flat, but slight undulations are still visible in the lateral direction. This is mainly due to the influence of the lane line risk field, used to characterize the risk change trend away from the lane center, and also reflects the soft constraint effect of the lane lines on vehicle movement. Overall, the risk field forms a spatial distribution pattern of "low in the middle and high on both sides," reflecting the rigid constraint characteristics of the road boundaries and the guiding role of lane lines in the lateral stability of vehicles.

[0083] See Figure 4The horizontal and vertical axes represent the longitudinal and lateral positions of the road, respectively, with colors ranging from blue to red indicating a gradual increase in risk value. It can be seen that the risk value reaches its highest point near the location of the obstacle vehicle, exhibiting a potential field distribution pattern that gradually decreases from the center outwards. Simultaneously, the vehicle risk field displays significant directional and asymmetric characteristics spatially, meaning that the risk spreads more extensively and has a more significant impact along the vehicle's direction of movement and the direction of potential conflict. This distribution characteristic not only effectively characterizes the local high-risk area around the obstacle vehicle but also reflects, to some extent, the dynamic impact of the obstacle vehicle's current movement on the surrounding spatial risk distribution.

[0084] Step 5: Generate candidate trajectory clusters for autonomous vehicles using a fifth-order polynomial. Combine the spatiotemporal distribution of the future comprehensive risk field to evaluate the cumulative risk of each candidate trajectory in the candidate trajectory cluster over the entire future period, and select a set of low-risk trajectories that meet the safety threshold. Introduce a comfort evaluation function to further screen and evaluate the low-risk trajectory set to obtain the optimal planning trajectory that is both safe and comfortable.

[0085] Trajectory planning is performed in a two-dimensional Cartesian coordinate system with the longitudinal direction of the road as the transverse axis and the direction perpendicular to the longitudinal direction of the road as the longitudinal axis. The vehicle's direction of travel is denoted as... The longitudinal position of the vehicle, perpendicular to the direction of travel, is denoted as... Corresponding to the vehicle's lateral position, the 3D trajectory planning is decomposed into the process where the vehicle's lateral position changes as its longitudinal position changes. Curve, using Indicates, and vertical position As time goes by changing Thread, using It means that, with To generate candidate trajectory clusters covering various driving behaviors, discrete sampling of lateral positions is performed at intervals. Trajectory planning is then performed along the road longitudinal direction. Axis, perpendicular to the longitudinal direction of the road The operation is performed in a two-dimensional Cartesian coordinate system, where, Indicates the longitudinal position of the vehicle along the direction of travel. This represents the lateral position of the vehicle relative to the road centerline. Corresponding to the vehicle's lateral position, through coordinate transformation, the 3D trajectory planning problem is decomposed into two independent 1D planning problems: one where the lateral position changes with the longitudinal position... Curve, using The second is the change in vertical position over time. Curve, using It means that, with Discrete sampling of lateral position for step size generates a candidate trajectory cluster covering various driving behaviors;

[0086] ;

[0087] ;

[0088] In the formula, ~ and ~ These are the coefficients of a fifth-order polynomial, used to describe the lateral and longitudinal motion characteristics of the vehicle. The lower-order coefficients reflect the position and its changing trend, while the higher-order coefficients are used to ensure the smoothness of the trajectory within the planned range.

[0089] Based on the future comprehensive risk field, the candidate trajectory cluster is risk-assessed, and low-risk paths in the low-risk trajectory set that meet the safety threshold are selected. The paths are then evaluated through a comfort evaluation function to finally obtain the optimal planned trajectory for the autonomous vehicle.

[0090] Based on the spatiotemporal distribution of the future comprehensive risk field, risk assessment is performed on the generated candidate trajectory clusters. The comprehensive risk field value corresponding to each sampling point along the driving path of each candidate trajectory is statistically calculated and accumulated as the risk assessment index of the candidate trajectory.

[0091] Based on risk assessment indicators, a set of low-risk candidate trajectories whose risk assessment values ​​meet the safety threshold is selected. A comfort assessment function is then introduced for secondary screening to finally determine the safe and comfortable optimal planned trajectory for the autonomous vehicle. The comfort assessment function is defined as follows:

[0092] ;

[0093] In the formula, It's about the comfort of the trajectory. , , It is the influence coefficient. It is a trajectory segment. It is the curvature corresponding to the time trajectory segment. It is the rate of change of curvature corresponding to the trajectory segment. and These are the start and end positions of the trajectory planning cycle. It represents a small increment in arc length along the longitudinal direction of the road.

[0094] See Figure 5The horizontal and vertical axes represent the lateral and longitudinal positions of vehicles on the road, respectively. The vertical direction and color changes collectively characterize the risk value (gradually increasing from blue to red). It can be observed that the road as a whole exhibits a basic distribution trend with the center of the lanes being low-risk and the side boundaries being high-risk. This is mainly due to the constraint effect of the risk field at the road boundaries. Based on this, multiple local high-risk "protrusions" distributed along the longitudinal direction of the road correspond to the positions of different obstacle vehicles. Their risk peaks significantly increase within the neighborhood of each vehicle and gradually decrease towards the surrounding space. Simultaneously, the risk fields of each vehicle spatially overlap and couple, resulting in a risk superposition effect in some areas, forming a more complex risk distribution pattern. This comprehensive risk field not only reflects the global constraints of the road structure on vehicle movement but also characterizes the local dynamic risk features under the combined action of multiple obstacle vehicles.

[0095] See Figure 6 The horizontal axis represents the longitudinal position, and the vertical axis represents the lateral position. It can be seen that each candidate trajectory starts from the same initial state and gradually evolves smoothly towards different lateral target positions as the longitudinal distance increases, forming multiple continuous and smooth candidate paths covering straight lines and different degrees of lateral offset.

[0096] See Figure 7 First, multiple candidate trajectories, or candidate trajectory clusters, are generated based on the current state of the autonomous vehicle. Each trajectory corresponds to its possible driving path within the next 3 seconds. Simultaneously, the future comprehensive risk field is calculated by inputting the future motion state of obstacle vehicles, output from the vehicle state prediction model, into the risk field model. Then, each candidate trajectory of the autonomous vehicle is matched with the future comprehensive risk field and a risk assessment is performed. The risk value of each trajectory in the area it traverses within different time periods is calculated, and its cumulative risk over the entire prediction time domain is comprehensively assessed, thereby eliminating high-risk trajectories and obtaining a set of low-risk trajectories. Based on this, a comfort evaluation function is further used to filter the low-risk trajectories, ultimately obtaining a candidate trajectory that balances both safety and comfort as the optimal planning trajectory for the autonomous vehicle.

[0097] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for autonomous driving risk perception and trajectory planning that integrates vehicle state prediction, characterized in that, Includes the following steps: Step 1: Collect historical motion information and static road structure information of the obstacle vehicle and its surrounding vehicles, perform unified vectorization processing, and further standardize and preprocess based on the processed historical motion information to obtain the spatiotemporal feature representation of the obstacle vehicle and its surrounding vehicles. Step 2: Combine Transformer and multi-task learning framework to build a vehicle state prediction model. Use the spatiotemporal features of the obstacle vehicle and its surrounding vehicles as the input features of the vehicle state prediction model to predict the future motion state of the obstacle vehicle. Predicting the future motion state of an obstacle vehicle includes its lateral trajectory position, longitudinal trajectory position, longitudinal velocity, longitudinal acceleration, and heading angle, which serves as the prediction task. The input features are embedded and positionally encoded in the input layer of the vehicle state prediction model. The trajectory information and vehicle interaction features are modeled using the Transformer encoder in the shared layer of the vehicle state prediction model, extracting a feature matrix reflecting temporal relationships and multi-vehicle interaction relationships. This matrix serves as the shared feature for the prediction task. The task-specific layers of the vehicle state prediction model then generate prediction results for the obstacle vehicle's trajectory position, velocity, acceleration, and heading angle in future time periods, obtaining temporally continuous and state-aligned multidimensional vehicle state prediction information. Step 3: Based on static road structure information and historical motion information, and combined with artificial potential field theory, construct a risk field model, including a road risk field model and an obstacle vehicle risk field model, and construct a comprehensive risk field model by weighted superposition. Step 4: Input the vehicle state prediction information output by the vehicle state prediction model into the obstacle vehicle risk field model to obtain the obstacle vehicle risk field for the future time period; input the static road structure information into the road risk field model to obtain the global road risk field; the obstacle vehicle risk field for the future time period and the global road risk field together constitute the future comprehensive risk field. Step 5: Generate candidate trajectory clusters for autonomous vehicles using a fifth-order polynomial. Combine the spatiotemporal distribution of the future comprehensive risk field to evaluate the cumulative risk of each candidate trajectory in the candidate trajectory cluster over the entire future period, and select a set of low-risk trajectories that meet the safety threshold. Furthermore, a comfort evaluation function is introduced to further screen and evaluate the low-risk trajectory set to obtain the optimal planned trajectory.

2. The autonomous driving risk perception and trajectory planning method integrating vehicle state prediction according to claim 1, characterized in that, In step 1, the historical motion information includes the trajectory position, speed, acceleration, and heading angle of the obstacle vehicle and its surrounding vehicles. The relative distance and relative speed between the obstacle vehicle and its surrounding vehicles are calculated to form vehicle interaction features. Subsequently, the trajectory coordinates, speed, acceleration, heading angle, and vehicle interaction features are standardized to eliminate dimensional differences. The standardized historical motion information and vehicle interaction features constitute the spatiotemporal feature representation of the obstacle vehicle and its surrounding vehicles and serve as input features for the vehicle state prediction model. In addition, the acquired static road structure information includes the location of road boundary lines and lane lines, providing basic data for the subsequent construction of the road risk field model.

3. The autonomous driving risk perception and trajectory planning method integrating vehicle state prediction according to claim 1, characterized in that, In step 2, the vehicle state prediction model adopts a multi-task learning framework, including an input layer, a shared layer, and a task-specific layer. The input layer performs embedding processing and position encoding, the shared layer is composed of a Transformer encoder, and there are 5 task-specific layers, each corresponding to a prediction task. Each task-specific layer uses a multilayer perceptron. The vehicle state prediction model uses 4 seconds of historical motion information as input to predict the vehicle trajectory position, speed, and acceleration information for the next 3 seconds. The vehicle state prediction model obtains multi-dimensional continuous state prediction information of the obstacle vehicle in the future time period. This multi-dimensional continuous state prediction information is used for subsequent construction and evaluation of the future comprehensive risk field.

4. The autonomous driving risk perception and trajectory planning method integrating vehicle state prediction according to claim 1, characterized in that, In step 3, the road risk field model includes two components: the lane line risk field. and road boundary risk field The details are as follows: ； ； ； Road risk field for: ； In the formula, It is a lane line risk field. It is a risk field at the road boundary. It is the total number of lane dividing lines. It is the risk field coefficient of the lane dividing line. It is the safety distance coefficient for lane dividing lines. These are the vertical coordinates of point A. It is the first Lane dividing line Axis position coordinates, It is from point A to the... Lane dividing line Distance of axis position, It is the risk field coefficient of the lane boundary line.

5. The autonomous driving risk perception and trajectory planning method integrating vehicle state prediction according to claim 1, characterized in that, In step 3, when constructing the obstacle vehicle risk field model, a corrected distance is introduced to characterize the changes in safety risk when surrounding vehicles approach the obstacle vehicle from different angles in a dynamic driving environment. By correcting the distance through coordinate rotation and velocity, a pseudo-distance for the vehicle is obtained. An equivalent mass related to vehicle speed is introduced to characterize the risk intensity of the vehicle under dynamic driving conditions. In a two-dimensional Cartesian coordinate system, let... The axis represents the longitudinal direction of the road, corresponding to the horizontal direction in which vehicles travel along the road. The axis is the horizontal direction of the road, corresponding to the vertical direction of vehicle movement; The obstacle vehicle risk field is obtained, and finally, the comprehensive risk field is obtained by weighted superposition of the road risk field and the obstacle vehicle risk field. , ； ； ； ； ； ； In the formula, It is the vehicle horizontal safety distance coefficient. It is the vehicle vertical safety distance coefficient. It is a distance correction, ( , ( ) represents the coordinates of the obstacle vehicle's center of mass in space. , ) is the original position ( , Corrected coordinates of ) It is the current acceleration of the vehicle in the obstacle. It is the current speed of the vehicle in the obstacle. It is a safety factor related to speed. It is a safety factor used to adjust the intensity of risk. It is the heading angle of the obstacle vehicle in its current state. For equivalent quality, For actual quality, It is the clockwise angle formed by the line connecting any point around the obstacle vehicle to the vehicle's center of mass and the direction of the vehicle's motion. It is the first A risk area for vehicles with obstacles. It is the number of vehicles obstructing traffic. and These are the weights for the road risk field and the obstacle vehicle risk field, respectively. It is a road risk area. It is a natural constant.

6. The autonomous driving risk perception and trajectory planning method integrating vehicle state prediction according to claim 1, characterized in that, In step 4, the vehicle state prediction information includes trajectory coordinates, speed, acceleration, and heading angle; the obstacle vehicle risk field for the future time period is used to characterize the risk distribution of obstacle vehicles in the future time period. Since the lateral position of the road boundary line and lane line remains unchanged in the future time period, the global road risk field distribution is constant in time and is used as global information to be shared by all vehicles. The obstacle vehicle risk field for the future time period is combined with the global road risk field to obtain the future comprehensive risk field.

7. The autonomous driving risk perception and trajectory planning method integrating vehicle state prediction according to claim 1, characterized in that, In step 5, trajectory planning is performed in a two-dimensional Cartesian coordinate system with the longitudinal direction of the road as the lateral axis and the direction perpendicular to the longitudinal direction of the road as the longitudinal axis. The vehicle's travel direction is denoted as... The longitudinal position of the vehicle, perpendicular to the direction of travel, is denoted as... Corresponding to the vehicle's lateral position, the 3D trajectory planning is decomposed into the process where the vehicle's lateral position changes as its longitudinal position changes. Curve, using Indicates, and vertical position As time goes by changing Thread, using It means that, with Discrete sampling of lateral positions at intervals generates candidate trajectory clusters covering various driving behaviors.

8. The autonomous driving risk perception and trajectory planning method integrating vehicle state prediction according to claim 7, characterized in that, In step 5, trajectory planning is performed with the road longitudinal direction as the reference point. Axis, perpendicular to the longitudinal direction of the road The operation is performed in a two-dimensional Cartesian coordinate system, where, Indicates the longitudinal position of the vehicle along the direction of travel. This represents the lateral position of the vehicle relative to the road centerline. Corresponding to the vehicle's lateral position, through coordinate transformation, the 3D trajectory planning problem is decomposed into two independent 1D planning problems: one where the lateral position changes with the longitudinal position... Curve, using The second is the change in vertical position over time. Curve, using It means that, with Discrete sampling of lateral position for step size generates a candidate trajectory cluster covering various driving behaviors; ； ； In the formula, ~ and ~ These are the coefficients of a fifth-order polynomial, used to describe the lateral and longitudinal motion characteristics of the vehicle. The lower-order coefficients reflect the position and its changing trend, while the higher-order coefficients are used to ensure the smoothness of the trajectory within the planned range. Based on the future comprehensive risk field, the candidate trajectory cluster is risk-assessed, and low-risk paths in the low-risk trajectory set that meet the safety threshold are selected. The paths are then evaluated through a comfort evaluation function to finally obtain the optimal planned trajectory for the autonomous vehicle. Based on the spatiotemporal distribution of the future comprehensive risk field, risk assessment is performed on the generated candidate trajectory clusters. The comprehensive risk field value corresponding to each sampling point along the driving path of each candidate trajectory is statistically calculated and accumulated as the risk assessment index of the candidate trajectory. Based on risk assessment indicators, a set of low-risk candidate trajectories whose risk assessment values ​​meet the safety threshold is selected. A comfort assessment function is then introduced for secondary screening to finally determine the optimal planned trajectory for the autonomous vehicle. The comfort assessment function is defined as follows: ； In the formula, It's about the comfort of the trajectory. , , It is the influence coefficient. It is a trajectory segment. It is the curvature corresponding to the time trajectory segment. It is the rate of change of curvature corresponding to the trajectory segment. and These are the start and end positions of the trajectory planning cycle. It represents a small increment in arc length along the longitudinal direction of the road.

Images