Urban and rural road dynamic optimization planning system based on ai and multi-source data

Abstract

The application discloses an urban and rural road dynamic optimization planning system based on AI and multi-source data, and relates to the technical field of traffic engineering. The method acquires multi-source vehicle space-time trajectory flow, constructs a physical filter, and generates a dynamic compliance trajectory sequence set. Then, the sequence set is subjected to time sequence feature extraction and maneuvering feature, and the maneuvering feature is converted into a geometric conflict area in two-dimensional space. The motion state change points in the trajectory are extracted as dynamic semantic anchor points, which are rigidly transformed and aligned with a preset three-dimensional terrain model, and a to-be-processed voxel subspace is generated based on the initial speed of the vehicle. In the subspace, a Boolean difference set of a vehicle dynamics safety envelope body and a driver's field of view body is constructed, a field of view gain score model of normalized impedance is used to calculate the priority of removing environmental voxels, and an environmental element removal list is generated. The application realizes the accurate positioning and minimally invasive intervention of urban and rural road hidden obstacles by inversely reconstructing the road geometric environment from sparse trajectory data.

Description

Technical Field

[0001] This invention relates to the field of traffic engineering technology, and in particular to a dynamic optimization planning system for urban and rural roads based on AI and multi-source data. Background Technology

[0002] In the planning and management of urban and rural road networks, existing technologies face a technical challenge: establishing a quantitative mapping between the microscopic dynamics of vehicles and the macroscopic static physical environment of roads, especially in the absence of high-precision surveying data and roadside sensing equipment. This binary separation between sensing data and physical entities prevents planning systems from accurately locating geometric conflict areas causing insufficient visibility or traffic obstruction using the widely available sparse vehicle spatiotemporal trajectory data. Specifically, existing technologies either rely on expensive lidar scanning to directly acquire physical environment information, which is not economically feasible in large-scale urban and rural road networks; or they can only perform macroscopic traffic statistics on trajectory data, failing to uncover the physical constraints behind trajectory fluctuations and thus unable to identify hidden physical obstacles such as vegetation obstruction and limited visibility due to embankments at a three-dimensional spatial scale. Therefore, existing technologies cannot provide minimally invasive intervention solutions for specific environmental voxels, and can only adopt extensive expansion or reconstruction methods, resulting in a waste of engineering resources and a lack of targeted management capabilities. Summary of the Invention

[0003] This invention provides a dynamic optimization planning system for urban and rural roads based on AI and multi-source data. It aims to solve the technical problem of how to use sparse and heterogeneous vehicle spatiotemporal trajectory data to reversely construct a quantitative mapping between the micro-dynamic behavior of vehicles and the macro-physical environment of roads under the condition of lack of high-precision surveying data and roadside sensing equipment, so as to accurately locate and eliminate geometric conflict areas that cause insufficient line of sight in three-dimensional voxel space.

[0004] In view of the above problems, this invention provides a dynamic optimization planning method for urban and rural roads based on AI and multi-source data, including the following steps: Dynamics compliance verification: Obtain multi-source vehicle spatiotemporal trajectory streams, construct a physical filter containing vehicle kinematic geometric constraints, perform point-by-point verification on the spatiotemporal trajectory streams, eliminate trajectories that do not meet the vehicle kinematic geometric constraints, and generate a set of dynamically compliant trajectory sequences; Geometric conflict inversion: Temporal features are extracted from the dynamic compliance trajectory sequence set to identify maneuver features that characterize the driver's evasive behavior, and the maneuver features are transformed into geometric conflict regions in two-dimensional space through an inverse mapping function; Heterogeneous spatiotemporal semantic anchor registration: Introducing a pre-set three-dimensional terrain model, extracting motion state change points from the dynamic compliance trajectory sequence set as dynamic semantic anchors, performing rigid transformation alignment with static feature points in the pre-set three-dimensional terrain model, and generating a voxel subspace to be processed containing the geometric conflict region based on the vehicle's initial velocity. Intervention decision generation: Within the voxel subspace to be processed, construct the Boolean difference set between the vehicle dynamics safety envelope and the driver's field of vision, calculate the ratio of the volume gain to the removal cost of removing environmental voxels to the driver's field of vision, and generate an environmental element removal list.

[0005] Furthermore, in the dynamics compliance verification step, the construction of a physical filter containing vehicle kinematic geometric constraints specifically includes: The lateral acceleration threshold and the heading angle change rate threshold are set based on the Ackermann steering geometry principle; When the instantaneous motion state of a trajectory point exceeds the lateral acceleration threshold or the heading angle change rate threshold, the trajectory is determined to be an unrestricted off-road trajectory and is discarded.

[0006] Furthermore, in the geometric conflict inversion step, the specific implementation of the geometric conflict inversion includes: Deep temporal neural networks are used to identify the S-shaped avoidance characteristics concentrated in straight topology or the turning maneuver characteristics in curved topology of the dynamic compliance trajectory sequence. Based on the occurrence frequency and spatial distribution density of the aforementioned features, a probability heatmap of the geometric conflict region is generated.

[0007] Furthermore, in the heterogeneous spatiotemporal semantic anchor registration step, the heterogeneous spatiotemporal semantic anchor registration specifically includes: The stopping or deceleration starting point in the trajectory is identified as the dynamic semantic anchor point; The coordinate system transformation matrix is ​​solved by iteratively matching the dynamic semantic anchor points with the coordinates of the curb stones or stop lines in the three-dimensional terrain model to obtain the nearest point. Using the center of the geometric conflict region as the origin and the braking safety distance corresponding to the initial velocity of the vehicle as the length, a three-dimensional bounding box is constructed, and the environmental voxels falling into the three-dimensional bounding box are defined as the voxel subspace to be processed.

[0008] Furthermore, in the intervention decision generation, the generated environmental element removal list adopts a normalized impedance-based view gain scoring model for voxels in the voxel subspace to be processed. Removal priority The calculation formula is: in: To remove voxels The volume increment of the driver's field of vision is described later; The preset reference view volume; The visual kinetic interference index characterizes voxels. The occlusion ratio of the vehicle dynamics safety envelope; It is the hyperbolic tangent function; and The preset adjustment coefficient; The impedance modulus is removed from the virtual environment obtained by inverting the avoidance urgency based on the aforementioned maneuver characteristics. This is the basic starting impedance constant used to prevent the denominator from being zero; This is the preset reference impedance; This is a cost sensitivity index.

[0009] A second aspect of the present invention provides a dynamic optimization planning system for urban and rural roads based on AI and multi-source data, the system comprising: processor; A memory storing computer-executable instructions that, when executed by the processor, cause the processor to perform the steps of the method described in any of the above schemes.

[0010] The technical solution provided in this application has at least the following technical effects: by establishing a reverse reconstruction mechanism from vehicle dynamics behavior characteristics to road physical geometry, it solves the problem of alignment and mapping of heterogeneous spatiotemporal data in physical space. It can accurately invert and locate geometric conflicts such as line-of-sight occlusion in the road based solely on compliant vehicle trajectory data without relying on roadside sensing hardware. Furthermore, it decouples the field-of-sight gain and engineering removal cost based on a normalized impedance scoring model, thereby directly outputting an environmental element removal list that takes into account both safety benefits and engineering economy. This achieves accurate perception and minimally invasive intervention of the physical environment of urban and rural roads under low data cost constraints. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall structure of the urban and rural road dynamic optimization planning system based on AI and multi-source data provided in an embodiment of the present invention. Detailed Implementation

[0012] The above technical solutions will now be described in detail with reference to the accompanying drawings and specific embodiments to provide a better understanding of them. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments used only to explain the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the drawings, not all of them.

[0013] For examples, please refer to Figure 1 This invention provides a dynamic optimization planning system for urban and rural roads based on AI and multi-source data. This system achieves reverse reconstruction and minimally invasive management of the road physical environment by executing a dynamic optimization planning method based on AI and multi-source data. Logically, this method includes the following four main processing steps: First, there is the dynamics compliance verification: The system acquires the spatiotemporal trajectory streams of multi-source vehicles, constructs a physical filter containing vehicle kinematic geometric constraints, performs point-by-point verification on the spatiotemporal trajectory streams of multi-source vehicles, eliminates trajectories that do not meet the vehicle kinematic geometric constraints, and generates a set of dynamically compliant trajectory sequences. Secondly, geometric conflict inversion: the system extracts temporal features from the dynamic compliance trajectory sequence set, identifies the maneuver features that characterize the driver's evasive behavior, and transforms the maneuver features into geometric conflict regions in two-dimensional space through an inverse mapping function; Next is the heterogeneous spatiotemporal semantic anchor registration: the system extracts the motion state change points in the dynamic compliance trajectory sequence set as dynamic semantic anchors, performs rigid transformation alignment with the static feature points in the preset 3D terrain model, and generates a voxel subspace containing geometric conflict areas based on the vehicle's initial velocity. Finally, intervention decision generation: within the voxel subspace to be processed, the system constructs the Boolean difference set between the vehicle dynamics safety envelope and the driver's field of vision, calculates the ratio of the volume gain to the removal cost of removing environmental voxels to the driver's field of vision, and generates an environmental element removal list.

[0014] The dynamic optimization planning method for urban and rural roads based on AI and multi-source data provided by this invention is implemented by an optimization planning system. The entire process of the optimization planning system begins with the digital reconstruction of vehicle behavior data from the physical world. Its core logic lies in establishing a rigorous data processing command chain to ensure that only high-quality data conforming to physical laws can enter subsequent decision-making stages. The optimization planning system first initiates the urban and rural road optimization planning process based on vehicle dynamics inversion. Through standardized interfaces, the optimization planning system continuously accesses multi-source vehicle spatiotemporal trajectory streams and sequentially schedules four sub-tasks: dynamics compliance verification, geometric conflict inversion, heterogeneous spatiotemporal semantic anchor point registration, and intervention decision generation, to achieve accurate positioning and repair of geometric conflict areas in urban and rural roads.

[0015] After the multi-source vehicle spatiotemporal trajectory streams are integrated into the optimization planning system, the standardized access and preprocessing procedures for heterogeneous trajectory streams are immediately initiated. Before entering the calculation process, the raw data packets containing timestamps, longitude, latitude, instantaneous speed, and heading angle dimensions undergo coordinate system transformation. Raw positioning data in WGS-84 or GCJ-02 coordinate formats are uniformly converted to universal transverse Mercator projection coordinates, thus mapping spherical latitude and longitude values ​​to meter-level measurements in a Cartesian coordinate system. Following the coordinate transformation, serialization based on the device's unique identifier and timestamp is performed, duplicate data points are removed, and the originally discrete data points are reassembled into a continuous spatiotemporal trajectory stream, establishing the data foundation for subsequent dynamic analysis.

[0016] After the multi-source vehicle spatiotemporal trajectory stream is standardized, the instantaneous state calculation process based on Ackerman geometry begins. The discrete-point difference algorithm is applied to each trajectory point in the multi-source vehicle spatiotemporal trajectory stream. The ratio of the change in Euclidean distance between two adjacent trajectory points to the timestamp difference is calculated and used to verify or regenerate instantaneous velocity values. Simultaneously, the derivative of the heading angle with respect to time is calculated to obtain the rate of change of the heading angle. Based on vehicle dynamics principles, the instantaneous lateral acceleration is derived by multiplying the instantaneous velocity by the rate of change of the heading angle. This calculation path does not rely on onboard sensors; it obtains the physical parameters describing the vehicle's motion state solely through second-order difference derivation of position information, providing quantified input variables for subsequent physical constraint filters.

[0017] Along with the generation of instantaneous physical state parameters, a physical filter incorporating vehicle kinematic geometric constraints is constructed and put into operation. Based on the Ackermann steering geometry principle, lateral acceleration thresholds and heading angle change rate thresholds are pre-set in the physical filter. The lateral acceleration threshold is set to 0.4g, which corresponds to the lateral adhesion limit of a normal vehicle driving on a non-off-road surface. The physical filter performs point-by-point verification on the spatiotemporal trajectory flow of multi-source vehicles. Once the instantaneous lateral acceleration value of a trajectory point exceeds the lateral acceleration threshold, or the turning radius derived from the heading angle change rate of the trajectory point is less than the minimum turning radius allowed by the vehicle's mechanical structure, the entire trajectory containing that trajectory point is determined to be an unrestricted off-road trajectory. Unrestricted off-road trajectories are physically eliminated by the optimization planning system, and the remaining trajectories that are not eliminated constitute a dynamically compliant trajectory sequence set. The dynamically compliant trajectory sequence set is output to the next stage as the sole data source for geometric conflict inversion, ensuring that the input data of the inversion algorithm strictly conforms to the physical motion laws of road vehicles.

[0018] After the dynamic compliance trajectory sequence set is transmitted to the geometric conflict inversion module, the deep temporal feature extraction program for the geometric conflict region is immediately initiated. A deep temporal neural network is built as the core computing engine, and a long short-term memory network architecture or a transformer encoder architecture is selected as the specific implementation form of the deep temporal neural network to capture the contextual dependencies in long sequence data. Each trajectory in the dynamic compliance trajectory sequence set is reconstructed into a multi-dimensional feature tensor, and the input vector dimension is defined as including the instantaneous velocity value, instantaneous lateral acceleration value, and rate of curvature change value of the trajectory point. The multi-dimensional feature tensor is batch-input into the deep temporal neural network, and the time series data is encoded through the network's internal gating mechanism or self-attention mechanism, thereby mapping the low-dimensional motion state data into a high-dimensional hidden layer feature vector.

[0019] Based on the generation of high-dimensional hidden layer feature vectors, the identification and classification of maneuver features begin. The deep temporal neural network performs pattern matching on trajectory behaviors according to pre-defined classification logic. On road segments marked as straight roads in the base map data, S-shaped avoidance features are specifically identified. These features are characterized by a continuous sinusoidal deviation of the vehicle trajectory in the lateral direction, accompanied by a significant decrease in longitudinal speed. This feature is determined by the system to be an emergency lane change by the driver to avoid sudden obstacles on the road. On road segments marked as curved topologies in the base map data, reversing maneuver features are specifically identified. These features are characterized by a vehicle's instantaneous speed dropping to zero, followed by a reversal of more than 90 degrees in the heading angle. This feature is determined by the system to be a large vehicle reversing to adjust its course due to insufficient turning radius. Through the above pattern matching, the specific driving behavior of each trajectory is assigned a clear maneuver feature label.

[0020] With the determination of maneuver feature labels, the generation process of a probability heatmap for two-dimensional geometric conflict regions is triggered. An inverse mapping function is defined and applied to the set of maneuver feature labels. The inverse mapping function uses the frequency and spatial distribution density of S-shaped avoidance features or weaving maneuvers identified within a specific region as independent variables to calculate the conflict probability value of the presence of physical obstacles in that region. The conflict probability value is mapped onto a two-dimensional grid map, forming a probability heatmap covering the target road network. Connected regions in the probability heatmap whose values ​​exceed a preset confidence threshold are extracted and defined as geometric conflict regions. The center coordinates and coverage area of ​​the geometric conflict regions are output as subsequent heterogeneous spatiotemporal semantic anchor point registration and two-dimensional target points for generating the voxel subspace to be processed.

[0021] Once the two-dimensional target point locations in the geometric conflict region are determined, the automatic extraction program for dynamic semantic anchor points is initiated within the dynamic compliance trajectory sequence set. Trajectory points exhibiting local extrema in longitudinal deceleration values ​​are identified as the starting points of rapid deceleration, and trajectory points where the instantaneous velocity value drops to zero are identified as stopping and jerk points. The starting points of rapid deceleration and stopping and jerk points are set together and defined as points of change in motion state, which are used as dynamic semantic anchor points. These points characterize the specific position coordinates of the vehicle in physical space under mandatory geometric constraints or traffic rule constraints.

[0022] Along with the generation of dynamic semantic anchor points, a pre-defined 3D terrain model is loaded into the computational unit. This model includes static feature points such as curb coordinates and stop line coordinates. An iterative nearest-point algorithm is applied between the dynamic semantic anchor point set and the static feature point set. The rigid transformation matrix solution process aims to minimize the average Euclidean distance between the dynamic semantic anchor points and their spatially neighboring static feature points. Through iterative calculation, the rigid transformation matrix, containing rotation and translation parameters, is solved. This rigid transformation matrix is ​​then applied inversely to the original GPS coordinates of the geometrically conflicting region, mathematically correcting satellite positioning drift and thus achieving precise alignment of heterogeneous spatiotemporal data within a unified physical coordinate system.

[0023] After coordinate alignment, the dynamic cutting operation of the voxel subspace to be processed is initiated. The average initial velocity values ​​of all trajectory points within the geometric conflict area are extracted. The braking safety distance is calculated based on the physical kinematics formula, which divides the square of the average initial velocity value by twice the product of the road friction coefficient and the gravitational acceleration, and adds the product of the average initial velocity value and the driver's reaction time to obtain the physical distance required for the vehicle to come to a complete stop. A 3D bounding box is constructed with the center coordinates of the geometric conflict area as the origin, the calculated braking safety distance as the length, the road width as the width, and the preset scanning height as the height. The 3D bounding box is used as a spatial filter. Only environmental voxel data falling within the 3D bounding box is cut from the 3D terrain model and loaded. These loaded environmental voxels constitute the voxel subspace to be processed. The voxel subspace to be processed will serve as the physical computation object for subsequent intervention decision generation. The height of the bounding box is set according to the average driving eye height of a conventional vehicle (e.g., 1.5 meters to 2.5 meters) to ensure that the calculation range covers the driver's effective line of sight area.

[0024] Once the voxel subspace to be processed is loaded, the intervention decision generation program is initiated. The driver's field of view is first established in three-dimensional space. The driver's field of view simulates the cone of vision space emanating from the driver's eye point and covering the road ahead, representing the range of physical space that the driver can clearly observe in the current driving state.

[0025] Simultaneously, the vehicle dynamics safety envelope is constructed based on the vehicle's physical dimensions and braking safety distance. This envelope represents the physical scan space required for the vehicle to safely stop at its current speed. Boolean difference operations are applied between the driver's field of view and the vehicle dynamics safety envelope. Environmental voxels located within the vehicle dynamics safety envelope but blocking ray projection from the driver's field of view are identified and labeled; these labeled voxels constitute the occlusion voxel set. For each occlusion voxel in the occlusion voxel set... The normalized impedance field-gain scoring model is invoked to calculate the removal priority. The calculation process strictly follows the following mathematical formula: In this formula, Defined as the volume increment of the driver's field of vision after removing voxels, this increment characterizes the degree of improvement in physical visibility. Defined as a preset reference view volume, used to perform dimensionless processing on volume increments. Defined as the visual kinetic interference index, this index characterizes the proportion of voxels that obscure the vehicle's dynamic safety envelope; the higher the value, the greater the safety hazard. Defined as a hyperbolic tangent function, it is used to map the visual kinetic interference exponent to the saturation region, simulating the diminishing marginal utility of the line-of-sight gain. and Defined as a preset adjustment coefficient, it is used to adjust the sensitivity of the safety weight.

[0026] The denominator of the formula is used to calculate the removal cost. Defined as voxel The virtual environment removes the impedance modulus. In this embodiment, the modulus is derived based on the strength of the maneuvering characteristics identified in the geometric conflict inversion step. Specifically, if the maneuvering characteristics are characterized by high-frequency abrupt avoidance or turning back (e.g., extremely high rate of curvature change), the system determines that the location is a high-impedance rigid obstacle (e.g., a wall, guardrail) and assigns a higher impedance modulus. Values ​​(e.g., 100-1000); if the maneuvering characteristics are low-frequency, smooth avoidance, the system determines it as a low-impedance flexible obstacle (e.g., vegetation) and assigns a lower value. Values ​​(e.g., 1-10) ensure that the system can still accurately estimate the removal cost through the vehicle's 'behavioral feedback' even without visual perception data, thus avoiding recommending the removal of high-cost, rigid facilities. Defined as the basic starting impedance constant, this constant is set to a non-zero positive number to prevent the mathematical singularity of a zero denominator when the impedance modulus is removed by the environment. Defined as a preset reference impedance. Defined as a cost sensitivity index, this index, as a power of a power function, imposes a non-linear penalty on the removal recommendation of high-impedance objects. The removal priority is determined by this formula. It is calculated as a quantization ratio of field-of-view gain to nonlinear impedance cost.

[0027] With the completion of priority scoring, the generation and output of the environmental element removal list is executed as the final step in the process. A preset priority threshold is loaded into the comparator. The removal priority value of each occlusion voxel is compared with the priority threshold. Occlusion voxels with removal priority values ​​higher than the priority threshold are retained and aggregated. The spatial coordinates, physical volume, and corresponding suggested removal actions of the aggregated voxels are encapsulated and written into the environmental element removal list. The environmental element removal list is output by the optimization planning system, serving as the final decision-making basis for guiding urban and rural road intervention and repair projects.

[0028] The physical implementation of the urban and rural road dynamic optimization planning system based on AI and multi-source data relies on a specific hardware architecture, which consists of a modular processor and non-volatile memory. The modular processor is configured as the core computing platform of the system. The four logical steps mentioned above—dynamic compliance verification, geometric conflict inversion, heterogeneous spatiotemporal semantic anchor registration, and intervention decision generation—are mapped to specific computational tasks within the processor. For the deep temporal neural network inference task involved in the geometric conflict inversion step, a graphics processing unit (GPU) or tensor processing unit (TPU) is integrated into the modular processor to provide high-concurrency matrix multiplication acceleration capabilities, thereby ensuring computational efficiency for real-time feature extraction of large-scale trajectory sequence sets.

[0029] For the iterative nearest point algorithm and rigid transformation matrix solving tasks involved in the heterogeneous spatiotemporal semantic anchor point registration step, arithmetic logic units with high-precision floating-point operation capabilities are allocated to perform vector and matrix operations to meet the sub-meter accuracy requirements of coordinate alignment. A high-speed data communication connection is established between the memory and the modular processor to maintain the system's data throughput and instruction scheduling. A pre-built 3D terrain model library is persistently stored in this memory, providing static feature point data support for heterogeneous spatiotemporal semantic anchor point registration. Simultaneously, a computer-executable instruction set resides in the memory, encoding the aforementioned dynamic compliance verification logic, geometric conflict inversion algorithm logic, anchor point registration logic, and normalized impedance scoring logic. When the optimization planning system starts, the modular processor reads and loads the computer-executable instructions from the memory, driving the multi-source vehicle spatiotemporal trajectory flow to and from various logic modules according to a predetermined timing sequence. Finally, the calculated environmental element removal list is written back to the memory or output through an external interface, thus completing a complete hardware execution closed loop from data input to physical decision output.

[0030] The above 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 with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic optimization planning method for urban and rural roads based on AI and multi-source data, characterized in that, Includes the following steps: Dynamics compliance verification: Obtain multi-source vehicle spatiotemporal trajectory streams, construct a physical filter containing vehicle kinematic geometric constraints, perform point-by-point verification on the spatiotemporal trajectory streams, eliminate trajectories that do not meet the vehicle kinematic geometric constraints, and generate a set of dynamically compliant trajectory sequences; Geometric conflict inversion: Temporal features are extracted from the dynamic compliance trajectory sequence set to identify maneuver features that characterize the driver's evasive behavior, and the maneuver features are transformed into geometric conflict regions in two-dimensional space through an inverse mapping function; Heterogeneous spatiotemporal semantic anchor registration: Introducing a pre-set three-dimensional terrain model, extracting motion state change points from the dynamic compliance trajectory sequence set as dynamic semantic anchors, performing rigid transformation alignment with static feature points in the pre-set three-dimensional terrain model, and generating a voxel subspace to be processed containing the geometric conflict region based on the vehicle's initial velocity. Intervention decision generation: Within the voxel subspace to be processed, construct the Boolean difference set between the vehicle dynamics safety envelope and the driver's field of vision, calculate the ratio of the volume gain to the removal cost of removing environmental voxels to the driver's field of vision, and generate an environmental element removal list.

2. The method according to claim 1, characterized in that, In the dynamics compliance verification step, the construction of a physical filter containing vehicle kinematic geometric constraints specifically includes: The lateral acceleration threshold and the heading angle change rate threshold are set based on the Ackermann steering geometry principle; When the instantaneous motion state of a trajectory point exceeds the lateral acceleration threshold or the heading angle change rate threshold, the trajectory is determined to be an unrestricted off-road trajectory and is discarded.

3. The method according to claim 1, characterized in that, The specific implementation of the geometric conflict inversion step includes: Deep temporal neural networks are used to identify the S-shaped avoidance characteristics concentrated in straight topology or the turning maneuver characteristics in curved topology of the dynamic compliance trajectory sequence. Based on the occurrence frequency and spatial distribution density of the aforementioned features, a probability heatmap of the geometric conflict region is generated.

4. The method according to claim 1, characterized in that, In the heterogeneous spatiotemporal semantic anchor registration step, the heterogeneous spatiotemporal semantic anchor registration specifically includes: The stopping or deceleration starting point in the trajectory is identified as the dynamic semantic anchor point; The coordinate system transformation matrix is ​​solved by iteratively matching the dynamic semantic anchor points with the coordinates of the curb stones or stop lines in the three-dimensional terrain model to obtain the nearest point. Using the center of the geometric conflict region as the origin and the braking safety distance corresponding to the initial velocity of the vehicle as the length, a three-dimensional bounding box is constructed, and the environmental voxels falling into the three-dimensional bounding box are defined as the voxel subspace to be processed.

5. The method according to claim 1, characterized in that, In the intervention decision generation step, the generation of the environmental element removal list adopts a normalized impedance field gain scoring model for voxels in the voxel subspace to be processed. Removal priority The calculation formula is: in: To remove voxels The volume increment of the driver's field of vision is described later; The preset reference view volume; The visual kinetic interference index characterizes voxels. The occlusion ratio of the vehicle dynamics safety envelope; It is the hyperbolic tangent function; and The preset adjustment coefficient; The impedance modulus is removed from the virtual environment obtained by inverting the avoidance urgency based on the aforementioned maneuver characteristics. This is the basic starting impedance constant used to prevent the denominator from being zero; This is the preset reference impedance; This is a cost sensitivity index.

6. A dynamic optimization planning system for urban and rural roads based on AI and multi-source data, characterized in that: The system includes: processor; A memory storing computer-executable instructions that, when executed by the processor, cause the processor to perform the method steps as described in any one of claims 1 to 5.

Images