Risk field modeling and path planning method and system based on three-dimensional semantic occupancy

Abstract

The application discloses a risk field modeling and path planning method and system based on three-dimensional semantic occupation, and the method comprises the following steps: acquiring a vehicle driving space, constructing a three-dimensional semantic occupation matrix according to the driving space; based on the constructed three-dimensional semantic occupation matrix, a three-dimensional risk field of the vehicle is established by using a three-dimensional Gaussian risk kernel function with multiple parameters, and the three-dimensional risk field is converted into a two-dimensional risk field; the vehicle path is segmented and planned according to the two-dimensional risk field, and a vehicle driving path is obtained. Through the risk field construction of the three-dimensional semantic occupation matrix, the global unified risk probability distribution modeling of various types of risk elements in the three-dimensional space is realized, the risk values of the trajectory points of the risk field and the global risk values are respectively integrated into the initial path planning of the first stage and the path optimization of the second stage, and the gradient descent iterative optimization of the initial path of the first stage is simultaneously realized by fusing the smoothness cost, so that the balance between the path comfort and the safety is ensured.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology for intelligent vehicles, and in particular to a risk field modeling and path planning method and system based on three-dimensional semantic occupancy. Background Technology

[0002] Path planning is a core function of autonomous driving systems, relying on environmental perception results to generate safe, feasible, and smooth driving trajectories. As the complexity of autonomous driving scenarios increases, path planning is gradually evolving from traditional obstacle avoidance methods based on geometric constraints to decision-making methods that incorporate environmental risk information. Early research primarily used empirical risk indicators such as Time-to-Care (TTC) and Time-to-Like (TLC), reflecting safety margins through critical time thresholds. However, these indicators are based on the relative motion state of the vehicle and a single adjacent vehicle and deterministic thresholds, making it difficult to consider the interactions between multiple traffic participants and the potential uncertainties of dynamic scenario evolution. Subsequent research has focused on probabilistic modeling-based risk assessment methods. Researchers have constructed risk field models in the form of probability distributions to integrate the uncertainties of dynamic and static environmental factors. For example, risk is represented as the product of the risk probability distribution and the expected severity, or risk calculation is based on ellipsoidal contours incorporating vehicle dynamic attributes. These methods also extend to modeling risk factors such as pedestrians and road boundaries. However, these methods are all element-oriented risk modeling approaches, relying on the detection and prediction of individual targets, making it difficult to directly characterize the global risk distribution in the environment. Furthermore, existing path planning methods typically combine MPC, RRT, or artificial potential field methods to generate low-risk trajectories. However, in complex and dynamic traffic scenarios, how to effectively integrate environmental perception results and risk information to serve route planning still requires further research.

[0003] The existing technology has the following drawbacks: 1. Insufficient Global Risk Modeling Capability: At the risk modeling level, existing element-oriented risk modeling is insufficient to characterize the global risk distribution. To address the problem of global risk characterization, some researchers have explored spatial risk field modeling, such as describing the scene as the boundary of the nearest passable area to the vehicle and quantifying the risk, or constructing a planar spatial risk field based on a 2D occupancy grid combined with convolutional operations. However, the former requires a high degree of continuity and accuracy in boundary perception, while the latter is limited to a two-dimensional plane and does not utilize three-dimensional information from environmental perception, making it impossible to achieve global unified risk modeling in three-dimensional space and difficult to form an overall scene risk representation with a continuous spatial structure.

[0004] 2. Significant information loss during the dimensionality reduction of the 3D risk field: At the risk field transformation level, directly applying 3D risk information to path planning leads to computational redundancy and reduced real-time performance. Using simple projection methods such as maximum value or mean value to compress the 3D risk field to 2D ignores the differences in the impact of obstacles of different heights on driving safety, thus losing crucial risk information.

[0005] 3. Path planning struggles to balance multiple performance aspects: At the path planning level, traditional planning algorithms do not effectively integrate risk constraints and trajectory optimization processes, resulting in problems such as low planning efficiency, generated trajectories that tend to cling to obstacles, and discontinuous trajectory curvature. They are unable to balance safety and comfort in complex urban scenarios. Summary of the Invention

[0006] To address the shortcomings of existing technologies, such as insufficient global risk modeling and lack of risk constraint integration in path planning, this invention provides a risk field modeling and path planning method and system based on three-dimensional semantic occupancy, thereby achieving a balance of multiple performance aspects in path planning.

[0007] Therefore, the technical solution adopted by the present invention is as follows: A risk field modeling and path planning method based on three-dimensional semantic occupancy is provided, the method comprising: Obtain the vehicle's driving space and construct a three-dimensional semantic occupancy matrix based on the driving space; Based on the constructed three-dimensional semantic occupancy matrix, a three-dimensional risk field for vehicles is established using a multi-parameter adjustable three-dimensional Gaussian risk kernel function, and the three-dimensional risk field is then converted into a two-dimensional risk field. The vehicle path is segmented and planned according to the two-dimensional risk field to obtain the vehicle driving path. Specifically, the segmented planning divides the drivable domain according to the two-dimensional risk field. A first-stage driving path is generated in the drivable domain by minimizing the fusion distance cost and trajectory point risk cost by a first cost function. The first-stage driving path is then iteratively optimized by gradient descent by minimizing the fusion smoothness cost and global risk cost by a second cost function to obtain the final vehicle driving path.

[0008] According to the above scheme, the three-dimensional semantic occupancy matrix is ​​obtained in the following way: The space within a certain range of the vehicle is obtained as the driving space, and the driving space is divided to obtain a three-dimensional grid set of the driving space; For each 3D grid, occupancy probability and semantic category prediction are performed, and a 3D semantic occupancy matrix is ​​generated based on the prediction results.

[0009] According to the above scheme, the multi-parameter adjustable three-dimensional Gaussian risk kernel function is calculated based on the variance parameters and assignment parameters of any point in the three-dimensional raster in the horizontal and vertical directions. The assignment parameters are adjusted according to the predicted semantic category.

[0010] According to the above scheme, the three-dimensional risk field of the vehicle is obtained in the following way: The 3D semantic occupancy matrix is ​​made continuous using the sigmoid function. The continuous 3D semantic occupancy matrix is ​​then convolved with a multi-parameter adjusted 3D Gaussian risk kernel function to obtain the vehicle's 3D risk field.

[0011] According to the above scheme, the three-dimensional risk field is converted into a two-dimensional risk field by using the probability density function of the Beta distribution as the vertical weighted projection function and projecting the three-dimensional risk field along the height direction; wherein the probability density function of the Beta distribution is calculated based on the maximum vertical height of the scene and the normalization coefficient of the adjustable shape parameter.

[0012] According to the above scheme, a risk threshold is specifically set for the drivable domain, and the area in the two-dimensional risk field with a risk value less than the risk threshold is taken as the drivable domain.

[0013] According to the above scheme, the first cost function is specifically calculated based on the cumulative cost of the vehicle trajectory point from the initial time to the previous time, the Euclidean distance of the vehicle trajectory point from the previous time to the current time, and the two-dimensional risk value of the vehicle trajectory point at the current time.

[0014] According to the above scheme, the smoothness cost is specifically calculated based on the coordinates of the vehicle trajectory points at the previous time, the current time, and the next time; the global risk cost is specifically calculated as the sum of the risk values ​​of all trajectory points on the path trajectory; and the second cost function is specifically calculated as the weighted sum of the smoothness cost and the global risk cost.

[0015] According to the above scheme, the gradient descent iterative optimization specifically calculates the gradient value of the second cost function in each iteration, and calculates the difference between the coordinates of the trajectory point before optimization and the product of the gradient value of the second cost function and the descent gradient step size in each iteration, so as to obtain the coordinates of the trajectory point after optimization. The iteration stops when the termination condition is met.

[0016] A risk field modeling and path planning system based on three-dimensional semantic occupancy is also provided, the system comprising: The matrix construction module is used to obtain the vehicle's driving space and construct a three-dimensional semantic occupancy matrix based on the driving space. The risk field establishment module is used to establish a vehicle three-dimensional risk field based on the constructed three-dimensional semantic occupancy matrix, using a multi-parameter adjustable three-dimensional Gaussian risk kernel function, and convert the three-dimensional risk field into a two-dimensional risk field. The path planning module is used to segment the vehicle path according to the two-dimensional risk field to obtain the vehicle driving path. Specifically, the segment planning is based on dividing the drivable domain according to the two-dimensional risk field. A first-stage driving path is generated in the drivable domain by minimizing the fusion distance cost and trajectory point risk cost of the first-stage driving path. The first-stage driving path is then iteratively optimized by minimizing the fusion smoothness cost and global risk cost of the second-stage driving path to obtain the final vehicle driving path.

[0017] The beneficial effects of this invention are as follows: By constructing a risk field using a three-dimensional semantic occupancy matrix, this invention models the global unified risk probability distribution of various risk elements in three-dimensional space, forming an overall scene risk representation with a continuous spatial structure. Furthermore, in the segmented planning of vehicle paths, distance cost is used in the first-stage path generation to measure the physical movement cost between trajectory point nodes, prioritizing shorter and straighter paths during trajectory point selection. Simultaneously, trajectory point risk cost is incorporated to increase the cost of high-risk trajectory points, prioritizing paths with lower risk values ​​and achieving safety constraints. In addition, in the second-stage optimization, smoothness cost is integrated to perform gradient descent iterative optimization of the initial path in the first stage, eliminating sawtooth oscillations in the initial path, reducing trajectory curvature, and combining global risk cost to ensure a balance between path comfort and safety.

[0018] Furthermore, this invention introduces the probability density function of the Beta distribution as the vertical weighted projection function, and flexibly controls the risk weights at different heights by adjusting the shape parameters. It projects the three-dimensional risk field along the height direction, avoiding the computational redundancy caused by directly using three-dimensional data, and solving the problem that simple projection methods such as maximum value and average value ignore the differences in height risk. This provides a continuous and clearly defined two-dimensional risk input for downstream path planning. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow of the risk field modeling and path planning method based on three-dimensional semantic occupancy according to an embodiment of the present invention; Figure 2 This is an overall flowchart of an embodiment of the present invention; Figure 3 This is a schematic diagram of the system structure of the risk field modeling and path planning system based on three-dimensional semantic occupancy, according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] To address the shortcomings of existing technologies, such as insufficient global unified risk modeling capabilities, severe information loss during 3D risk field dimensionality reduction, and difficulty in balancing various performance aspects in path planning, this invention provides a risk field modeling and path planning method based on 3D semantic occupancy, such as... Figure 1 As shown, the method includes: S1. Obtain the vehicle's driving space and construct a three-dimensional semantic occupancy matrix based on the driving space.

[0022] S2. Based on the constructed three-dimensional semantic occupancy matrix, a three-dimensional risk field for vehicles is established using a multi-parameter adjustable three-dimensional Gaussian risk kernel function, and the three-dimensional risk field is converted into a two-dimensional risk field.

[0023] S3. Based on the two-dimensional risk field, segment the vehicle path to obtain the vehicle driving path.

[0024] In this embodiment, in order to achieve accurate modeling of three-dimensional risks in a three-dimensional scene, the driving space around the vehicle is rasterized to construct a continuous three-dimensional semantic occupancy matrix that includes spatial location, occupancy probability, and semantic category.

[0025] Specifically, based on the actual perception range and risk modeling requirements, the space within a certain range of the vehicle is obtained as the driving space, which can be represented as: ,in , , These represent the relative distance ranges from the center of the vehicle in the horizontal, vertical, and longitudinal directions, respectively.

[0026] The driving space is uniformly divided to obtain a three-dimensional grid set of the driving space, which can be represented as: , Each grid The spatial coordinates of the center point are represented as:

[0027] in, , , These represent the resolution of the raster in three directions.

[0028] Furthermore, the occupancy probability and semantic category prediction for each 3D raster can be expressed as: The probability of occupying semantic category .

[0029] Based on the prediction results, using matrix operators Preliminary construction of a three-dimensional semantic occupancy matrix:

[0030] in, This is the occupancy probability threshold.

[0031] Specifically, in step S2, a multi-parameter adjustable three-dimensional Gaussian risk kernel function... Specifically, calculations are performed based on the variance parameters and assignment parameters of any point in the horizontal and vertical directions of the 3D raster, used to characterize the points located within the raster. The risk at any point in space The effect of this. The kernel function can be defined as:

[0032] in, , These are the variance parameters in the horizontal and vertical directions, respectively, which determine the extent of risk diffusion in the horizontal and vertical directions; amplitude parameters... Based on semantic category adjustment, differentiated expression of risk intensity for different categories of targets is achieved; preferably, in this embodiment, the risk magnitude of dynamic, static, and road three-dimensional risk kernel functions is used. The horizontal decay coefficients are 16, 6, and 20 respectively; for the dynamic and static three-dimensional risk kernel functions. The vertical attenuation coefficients are 1.2 and 1.0, respectively. The values ​​are 0.3 and 0.8 respectively.

[0033] Preferably, the vehicle three-dimensional risk field in this embodiment is obtained in the following manner: The 3D semantic occupancy matrix is ​​made continuous using the sigmoid function. The continuous 3D semantic occupancy matrix is ​​then convolved with a multi-parameter adjusted 3D Gaussian risk kernel function to obtain the vehicle's 3D risk field.

[0034] Specifically, using the sigmoid function to make the 3D semantic occupancy matrix continuous can suppress the spread of false risks in low-occupancy-probability regions. After continuous processing, the 3D semantic occupancy matrix can be updated as follows:

[0035] in, The amplitude coefficient of the sigmoid function is used to achieve an occupation probability exceeding a threshold. The rapid increase in subsequent risks causes risk interactions to skew towards high-occupancy probability regions. In this embodiment, we take... , .

[0036] Furthermore, the calculation process of the vehicle's three-dimensional risk field can be expressed as follows:

[0037] This three-dimensional risk field characterizes the superposition relationship of the influence of risk sources in space. Convolution operations are used to demonstrate the influence of each risk source (obstacle, pedestrian, vehicle, etc.) on the surrounding space within its neighborhood. This influence gradually decays with increasing spatial distance. By summing the risk values ​​of all risk sources across the entire space, the risk value at any location can be obtained. .

[0038] Preferably, in step S2, the conversion of the three-dimensional risk field into a two-dimensional risk field is specifically achieved by using the probability density function of the Beta distribution as the vertical weighted projection function and projecting the three-dimensional risk field along the height direction.

[0039] Specifically, the probability density function of the Beta distribution can be expressed as:

[0040] Three-dimensional risk field along the height direction Projection is performed to achieve projection onto a two-dimensional risk field. The transformation and projection formula is:

[0041] in, This represents the maximum height of the scene in the vertical direction. These are the normalization coefficients; , To adjust the shape parameters and achieve risk weight adjustment for obstacles of different heights, in this embodiment, the following is taken: .

[0042] Preferably, in step S3, to ensure the efficiency and overall safety of path planning, this embodiment of the invention designs a two-stage path planning scheme to meet the driving needs of complex urban scenarios, wherein the first stage uses a two-dimensional risk field. Based on this, a safe and drivable domain is quickly selected, and an initial drivable path is generated by minimizing the first cost function of the improved A* algorithm. This ensures the real-time nature of the planning and overall security.

[0043] Specifically, this embodiment sets a risk threshold. A threshold segmentation method was used to extract the two-dimensional risk field. Medium risk value is lower than The area is designated as a safe driving zone. , will driveable domain boundary As a hard constraint for path planning, it ensures that the generated path always stays within the safe zone.

[0044] Specifically, the first cost function of the improved A* algorithm It can be represented as:

[0045] in, The vehicle trajectory points from the initial moment at the starting point to the previous moment at the parent node. The cumulative cost; From the vehicle trajectory point at the previous moment to the vehicle trajectory point at the current moment The Euclidean distance; Vehicle trajectory point at the current moment Two-dimensional risk intensity at the location; This is a risk cost scaling factor, which is taken in this embodiment. This cost function proactively penalizes high-risk areas during path search, generating an initial drivable path. .

[0046] The first cost function in this embodiment uses cumulative cost to ensure the global optimality of path search. This allows the algorithm to consider not only the cost of the current step but also previous steps when expanding trajectory point nodes, avoiding shortsighted paths that fall into local optima and ensuring that the final generated path is the path with the lowest total cost from the starting point to the current node. At the same time, the Euclidean distance between the vehicle trajectory point at the previous moment and the vehicle trajectory point at the current moment is used to measure the physical movement cost between nodes, prompting the algorithm to prioritize shorter and straighter paths and avoid generating detours and tortuous trajectories. In addition, the risk cost of a single trajectory point is multiplied by the distance cost in the form of a coefficient, which is equivalent to increasing the unit distance cost of high-risk road segments and setting a "penalty tax" for paths in high-risk areas. This causes the algorithm to actively avoid areas with high R(n) when expanding nodes, achieving path safety constraints in the first stage. Furthermore, due to the existence of the risk penalty term, the algorithm will actively avoid high-risk areas during path search, reducing the number of expansions of nodes in these areas, reducing the number of invalid node traversals, and improving search efficiency in complex scenarios.

[0047] Furthermore, to ensure that the path planning output is a drivable trajectory that balances comfort and safety, the two-stage path planning scheme designed in this embodiment of the invention generates the final optimal path in the second stage through trajectory optimization based on gradient descent. This stage uses the initial drivable path... As an initial solution, a second cost function is constructed to minimize the smoothness cost and the global risk cost through multi-objective optimization. This function is then iteratively optimized using the gradient descent algorithm to generate a final trajectory that is both safe and smooth. .

[0048] Specifically, the second cost function It can be represented as:

[0049] in, As a trade-off for smoothness, it is used to measure the continuous differentiability of the trajectory, avoid broken lines and turning points, and ensure the smooth driving of the vehicle. This serves as the overall risk cost, used to measure the overall risk level of the trajectory and ensure that the trajectory stays away from high-risk areas. The cost-weighted coefficient is used to adjust the weights for smoothness and safety. In this embodiment, it is taken as... .

[0050] Smoothness typically requires the trajectory to physically correspond to the continuity of vehicle acceleration and its rate of change, directly impacting ride comfort. Mathematically, this requirement is often expressed as a constraint on the amplitude of the second or third derivative of the trajectory curve. This can be achieved by minimizing the sum of squares of the corresponding derivatives in the optimization objective. Consider a continuous trajectory. ,in For normalized trajectory parameters, the smoothness cost It can be represented as:

[0051] However, due to the initial drivable path Since it is a set of discrete trajectory points, continuous integrals cannot be directly calculated. Therefore, the continuous cost needs to be discretized through difference approximation. In this embodiment, let the discrete trajectory points be a set of discrete trajectory points. One, denoted as The second derivative can be approximated by the difference between three adjacent points, and the smoothness cost after discretization is... Updatable is represented as:

[0052] By using the smooth cost calculation of continuous cost discretization through differential approximation, the turning rate and turning amplitude of the overall driving path are optimized, avoiding the poor driving experience caused by large and frequent turns, and improving the smoothness and comfort of the driving path.

[0053] The global risk cost measures the total risk of all points on the trajectory, requiring the risk integral of the continuous trajectory to be converted into the sum of risks at discrete points. (Risk cost of a continuous trajectory) The integral of the risk value at each point on the trajectory:

[0054] Specifically, gradient descent iterative optimization calculates the gradient value of the second cost function in each iteration, and calculates the difference between the coordinates of the trajectory point before optimization and the product of the gradient value of the second cost function and the descent gradient step size in each iteration, to obtain the coordinates of the trajectory point after optimization. The iteration stops when the termination condition is met.

[0055] The gradient of the second cost function can be expressed as:

[0056]

[0057] The optimized trajectory point coordinates can be represented as:

[0058]

[0059] in, The iteration number is [number], and the iteration continues until the total cost function is reached. convergence, For a fixed update step size in gradient descent, this embodiment takes... The final optimal driving trajectory is obtained. .

[0060] In this embodiment, the termination condition is that the number of iterations reaches a certain number.

[0061] Specifically, the overall flowchart of this embodiment is as follows: Figure 2 As shown.

[0062] Furthermore, embodiments of the present invention also provide a risk field modeling and path planning system based on three-dimensional semantic occupancy, used to implement the risk field modeling and path planning method based on three-dimensional semantic occupancy of the embodiments of the present invention, such as... Figure 3 As shown, the system includes: The matrix construction module is used to obtain the vehicle's driving space and construct a three-dimensional semantic occupancy matrix based on the driving space. The risk field establishment module is used to establish a vehicle three-dimensional risk field based on the constructed three-dimensional semantic occupancy matrix, using a multi-parameter adjustable three-dimensional Gaussian risk kernel function, and convert the three-dimensional risk field into a two-dimensional risk field. The path planning module is used to segment the vehicle path according to the two-dimensional risk field to obtain the vehicle driving path. Specifically, the segmented planning divides the drivable domain according to the two-dimensional risk field. A first-stage driving path is generated in the drivable domain by minimizing the first cost function that integrates the distance cost between each trajectory point and the risk cost of the trajectory point. The first-stage driving path is then iteratively optimized by gradient descent by minimizing the second cost function that integrates the overall smoothness cost of the fused path and the global risk cost to obtain the final vehicle driving path.

[0063] The various modules or mechanisms of the system are mainly used to implement the various steps of the above method embodiments, and will not be described in detail here.

[0064] This invention constructs a risk field using a three-dimensional semantic occupancy matrix, globally unifying the risk probability distribution model for various risk elements in three-dimensional space. This forms an overall scene risk representation with a continuous spatial structure. Furthermore, during the segmented planning of vehicle paths, distance cost is used in the first-stage path generation to measure the physical movement cost between trajectory point nodes, prioritizing shorter and straighter paths. Simultaneously, trajectory point risk cost is incorporated to increase the cost of high-risk trajectory points, prioritizing paths with lower risk values ​​and achieving safety constraints. In the second-stage optimization, smoothness cost is integrated to iteratively optimize the initial path using gradient descent, eliminating sawtooth oscillations and reducing trajectory curvature. Combined with global risk cost, this ensures a balance between path comfort and safety.

[0065] Furthermore, this embodiment of the invention introduces the probability density function of the Beta distribution as the vertical weighted projection function, and flexibly controls the risk weights at different heights by adjusting the shape parameters. The three-dimensional risk field is projected along the height direction, avoiding the computational redundancy caused by directly using three-dimensional data, and solving the problem that simple projection methods such as maximum value and average value ignore the differences in height risk. This provides a continuous and clearly defined two-dimensional risk input for downstream path planning.

[0066] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0067] The sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0068] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A risk field modeling and path planning method based on three-dimensional semantic occupancy, characterized in that, The method includes: Obtain the vehicle's driving space and construct a three-dimensional semantic occupancy matrix based on the driving space; Based on the constructed three-dimensional semantic occupancy matrix, a three-dimensional risk field for vehicles is established using a multi-parameter adjustable three-dimensional Gaussian risk kernel function, and the three-dimensional risk field is then converted into a two-dimensional risk field. The vehicle path is segmented and planned according to the two-dimensional risk field to obtain the vehicle driving path. Specifically, the segmented planning divides the drivable domain according to the two-dimensional risk field. A first-stage driving path is generated in the drivable domain by minimizing the first cost function that integrates the distance cost between each trajectory point and the risk cost of the trajectory point. The first-stage driving path is then iteratively optimized by gradient descent by minimizing the second cost function that integrates the overall smoothness cost and the global risk cost of the fused path to obtain the final vehicle driving path.

2. The risk field modeling and path planning method based on three-dimensional semantic occupancy according to claim 1, characterized in that, The three-dimensional semantic occupancy matrix is ​​obtained in the following way: The space within a certain range of the vehicle is obtained as the driving space, and the driving space is divided to obtain a three-dimensional grid set of the driving space; For each 3D grid, occupancy probability and semantic category prediction are performed, and a 3D semantic occupancy matrix is ​​generated based on the prediction results.

3. The risk field modeling and path planning method based on three-dimensional semantic occupancy according to claim 2, characterized in that, The multi-parameter adjustable 3D Gaussian risk kernel function is calculated based on the variance parameters and assignment parameters of any point in the 3D raster in the horizontal and vertical directions. The assignment parameters are adjusted according to the predicted semantic category.

4. The risk field modeling and path planning method based on three-dimensional semantic occupancy according to claim 1, characterized in that, The vehicle's three-dimensional risk field is obtained in the following way: The 3D semantic occupancy matrix is ​​made continuous using the sigmoid function. The continuous 3D semantic occupancy matrix is ​​then convolved with a multi-parameter adjusted 3D Gaussian risk kernel function to obtain the vehicle's 3D risk field.

5. The risk field modeling and path planning method based on three-dimensional semantic occupancy according to claim 1, characterized in that, The conversion of a 3D risk field into a 2D risk field is achieved by using the probability density function of the Beta distribution as the vertical weighted projection function and projecting the 3D risk field along the height direction. The probability density function of the Beta distribution is calculated based on the maximum vertical height of the scene and the normalization coefficient of the adjustable shape parameters.

6. The risk field modeling and path planning method based on three-dimensional semantic occupancy according to claim 1, characterized in that, The drivable domain is specifically defined by a risk threshold, and the area in the two-dimensional risk field with a risk value less than the risk threshold is defined as the drivable domain.

7. The risk field modeling and path planning method based on three-dimensional semantic occupancy according to claim 1, characterized in that, The first cost function is specifically calculated based on the cumulative cost of the vehicle trajectory point from the initial time to the previous time, the Euclidean distance of the vehicle trajectory point from the previous time to the current time, and the two-dimensional risk value of the vehicle trajectory point at the current time.

8. The risk field modeling and path planning method based on three-dimensional semantic occupancy according to claim 1, characterized in that, The smoothness cost is specifically calculated based on the coordinates of the vehicle trajectory points at the previous time step, the current time step, and the next time step; the global risk cost is specifically calculated as the sum of the risk values ​​of all trajectory points on the path trajectory; the second cost function is specifically calculated as the weighted sum of the smoothness cost and the global risk cost.

9. The risk field modeling and path planning method based on three-dimensional semantic occupancy according to claim 1, characterized in that, The gradient descent iterative optimization specifically calculates the gradient value of the second cost function in each iteration, and calculates the difference between the coordinates of the trajectory point before optimization and the product of the gradient value of the second cost function and the descent gradient step size in each iteration, to obtain the coordinates of the trajectory point after optimization. The iteration stops when the termination condition is met.

10. A risk field modeling and path planning system based on three-dimensional semantic occupancy, characterized in that, The system includes: The matrix construction module is used to obtain the vehicle's driving space and construct a three-dimensional semantic occupancy matrix based on the driving space. The risk field establishment module is used to establish a vehicle three-dimensional risk field based on the constructed three-dimensional semantic occupancy matrix, using a multi-parameter adjustable three-dimensional Gaussian risk kernel function, and convert the three-dimensional risk field into a two-dimensional risk field. The path planning module is used to segment the vehicle path according to the two-dimensional risk field to obtain the vehicle driving path. Specifically, the segmented planning is based on dividing the drivable domain according to the two-dimensional risk field. A first-stage driving path is generated in the drivable domain by minimizing the first cost function that integrates the distance cost between each trajectory point and the risk cost of the trajectory point. The first-stage driving path is then iteratively optimized by gradient descent by minimizing the second cost function that integrates the overall smoothness cost of the fused path and the global risk cost to obtain the final vehicle driving path.

Images