Unstructured road trajectory prediction method fusing topological skeleton and interactive multi-model

Abstract

The application discloses a non-structured road track prediction method fusing a topological skeleton and interactive multi-models, and realizes high-precision, multi-modal and smooth and reliable prediction capability of a vehicle future track in a non-structured road scene without lane lines, with multiple obstacles and variable structures through four core steps of environment topological extraction, multi-model state estimation, intention multi-modal prediction and dynamic constraint track generation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation technology, specifically relating to a method for predicting unstructured road trajectories by integrating topological skeletons and interactive multi-models. Background Technology

[0002] In autonomous driving and assisted driving technologies, accurate prediction of the future trajectories of surrounding traffic participants (such as vehicles and pedestrians) is a prerequisite for obstacle avoidance and path generation by the decision-making and planning modules. In structured scenarios such as urban arterial roads or highways, existing technologies typically rely heavily on lane centerlines provided by high-resolution maps (HD Maps) as strong prior constraints, and relatively accurate long-term predictions can be achieved using lane-keeping assumptions. However, in unstructured environments such as open-pit mines, port yards, rural roads, or unmarked plazas, there are no clear lane markings or high-resolution map data. In these situations, vehicle movement exhibits a high degree of freedom and is often accompanied by complex variable curvature maneuvers.

[0003] For such scenarios, existing technologies mainly fall into two categories: deep learning-based methods and kinematic model-based extrapolation methods. The first method relies on massive amounts of historical trajectory data to train neural networks (such as LSTM and Transformer), but data acquisition is difficult in specific closed scenarios (such as mining areas), and the model's generalization ability and interpretability are poor. The second method, which is currently the mainstream in engineering applications, typically uses constant velocity (CV), constant turning rate (CTRV), or constant acceleration (CA) models to recursively extrapolate vehicle states. While this method has low computational cost and high real-time performance, it suffers from significant drawbacks in complex unstructured road conditions, including the inability of a single model to adapt to dynamic maneuvers, the "tail-drifting" phenomenon in long-term predictions of high-order models, and the lack of environmental topology guidance. Summary of the Invention

[0004] In view of this, the present invention proposes an unstructured road trajectory prediction method that integrates topological skeleton and interactive multi-model. Through four core steps, namely environmental topology extraction, multi-model state estimation, intention multi-modal prediction, and dynamic constraint trajectory generation, it achieves high-precision, multi-modal, smooth and reliable prediction of vehicle future trajectories in unstructured road scenarios with no lane lines, multiple obstacles, and variable structures.

[0005] To achieve the above objectives, this invention provides a method for predicting unstructured road trajectories that integrates topological skeletons and interactive multi-model approaches, comprising: S1. Extract the topological skeleton lines of unstructured roads based on environmental perception data; S2. An interactive multi-model extended Kalman filter is used to estimate and fuse the motion state of the target vehicle in parallel, and the output is a full state vector containing position, velocity, heading, acceleration and angular velocity. S3. Combining the topology skeleton line and the full state vector, multiple discrete intention target points are predicted through dynamic sector search and intent scoring mechanisms. S4. For each of the intended target points, generate a fifth-order Bézier curve trajectory based on asymmetric dynamic constraints.

[0006] Preferably, step S1 specifically includes: Acquire environmental perception data, and based on the environmental perception data, characterize the unstructured environment into a binary occupancy grid map. The state of each pixel in the binary occupancy grid map is represented by obstacle occupancy and free space, respectively. For each free space pixel in a binary occupied raster map, calculate its Euclidean distance to the nearest obstacle pixel to characterize its safety margin, and construct an Euclidean distance symbol field. A morphological thinning algorithm is performed on the Euclidean distance symbol field to iteratively erode the boundary pixels of the free region until the free region is thinned into a set of connected lines with a width of one pixel. This set of lines is used as an approximation of the generalized Vinograph on the discrete raster map and serves as the topological skeleton line. Mapping the topological skeleton lines in the grid coordinates to the world coordinate system where the vehicle is located yields the physical topological skeleton point set, which represents the traversable topological structure of the environment.

[0007] Preferably, the interactive multi-model extended Kalman filter includes a constant velocity model, a constant acceleration model, and a constant angular acceleration variable curvature model. All models are described using discrete-time state transition equations, wherein: The constant speed model is used to describe the state of a vehicle when it is traveling straight or in steady-state cruise, and to determine the vehicle's motion state with respect to position, speed, and heading. The constant acceleration model is used to describe the state of a vehicle during longitudinal acceleration or emergency braking and to determine the vehicle's motion state with respect to longitudinal acceleration. The constant angular acceleration variable curvature model is used to describe the state of a vehicle when it is making a sharp turn, U-turn, or S-shaped obstacle avoidance, and to determine the vehicle's motion state with respect to angular velocity and angular acceleration.

[0008] Preferably, the parallel estimation and fusion of the motion state of the target vehicle includes: Model definition and initialization: Configure an extended Kalman filter for each of the constant velocity model, constant acceleration model, and constant angular acceleration variable curvature model; define the system observation vector as the observation vector directly provided by the sensor; and define the system full state vector used for the final fusion output as the motion state output of the three models. Model interaction: At each time step, based on the preset Markov model transition probability matrix, the mixture probability between the models at the previous time step is calculated. This is used to redistribute historical information according to the credibility of each model to prevent state jumps when switching models. Parallel filtering and observation update: Three extended Kalman filters work in parallel based on the current observations, each calculating its own observation residual. At each estimation point, the extended Kalman filter performs a first-order Taylor expansion of the nonlinear state transition function by calculating the Jacobian matrix to achieve local linearization of the system. Model probability update: The likelihood function is calculated based on the observation residuals of each model. The smaller the residual between the predicted value and the observed value of the model, the higher its likelihood and the greater the corresponding model probability, thereby enabling the system to automatically recognize the current driving intention. Weighted fusion output: The final output is the weighted sum of the three model state estimates as the fused state.

[0009] Preferably, step S3 includes: Based on the vehicle's current position, speed, and heading angle output in step S2, a dynamically changing forward search region is constructed. The search region is a fan-shaped area with the current position as the vertex and the current heading as the central axis. The search radius of the fan-shaped area is dynamically determined by the product of the current speed, the preset prediction time, and a distance redundancy coefficient greater than one. Its horizontal subtended angle is determined by a preset field of view half-angle threshold. Based on the topology skeleton line output in step S1, all points that are simultaneously located within the sector area and within the preset distance and angle thresholds are selected to form an initial candidate target point set. A comprehensive intent score is calculated for each point in the initial candidate target point set, wherein the intent score is a weighted combination of two parts: direction consistency score and distance matching score. The direction consistency score measures the degree of agreement between the direction of the candidate point relative to the vehicle's current position and the vehicle's current heading angle. The more consistent they are, the higher the score. The distance matching score assesses whether the distance between the candidate point and the vehicle's current position is within the theoretically achievable range based on the current speed and prediction time. The more the distance matches, the higher the score. Non-maximum suppression filtering is performed on the candidate point set after scoring. This filtering process is carried out iteratively: each time, the point with the highest intent score is selected from the current candidate point set as one of the valid intent target points and removed from the set. At the same time, all other candidate points within a certain distance around the point are removed to avoid redundant points that are too dense in space being repeatedly selected as different intents. Repeat the selection and removal process described above until the candidate point set is empty, or the preset limit for the number of target points has been reached. The final output consists of multiple discrete intention target points, representing different path branches that the vehicle may travel to in the future prediction period.

[0010] Preferably, step S4 includes: For each intention target point output in step S3, a future trajectory is generated smoothly from the current state of the vehicle to the target point using a fifth-order Bézier curve as the basic mathematical model. The fifth-order Bézier curve is fully defined by six control points. Calculate the first three control points of the Bézier curve using the fully fused state vector: The first control point is obtained by using the position coordinate components in the vehicle's fully fused state vector. The second control point is obtained by performing a scalar multiplication operation on the vehicle speed vector based on the total prediction time and the order characteristics of the Bézier curve, and then performing a vector addition operation with the first control point. The third control point is obtained by constructing a second-order derivative constraint equation for the Bézier curve with the vehicle's resultant acceleration vector as a known condition and solving the equation. Calculate the last three control points of the Bézier curve using the intended target point and the reduced-order motion assumption: The sixth control point is obtained by directly mapping the coordinates of the intended target point. By calculating the unit direction vector from the first control point to the sixth control point, and performing scalar and vector multiplication operations on it with the terminal predicted velocity value, the predicted total duration and the Bézier curve order parameter, an offset vector is generated. Then, a vector subtraction operation is performed on the coordinates of the sixth control point and the offset vector to obtain the fifth control point. The fourth control point is obtained by substituting the coordinates of the fifth and sixth control points into the linear smooth constraint relationship that characterizes the change of zero acceleration at the end point and performing a linear combination operation. After obtaining all six control points, discretization interpolation is performed using the Bézier curve formula, and the final output is a continuous, smooth future trajectory sequence that conforms to asymmetric dynamic constraints for each intended target point.

[0011] The present invention has achieved at least the following beneficial effects: 1. This invention, by combining topological skeleton lines with environmental perception data, and integrating dynamic sector search and intent scoring mechanisms, can accurately predict multiple discrete target points (such as different path branches at intersections) that a vehicle may travel to in the future. This improves the diversity and robustness of trajectory prediction on unstructured roads (without lane lines, multiple obstacles) and avoids the limitations of a single predicted path.

[0012] 2. This invention employs an interactive multi-model extended Kalman filter (IMM-EKF), fusing three motion models: constant velocity, constant acceleration, and constant angular acceleration with variable curvature, to identify vehicle motion patterns (such as straight-line driving, acceleration, and turning) in real time. The system can adaptively switch model weights to achieve accurate estimation of all states (position, velocity, heading, acceleration, and angular velocity), improving the tracking accuracy for complex driving behaviors.

[0013] 3. This invention generates a fifth-order Bézier curve trajectory for each intended target point based on asymmetric dynamic constraints, achieving a smooth transition of "high-order anchoring at the near end and low-order degradation at the far end." This ensures the accuracy of transient responses such as sharp turns while avoiding trajectory divergence (such as "fishtailing") caused by long-term integration, outputting a continuous and smooth trajectory that conforms to the vehicle's dynamic characteristics.

[0014] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0015] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration: Figure 1 This is a flowchart illustrating the steps of the unstructured road trajectory prediction method that integrates topological skeleton and interactive multi-model in an embodiment of the present invention. Detailed Implementation

[0016] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0017] To achieve the above objectives, the present invention provides the following technical solution: The unstructured road trajectory prediction method that integrates topological skeleton and interactive multi-model provided by this invention refers to... Figure 1 ,include: S1. Extract the topological skeleton lines of unstructured roads based on environmental perception data; S2. An interactive multi-model extended Kalman filter is used to estimate and fuse the motion state of the target vehicle in parallel, and the output is a full state vector containing position, velocity, heading, acceleration and angular velocity. S3. Combining the topology skeleton line and the full state vector, multiple discrete intention target points are predicted through dynamic sector search and intent scoring mechanisms. S4. For each of the intended target points, generate a fifth-order Bézier curve trajectory based on asymmetric dynamic constraints.

[0018] The working principle and beneficial effects of the above technical solution are as follows: By combining topological skeleton lines with environmental perception data, and integrating dynamic sector search and intent scoring mechanisms, this invention can accurately predict multiple discrete target points (such as different path branches at intersections) that a vehicle may travel to in the future. This improves the diversity and robustness of trajectory prediction in unstructured roads (without lane lines, multiple obstacles) and avoids the limitations of a single predicted path.

[0019] This invention employs an interactive multi-model extended Kalman filter (IMM-EKF) that integrates three motion models: constant velocity, constant acceleration, and constant angular acceleration with variable curvature, to identify vehicle motion patterns (such as straight-line driving, acceleration, and turning) in real time. The system can adaptively switch model weights to achieve accurate estimation of all states (position, velocity, heading, acceleration, and angular velocity), thus improving the tracking accuracy for complex driving behaviors.

[0020] This invention generates a fifth-order Bézier curve trajectory for each intended target point based on asymmetric dynamic constraints, achieving a smooth transition of "high-order anchoring at the near end and low-order degradation at the far end." This ensures the accuracy of transient responses such as sharp turns while avoiding trajectory divergence (such as "fishtailing") caused by long-term integration, outputting a continuous and smooth trajectory that conforms to the vehicle's dynamic characteristics.

[0021] In one specific embodiment, the unstructured road vehicle trajectory prediction method proposed in this invention, which integrates topological skeleton guidance and interactive multi-model estimation, is mainly executed by an onboard computing unit or a cloud server. The method specifically includes the following four steps: Step 1: Construction of Unstructured Environment Topology Skeleton Based on Morphological Refinement. This step is performed by the environment perception and processing module. For raster maps without lane lines, a mature morphological refinement algorithm is used to extract the discrete approximation of the Generalized Voronoi Diagram (GVD)—the skeleton lines—to construct a walkable road topology network. The specific implementation is as follows: Definition of environmental data The system receives a local environment binary occupancy grid map constructed by onboard sensors, denoted as For any pixel in the map Its state is defined as follows: Construction of Euclidean distance field (ESDF) definition The set of all free space pixels. Let this be the set of all obstacle pixels. To quantify the vehicle's "safety margin" at its current position and determine the center of the road, we need to consider the set... Perform a Euclidean distance transformation on each pixel. Define the distance function. Its value is equal to that of the pixel. The straight-line Euclidean distance to the nearest obstacle boundary point: After this processing, the original binary map is transformed into a grayscale distance field map. In the distance field, the larger the value, the farther away from the obstacle, and the local maximum point corresponds to the centerline position of the road.

[0022] Skeleton Extraction Based on Morphological Refinement Based on the distance field A morphological thinning algorithm was used to extract the road skeleton. This algorithm simulates a "layer-by-layer peeling" process. While maintaining the overall connectivity of the image, iterative erosion is performed on the edges of free regions in the distance field. Each iteration removes the outermost boundary pixels until the free region shrinks to a single-pixel-width line that can no longer be removed. The final pixel set is obtained as follows. This is a discrete approximation of the Generalized Vinograph (GVD). The skeleton line naturally lies at the midpoint of all obstacles, maximizing the representation of the traversable topology of unstructured roads. Finally, Mapping to the world coordinate system yields the physical topology skeleton point set. .

[0023] Step 2: State estimation based on Interactive Multi-Model Extended Kalman Filter (IMM-EKF). To address the variable vehicle motion patterns in unstructured roads (e.g., switching from constant speed straight driving to sharp turns), this step designs an Interactive Multi-Model (IMM) estimator. This estimator runs three filters with different dynamic characteristics in parallel, achieving robust estimation of the vehicle's full motion state through probability-weighted fusion.

[0024] Dynamic description of motion modes This system defines three motion modes, each corresponding to different driving behaviors of the vehicle. All models employ discrete-time state transition equations. Describe, in which This represents the sampling time interval.

[0025] (1) Mode A: Constant velocity model (CVModel) Applicable operating conditions: When the vehicle is driving straight or cruising in a steady state.

[0026] State vector: (Location ,speed ,course ).

[0027] Dynamic equations: (2) Mode B: Constant Acceleration Model (CAModel) Applicable operating conditions: When the vehicle is in the longitudinal acceleration or emergency braking phase.

[0028] State vector: (Increase longitudinal acceleration) ).

[0029] Dynamic equations: (3) Mode C: Constant angular acceleration variable curvature model (CYRA Model) Applicable working conditions: When the vehicle is making a sharp turn, making a U-turn, or going around an obstacle in an S-shape.

[0030] State vector: (Increase angular velocity) and angular acceleration ).

[0031] Dynamic equations: State estimation and fusion principle (1) Definitions of state variables and observables System full state vector: To unify the output, the fused state vector is defined to include the state components of all sub-models. .

[0032] System observations: The observation vectors directly provided by the sensors are: .

[0033] (2) Application of Extended Kalman Filter (EKF) In this system, the vehicle's position update equation (such as...) The system exhibits significant state coupling and nonlinear characteristics. That is, the change in position depends not only on the velocity but also on the trigonometric function value of the heading angle in a nonlinear relationship. Traditional Kalman filtering (KF) is only applicable to linear systems. To address the aforementioned nonlinearity issues, this invention employs extended Kalman filtering (EKF). Its core principle is to perform a first-order Taylor expansion of the nonlinear function at each estimated time point by calculating the Jacobian matrix, thereby achieving local linearization of the system and ensuring the tracking accuracy for variable curvature motion.

[0034] (3) IMM Algorithm Fusion Process Interaction: Based on the Markov transition probability matrix, calculate the mixture probability between the models at the previous time step. This step aims to redistribute "historical information" according to the model's credibility, preventing state jumps during model switching.

[0035] Parallel filtering: Three EKF filters operate in parallel, each based on the current observations. Calculate the respective observation residuals (Innovation).

[0036] Probability Update: Calculate the likelihood function for each model. The smaller the residual between the predicted and observed values ​​for a model, the higher its likelihood, and the higher the corresponding model probability. The larger the residual, the greater the weight. For example, when a vehicle makes a sharp turn, the residual of the CYRA model is the smallest, and its weight will automatically increase.

[0037] Weighted fusion: The final output state The weighted sum of the three model state estimates: This mechanism ensures that the system can automatically identify the current driving intention and output the state estimate that best matches the current motion characteristics.

[0038] Step 3: Multimodal Intent Prediction Based on Topological Scoring and Nonmaximum Suppression. This step is performed by the intent prediction module. This module combines the physical topological skeleton extracted in Step 1. Full vehicle fusion state output from step two By using spatiotemporal consistency scoring and multimodal screening mechanisms, the vehicle's performance within a finite time domain is predicted. A set of discrete target locations that may be reached in 3 to 5 seconds.

[0039] Dynamic sector search for candidate target points To determine an effective search space in complex unstructured road networks, the system constructs dynamic search sectors based on the vehicle's current motion state. The vehicle's current fused position is defined as... The magnitude of the velocity vector is The heading angle is The construction of the search sector must satisfy the following two geometric constraints: (1) Distance constraint (vertical coverage): Set the dynamic search radius : in This is the distance redundancy factor (typically ranging from 1.2 to 1.5), used to cover potential vehicle acceleration behavior and prevent missed detections.

[0040] (2) Field of view constraint (lateral coverage): Traverse the physical topology skeleton point set Each point in If it satisfies the following formula, then it is added to the candidate target point set. : in, This represents the phase angle of a vector. The half-angle threshold of the forward search field of view (e.g.) ).

[0041] Intent scoring function based on spatiotemporal consistency To quantify the posterior probability of a vehicle heading to each candidate point, an intent scoring function is established. This function is a weighted average of a direction consistency component and a distance matching component. For the candidate set... any point in The scoring formula is as follows: Directional consistency components : Measures the degree of alignment between the candidate point's orientation and the vehicle's current instantaneous speed direction.

[0042] in, displacement vector The direction angle. This value is distributed across... The interval indicates that the more a point matches the current course, the higher its score.

[0043] Distance matching component This measures whether the distance between the candidate point and the vehicle conforms to the current kinematic calculation. It is described using a Gaussian kernel function. in, This is the bandwidth parameter of the Gaussian kernel. This parameter reflects the vehicle's... The probability of "just arriving" at that point within a given timeframe.

[0044] Multimodal screening based on nonmaximum suppression (NMS) Since skeleton points are densely distributed in space, directly selecting high-scoring points will cause the prediction results to cluster on a single path. To extract independent intentions representing different path branches (such as different destinations at forks in the road), this system introduces the Non-Maximum Suppression (NMS) algorithm. The algorithm flow is as follows: Sorting: arranging the candidate point set According to the rating Sort from highest to lowest.

[0045] Initialization: Create an empty set Used to store the ultimate intent target.

[0046] Iterative filtering: like If empty, the loop ends.

[0047] take out The highest scorer in the current middle Add it to the set .

[0048] Suppress redundancy: Traversal For all remaining points in the array, calculate their intersection with the array. The Euclidean distance. If the distance is less than the set spatial suppression threshold. (For example, 3.0 meters), then it is considered a redundant point of the same intent pattern and is removed from... Removed from the middle.

[0049] Repeat the steps above.

[0050] Output: From the set Before the election These points serve as the final set of intended target points. .

[0051] Step 4: Bezier trajectory generation based on asymmetric dynamic constraints. This step is performed by the trajectory generation module. For each intended target point, a fifth-order Bezier curve connecting the vehicle's current state and the target point is generated. In existing technologies, although high-order models (such as CYRA) can accurately estimate the current angular acceleration, they lead to a cubic amplification of the error during long-term integration (fishtailing). This step proposes an asymmetric strategy of "high-order anchoring at the near end and low-order degradation at the far end": proximal ( ): Strictly anchoring step two, the IMM-EKF output includes angular acceleration. The higher-order state ensures transient response to sharp turns.

[0052] remote( Forced adoption of reduced-order steady-state assumption (ignoring) This eliminates the divergence sources of long-time integration from a mechanistic perspective.

[0053] Definition of a fifth-order Bézier curve Predicted trajectory Composed of 6 control points Define, where the normalized time parameter The equation of the curve is: in is the binomial coefficient.

[0054] Proximal control point calculation (higher-order dynamic anchoring) Using the full fused state vector output by IMM-EKF in step two Calculate the first three control points. This process explicitly incorporates the angular velocities estimated by the CYRA model. and angular acceleration .

[0055] (Position constraints): (First derivative / velocity constraint): Determined by the vehicle's current velocity vector, ensuring tangential continuity of the trajectory. This represents the optimal velocity estimate obtained after step two using IMM-EKF filtering.

[0056] (Second derivative / acceleration constraint): The resultant acceleration vector after IMM-EKF filtering. Decision. The resultant acceleration includes longitudinal acceleration. (Tangential) and centripetal acceleration (Normal). Based on the properties of the second derivative of Bézier curves. Solving inversely, we get: Wherein, the resultant acceleration vector The components in the world coordinate system are: Remote control point calculation (asymmetric order reduction / anti-drift) Calculate the last three control points At that time, a reduced-order assumption is enforced, that is, it is assumed that the vehicle has entered a steady state of motion when it reaches the target point, and the angular acceleration is... .

[0057] (Target position constraints): (End-of-course tangential constraint): Assume that the vehicle's heading when it reaches the target point is consistent with the direction of the line connecting the vehicle and the target point (homing assumption). Define the end-of-course tangential unit vector. .

[0058] in To predict the terminal velocity, the current velocity can be used. Or based on longitudinal acceleration Calculation.

[0059] (End curvature constraint): To prevent violent oscillations at the end of the trajectory, a "zero acceleration change" or "linear smooth transition" constraint is applied here: This constraint implicitly assumes that the terminal acceleration is 0, thereby forcibly straightening the end of the trajectory and achieving the beneficial effect of preventing tail-wagging.

[0060] Trajectory Interpolation and Output Substitute the calculated control points into the Bessel equations, with a fixed time step (e.g., Discretization interpolation is performed to output a set of multiple spatiotemporally continuous predicted trajectories. Each trajectory contains a future position sequence, velocity sequence, and heading sequence.

[0061] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A method for predicting unstructured road trajectories by integrating topological skeletons and interactive multi-models, characterized in that, include: S1. Extract the topological skeleton lines of unstructured roads based on environmental perception data; S2. An interactive multi-model extended Kalman filter is used to estimate and fuse the motion state of the target vehicle in parallel, and the output is a full state vector containing position, velocity, heading, acceleration and angular velocity. S3. Combining the topology skeleton line and the full state vector, multiple discrete intention target points are predicted through dynamic sector search and intent scoring mechanisms. S4. For each of the intended target points, generate a fifth-order Bézier curve trajectory based on asymmetric dynamic constraints.

2. The method for predicting unstructured road trajectories by fusing topological skeletons and interactive multi-models according to claim 1, characterized in that, Step S1 specifically includes: Acquire environmental perception data, and based on the environmental perception data, characterize the unstructured environment into a binary occupancy grid map. The state of each pixel in the binary occupancy grid map is represented by obstacle occupancy and free space, respectively. For each free space pixel in a binary occupied raster map, calculate its Euclidean distance to the nearest obstacle pixel to characterize its safety margin, and construct an Euclidean distance symbol field. A morphological thinning algorithm is performed on the Euclidean distance symbol field to iteratively erode the boundary pixels of the free region until the free region is thinned into a set of connected lines with a width of one pixel. This set of lines is used as an approximation of the generalized Vinograph on the discrete raster map and serves as the topological skeleton line. Mapping the topological skeleton lines in the grid coordinates to the world coordinate system where the vehicle is located yields the physical topological skeleton point set, which represents the traversable topological structure of the environment.

3. The method for predicting unstructured road trajectories by fusing topological skeletons and interactive multi-models according to claim 1, characterized in that, The interactive multi-model extended Kalman filter includes constant velocity, constant acceleration, and constant angular acceleration variable curvature models. All models are described using discrete-time state transition equations, where: The constant speed model is used to describe the state of a vehicle when it is traveling straight or in steady-state cruise, and to determine the vehicle's motion state with respect to position, speed, and heading. The constant acceleration model is used to describe the state of a vehicle during longitudinal acceleration or emergency braking and to determine the vehicle's motion state with respect to longitudinal acceleration. The constant angular acceleration variable curvature model is used to describe the state of a vehicle when it is making a sharp turn, U-turn, or S-shaped obstacle avoidance, and to determine the vehicle's motion state with respect to angular velocity and angular acceleration.

4. The method for predicting unstructured road trajectories by fusing topological skeletons and interactive multi-models according to claim 3, characterized in that, Parallel estimation and fusion of the motion state of the target vehicle includes: Model definition and initialization: Configure an extended Kalman filter for each of the constant velocity model, constant acceleration model, and constant angular acceleration variable curvature model; define the system observation vector as the observation vector directly provided by the sensor; and define the system full state vector used for the final fusion output as the motion state output of the three models. Model interaction: At each time step, based on the preset Markov model transition probability matrix, the mixture probability between the models at the previous time step is calculated. This is used to redistribute historical information according to the credibility of each model to prevent state jumps when switching models. Parallel filtering and observation update: Three extended Kalman filters work in parallel based on the current observations, each calculating its own observation residual. At each estimation point, the extended Kalman filter performs a first-order Taylor expansion of the nonlinear state transition function by calculating the Jacobian matrix to achieve local linearization of the system. Model probability update: The likelihood function is calculated based on the observation residuals of each model. The smaller the residual between the predicted value and the observed value of the model, the higher its likelihood and the greater the corresponding model probability, thereby enabling the system to automatically recognize the current driving intention. Weighted fusion output: The final output is the weighted sum of the three model state estimates as the fused state.

5. The method for predicting unstructured road trajectories by fusing topological skeletons and interactive multi-models according to claim 1, characterized in that, Step S3 includes: Based on the vehicle's current position, speed, and heading angle output in step S2, a dynamically changing forward search region is constructed. The search region is a fan-shaped area with the current position as the vertex and the current heading as the central axis. The search radius of the fan-shaped area is dynamically determined by the product of the current speed, the preset prediction time, and a distance redundancy coefficient greater than one. Its horizontal subtended angle is determined by a preset field of view half-angle threshold. Based on the topology skeleton line output in step S1, all points that are simultaneously located within the sector area and within the preset distance and angle thresholds are selected to form an initial candidate target point set. A comprehensive intent score is calculated for each point in the initial candidate target point set, wherein the intent score is a weighted combination of two parts: direction consistency score and distance matching score. The direction consistency score measures the degree of agreement between the direction of the candidate point relative to the vehicle's current position and the vehicle's current heading angle. The more consistent they are, the higher the score. The distance matching score assesses whether the distance between the candidate point and the vehicle's current position is within the theoretically achievable range based on the current speed and prediction time. The more the distance matches, the higher the score. Non-maximum suppression filtering is performed on the candidate point set after scoring. This filtering process is carried out iteratively: each time, the point with the highest intent score is selected from the current candidate point set as one of the valid intent target points and removed from the set. At the same time, all other candidate points within a certain distance around the point are removed to avoid redundant points that are too dense in space being repeatedly selected as different intents. Repeat the selection and removal process above until the candidate point set is empty, or the preset limit for the number of target points has been reached. The final output consists of multiple discrete intention target points, representing different path branches that the vehicle may travel to in the future prediction period.

6. The method for predicting unstructured road trajectories by fusing topological skeletons and interactive multi-models according to claim 1, characterized in that, Step S4 includes: For each intention target point output in step S3, a future trajectory is generated smoothly from the current state of the vehicle to the target point using a fifth-order Bézier curve as the basic mathematical model. The fifth-order Bézier curve is fully defined by six control points. Calculate the first three control points of the Bézier curve using the fully fused state vector: The first control point is obtained by using the position coordinate components in the vehicle's fully fused state vector. The second control point is obtained by performing a scalar multiplication operation on the vehicle speed vector based on the total prediction time and the order characteristics of the Bézier curve, and then performing a vector addition operation with the first control point. The third control point is obtained by constructing a second-order derivative constraint equation for the Bézier curve with the vehicle's resultant acceleration vector as a known condition and solving the equation. Calculate the last three control points of the Bézier curve using the intended target point and the reduced-order motion assumption: The sixth control point is obtained by directly mapping the coordinates of the intended target point. By calculating the unit direction vector from the first control point to the sixth control point, and performing scalar and vector multiplication operations on it with the terminal predicted velocity value, the predicted total duration and the Bézier curve order parameter, an offset vector is generated. Then, a vector subtraction operation is performed on the coordinates of the sixth control point and the offset vector to obtain the fifth control point. The fourth control point is obtained by substituting the coordinates of the fifth and sixth control points into the linear smooth constraint relationship that characterizes the change of zero acceleration at the end point and performing a linear combination operation. After obtaining all six control points, discretization interpolation is performed using the Bézier curve formula, and the final output is a continuous, smooth future trajectory sequence that conforms to asymmetric dynamic constraints for each intended target point.

Images