A method for estimating the drivability of off-road terrain under conditions of motion uncertainty

By employing single-frame point cloud distribution and a self-supervised learning network in off-road terrain to update the feature matrix in real time, the accuracy problem of off-road terrain accessibility estimation under motion uncertainty is solved, achieving more stable and accurate off-road terrain accessibility estimation.

CN121121678BActive Publication Date: 2026-11-10BEIJING INST OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202511232754.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-09-01
Publication Date
2026-11-10
Estimated Expiration
2045-09-01

AI Technical Summary

Technical Problem

Under conditions of motion uncertainty, existing off-road terrain accessibility estimation methods suffer from reduced point cloud stitching accuracy and feature extraction reliability due to pose estimation errors, thus affecting the accuracy of off-road terrain accessibility estimation.

Method used

Feature extraction is performed using single-frame point cloud distribution. By utilizing information from LiDAR and inertial measurement units, and through a self-supervised learning network, the feature matrix is ​​updated in real time, reducing the impact of pose estimation errors on feature extraction and improving the accuracy of toll cost prediction.

Benefits of technology

It improves the stability and accuracy of off-road terrain accessibility estimation under motion uncertainty conditions, reduces the impact of attitude estimation error on feature extraction, and avoids point cloud stitching error.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121121678B_ABST
    Figure CN121121678B_ABST
Patent Text Reader

Abstract

This invention provides a method for estimating the drivability of off-road terrain under motion uncertainty conditions. This method is based on the extraction of features from the distribution and reflection characteristics of single-frame point clouds, thus reducing the impact of attitude estimation errors, avoiding point cloud stitching errors under motion uncertainty conditions, improving the reliability of feature extraction, and enhancing the stability and accuracy of off-road terrain drivability estimation under motion uncertainty conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of intelligent vehicle technology, and specifically relates to a method for estimating off-road terrain drivability under motion uncertainty conditions. Background Technology

[0002] Off-road terrain accessibility estimation is a fundamental prerequisite for the safe operation of Unmanned Ground Vehicles (UGVs) in off-road terrain, directly impacting their mission performance. Using sensor information, the geometric features and physical properties of off-road terrain can be evaluated, generating a access cost matrix. This provides accurate accessibility risk warnings for UGV trajectory planning and motion control, enabling them to autonomously avoid high-risk areas. Therefore, researchers have conducted extensive research on off-road terrain accessibility estimation based on sensor perception information. The research approach has evolved from early heuristic estimation methods based on geometric features to an end-to-end accessibility estimation framework based on self-supervised learning.

[0003] The self-supervised learning-based off-road terrain accessibility estimation method is based on constructing a dataset that uses environmental perception information from external sensors such as LiDAR as input and UGV motion states observed by ontological sensors such as Inertial Measurement Units (IMUs) as labels. Then, based on this dataset and considering the interaction between the ground and the platform, a neural network is used to autonomously learn the mapping relationship between the terrain traversed by the UGV and the UGV's motion state.

[0004] Existing self-supervised learning methods for estimating off-road terrain accessibility typically stitch together multiple frames of point clouds acquired by LiDAR based on the pose transformation matrix of the UGV during its movement, and extract statistical features from the point clouds as input to the accessibility cost prediction network. However, UGVs traveling at high speeds in off-road terrain are inevitably affected by road surface unevenness, resulting in significant dynamic bumps. This bumping introduces motion uncertainty, which in turn increases pose estimation error. Pose estimation error further affects the accuracy of the pose transformation matrix, thereby reducing the accuracy of point cloud stitching and the reliability of point cloud statistical feature extraction, ultimately impacting the accuracy of off-road terrain accessibility estimation. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a method for estimating the drivability of off-road terrain under motion uncertainty conditions. This method uses the distribution of a single-frame point cloud for feature extraction to reduce the impact of pose estimation error on the reliability of feature extraction, improve the accuracy of drivability cost prediction under motion uncertainty conditions, and achieve stable and accurate off-road terrain drivability estimation.

[0006] A method for estimating off-road terrain drivability under motion uncertainty conditions includes:

[0007] Step 1: Install a lidar on the ground unmanned platform UGV to collect lidar point cloud information in real time, and install an inertial measurement unit to collect the position and attitude information of the unmanned platform UGV in real time.

[0008] Step 2: Define the local coordinate system of the UGV (Unmanned Ground Vehicle) platform. The XY plane is then discretized into a raster; a feature matrix Ω is defined, where each element is the eigenvalue ω of each raster. u,v =(ρ u,v ,I u,v ), where ρ u,v I represents the roughness feature value of the raster with index (u,v). u,v This represents the reflectance intensity characteristic value of the raster with index (u,v);

[0009] Step 3: Define the radar coordinate system

[0010] Step 4: After the lidar receives a new frame of point cloud, it performs a structured representation of the point cloud in the lidar coordinate system.

[0011] Step 5: Extract features from the point cloud in the radar coordinate system, including the roughness and reflection intensity of the laser points;

[0012] Step 6: Calculate the point cloud P in the k-th frame based on the pose information. k The coordinates in the local coordinate system are represented as B. k ;

[0013] Step 7: Based on coordinate B k Determine the index of each laser point in the discrete grid system;

[0014] Step 8: Update the feature matrix Ω based on the roughness and reflection intensity extracted from the point cloud of the k-th frame in Step 4;

[0015] Step 9: Based on the position and attitude information of the ground unmanned platform UGV at time k, determine the grid index occupied by the UGV in the XY plane;

[0016] Step 10: Use the current driving speed and the feature value of the occupied grid of the ground unmanned platform UGV as input data for the self-supervised learning network dataset;

[0017] Step 11: Use the passage cost of the ground unmanned platform UGV at each moment as the label of the self-supervised learning network dataset;

[0018] Step 12: Repeat steps 4 to 11 to obtain a self-supervised learning dataset during the operation of the UGV (Unmanned Ground Vehicle) platform.

[0019] Step 13: Train the self-supervised learning network using the dataset;

[0020] Step 14: During the UGV's operation on the ground, the trained self-supervised learning network is used to predict the passage cost of its planned trajectory, thereby estimating the drivability of off-road terrain.

[0021] Preferably, in step 2, a local coordinate system for the ground-based unmanned platform (UGV) is defined. The initial position of the ground unmanned platform UGV is selected as the origin O. L The Z-axis points upwards along the direction of gravity, the XY plane is perpendicular to the Z-axis, and the X and Y axes are orthogonal to each other. The X-axis direction is aligned with the initial motion direction of the ground-based unmanned platform (UGV). In step 3, a radar coordinate system is defined. The location of the lidar is selected as the origin O of the coordinate system. The installation plane of the lidar is parallel to the motion plane of the ground unmanned platform UGV. The x-axis points directly in front of the UGV and is parallel to the roll axis of the UGV. The z-axis is vertically upward and parallel to the yaw axis of the ground unmanned platform UGV. The y-axis is determined according to the right-hand rule.

[0022] Preferably, step 4 specifically includes: after the lidar receives a new frame of point cloud, let the position coordinates of the collected point cloud in the lidar coordinate system at time k be represented as a matrix. Where m represents the number of LiDAR scan lines, and n represents the maximum number of sampling points on a single scan line;

[0023] In the radar coordinate system, the coordinates of any laser point in the point cloud matrix of the k-th frame are represented as follows: Where i∈[1,m] represents the number of the i-th scan line, and j∈[1,n] represents the j-th sampling point on the scan line;

[0024] If the lidar is located at the origin (0,0,0), then the distance from each point in the point cloud to the lidar is expressed as: In the formula, ||·|| represents the second norm of the vector.

[0025] Preferably, in step 5, for a certain laser point Its roughness The calculation formula is as follows:

[0026]

[0027] Where l represents the laser point The number of adjacent points on the same scan line.

[0028] Preferably, step 8 is specifically as follows:

[0029] Step 8.1: Based on the index of each laser point in the point cloud within the discrete grid system, count the laser observation points corresponding to each grid in the discrete grid system; calculate the average roughness of these laser observation points. and average reflection intensity That is, the point cloud feature observation value of the kth frame within this grid;

[0030] Step 8.2: If the original raster feature value is 0, then update the raster feature value. Otherwise, use Kalman filtering to update the raster feature values.

[0031] Preferably, step 9 specifically includes the following methods:

[0032] Step 9.1: Assume that the ground area occupied by the ground unmanned platform UGV is represented as a rectangular area S with a length of a and a width of b; according to the measurement information of the inertial measurement unit, the position of UGV in the local coordinate system at time k is the center of the rectangular area S, and the roll axis direction of UGV in the attitude of the local coordinate system at time k is the direction of the long side of the rectangular area S.

[0033] Step 9.2: Divide the rectangular region S into grids along the long and short sides with a resolution of γ, and calculate the coordinates of the center points of each grid obtained from the division of the rectangular region S in the local coordinate system. Based on these coordinates, determine the index of each grid center point in the discrete grid system of the local coordinate system. The set of these indices is the grid index occupied by UGV in the XY plane of the local coordinate system at time k.

[0034] Preferably, step 11 specifically includes:

[0035] Based on M predetermined time points, the travel cost of the UGV (Unmanned Ground Vehicle) navigating the current terrain is calculated as a label for self-supervised learning. The calculation formula is as follows:

[0036]

[0037] Where, ω y (k), ω x (k) and a z (k) represents the absolute values ​​of pitch angular velocity, roll angular velocity, and longitudinal acceleration of the ground unmanned platform UGV at time k of the inertial measurement unit strategy, respectively; c1, c2, and c3 are normalized parameters.

[0038] Preferably, in step 13, when training the self-supervised learning network, the mean squared error (MSE) is used as the loss function for training.

[0039]

[0040] in, For the h-th feature data in the training set, The corresponding UGV (Unmanned Ground Vehicle) speed is input. C represents the output of a self-supervised learning network. h H represents the corresponding passage cost, and H represents the number of data points in the training set.

[0041] Preferably, step 14 is performed as follows:

[0042] Step 14.1: Repeat steps 4 to 8 during the operation of the UGV (Unmanned Ground Vehicle) to update the feature matrix Ω in the discrete grid system.

[0043] Step 14.2: Based on the trajectory planned by the UGV (Unmanned Ground Vehicle) platform, determine the position and attitude information at the future time k'. Extract the feature value G according to steps 9 and 10. k' And combined with the planned driving speed v k' The passage cost f at this location is predicted using a self-supervised learning network. C (G k' ,v k' By predicting the access cost at various locations in off-road terrain, the accessibility of off-road terrain is estimated.

[0044] The present invention has the following beneficial effects:

[0045] This invention provides a method for estimating the drivability of off-road terrain under motion uncertainty conditions. This method extracts features based on the distribution and reflection characteristics of single-frame point clouds, thus reducing the impact of attitude estimation errors, avoiding point cloud stitching errors under motion uncertainty conditions, improving the reliability of feature extraction, and enhancing the stability and accuracy of off-road terrain drivability estimation under motion uncertainty conditions. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention.

[0047] Figure 2 This is a schematic diagram of the unmanned ground platform used in this invention.

[0048] Figure 3 This is a schematic diagram of the local coordinate system and the radar coordinate system in this invention.

[0049] Figure 4 This is a schematic diagram of how the UGV occupies the grid in the XY plane of the local coordinate system in this invention. Detailed Implementation

[0050] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0051] In this embodiment, the following is adopted: Figure 2 The SCOUT MINI 2.0 ground unmanned platform shown is shown.

[0052] Step 1: Install a velodyne 16-line lidar and an inertial measurement unit (IMU) on the SCOUT MINI 2.0 unmanned ground vehicle (UGV) and collect lidar information (point cloud) and IMU measurement information (position and attitude) in real time.

[0053] Step 2: Define the local coordinate system The initial position of the UGV is selected as the origin O. L The Z-axis points upwards along the direction of gravity, and the XY plane is perpendicular to the Z-axis. The X and Y axes are orthogonal to each other. Generally, the X-axis direction is aligned with the initial motion direction of the UGV. Figure 3 As shown.

[0054] Define a discrete grid system in the XY plane in the local coordinate system, setting the grid resolution to δ = 0.05 m, and the position of each axial boundary of the grid to δ*N. max In the formula N max =10000 represents the maximum number of grid cells along a single axis. Therefore, for any point (X,Y,Z) in the local coordinate system, its grid index (u,v) can be represented as:

[0055]

[0056] In the formula This indicates rounding down to the nearest integer.

[0057] In a discrete grid system, define the characteristic matrix. The elements in the matrix can be represented as raster eigenvalues ​​ω. u,v =(ρ u,v ,I u,v ), where ρ u,v I represents the roughness feature value of the raster with index (u,v). u,v This represents the reflectance intensity characteristic value of the raster at index (u,v). The initial values ​​of the characteristic matrix are set to NULL.

[0058] Step 3: Define the radar coordinate system The lidar position is selected as the origin O of the coordinate system. The lidar mounting plane is parallel to the UGV's motion plane. The x-axis points directly forward of the UGV and is parallel to the UGV's roll axis. The z-axis is perpendicular upward and parallel to the UGV's yaw axis. The y-axis direction is determined according to the right-hand rule. Figure 3 As shown.

[0059] Step 4: After receiving a new frame of point cloud, the lidar performs a structured representation of the original point cloud in the radar coordinate system. Let the acquired point cloud (the k-th frame point cloud) and its position coordinates in the radar coordinate system at time k be represented as a matrix. Where m represents the number of LiDAR scan lines and n represents the maximum number of sampling points on a single scan line. In this example, m = 16 and n = 1800.

[0060] Then, in the radar coordinate system, the coordinates of any laser point in the point cloud matrix of the k-th frame can be expressed as: Where i∈[1,m] represents the number of the i-th scan line, and j∈[1,n] represents the j-th sampling point on the scan line.

[0061] Since the lidar is located at the origin (0,0,0), the distance from each point in the point cloud to the lidar can be expressed as: In the formula, ||·|| represents the second norm of the vector.

[0062] Step 5: Extract features from the point cloud in the radar coordinate system. First, calculate the roughness of each laser point in the point cloud of this frame as a feature. For a given point... Its roughness The calculation formula is as follows:

[0063]

[0064] Where l represents the laser point The system collects l adjacent points on the same scan line (l = 10 in this example). Then, it collects the reflection intensity of each laser point in the point cloud of that frame. As another feature.

[0065] Step 6: Calculate the point cloud P in the k-th frame based on the pose information. k Coordinates in the local coordinate system are represented as The coordinates of any laser point in the point cloud in the local coordinate system can be expressed as: Where i∈[1,m] represents the number of the i-th scan line, and j∈[1,n] represents the j-th sampling point on the scan line.

[0066] Step 7: Based on the coordinates B of the point cloud in the local coordinate system of the k-th frame. k The index of each laser point in the point cloud in the discrete grid system is determined.

[0067] Step 8: Update the feature matrix Ω within the discrete raster system based on the features extracted from the point cloud of frame k. The specific steps are as follows:

[0068] Step 8.1 Calculate the observed values ​​of the features extracted from the point cloud in the k-th frame within the discrete grid system. Based on the indices of each laser point in the point cloud within the discrete grid system, count the laser observation points corresponding to each grid in the discrete grid system. Calculate the average roughness and reflection intensity of these laser points. and This refers to the point cloud feature observation value of the kth frame within the grid.

[0069] Step 8.2 Update the raster feature values ​​for points with laser observations. If the original raster feature value is (NULL, NULL), then the updated raster feature value... Otherwise, use Kalman filtering to update the raster feature values.

[0070] Step 9: Based on the position and attitude information of the UGV measured by the IMU at time k, determine the grid index occupied by the UGV in the XY plane of the local coordinate system at time k, such as... Figure 4 As shown. The specific method is as follows:

[0071] Step 9.1 Assume the ground area occupied by the UGV can be represented as a rectangular region S, with a length a = 0.6m (the longer side is defined as along the vehicle's forward direction) and a width b = 0.5m (the shorter side is defined as perpendicular to the vehicle's forward direction). According to the IMU's measurement information, the position of the UGV in the local coordinate system at time k is the center of the rectangular region S, and the roll axis direction of the UGV in the local coordinate system at time k is the direction of the longer side of the rectangular region S.

[0072] Step 9.2 The rectangular region S is divided into grids along its long and short sides with a resolution γ = 0.05 m. The coordinates of the center points of each grid obtained from the division of the rectangular region S in the local coordinate system are calculated. Based on these coordinates, the index of each grid center point in the discrete raster system of the local coordinate system is determined. The set of these indices is the raster index occupied by UGV in the XY plane of the local coordinate system at time k, which can be expressed as: Each element This represents the two-dimensional coordinate index of the m-th occupied grid cell.

[0073] Step 10: Determine the input data for the self-supervised learning dataset at time k.

[0074] Step 10.1 Calculate the current driving speed of the UGV based on the measurement information from the IMU, and use it as one of the network inputs for self-supervised learning.

[0075] Step 10.2 Based on the grid index occupied by UGV in the XY plane of the local coordinate system at time k. The feature values ​​of the UGV-occupied grids are obtained and used as one of the network inputs for self-supervised learning. If the feature value of an occupied grid is (NULL, NULL), the Bayesian generalized kernel inference method is used to complete the feature value of the grid based on the feature values ​​of its surrounding grids.

[0076] Step 11: Determine the labels of the self-supervised learning dataset at time k.

[0077] Based on the absolute value ω of the UGV pitch angular velocity measured by the IMU at the previous M time points y absolute value of roll angular velocity ω x Longitudinal acceleration a z The sequence (minus the gravity component) is used to calculate the travel cost of the UGV in the current terrain as a label for self-supervised learning. The calculation formula is as follows:

[0078]

[0079] Where c1, c2, and c3 are normalization parameters, respectively, for ω x ,ω y ,a z Normalize to the [0,1] interval.

[0080] Step 12: Repeat steps 4 to 11 to obtain the self-supervised learning dataset during the operation of the ground-based unmanned platform. This dataset includes two types of information: input data and labels. The input data consists of the UGV's driving speed and the feature values ​​of the grid cells occupied by the UGV at each time step; the labels are the UGV's passage cost at each time step.

[0081] Step 13: After obtaining the self-supervised learning dataset, use this dataset to train the prediction network. During training, use the Mean Squared Error (MSE) as the loss function.

[0082]

[0083] in, For the h-th feature data in the training set, The corresponding UGV (Unmanned Ground Vehicle) speed is input. C represents the output of a self-supervised learning network. h Let H be the corresponding passage cost, and H be the number of data points in the training set. As the number of iterations increases, when the value of the loss function gradually stabilizes and the performance of the prediction network no longer improves significantly, the convergent prediction network f can be obtained. C (·)

[0084] Step 14: During the UGV's operation, predict the travel cost of its planned trajectory to estimate the drivability of off-road terrain. The specific method is as follows:

[0085] Step 14.1 Repeat steps 4 to 8 during UGV operation to update the feature matrix Ω within the discrete grid system.

[0086] Step 14.2 Based on the trajectory planned by the UGV, determine the position and attitude information of the UGV at the future time k'. Extract the input G according to steps 9 and 10. k' And combined with the planned driving speed v k' This allows for real-time prediction of the passage cost f at that location. C (G k' ,v k' By predicting the access cost at various locations in off-road terrain, the accessibility of off-road terrain can be estimated.

[0087] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for estimating the drivability of off-road terrain under conditions of motion uncertainty, characterized in that, include: Step 1: Install a lidar on the ground unmanned platform UGV to collect lidar point cloud information in real time, and install an inertial measurement unit to collect the position and attitude information of the unmanned platform UGV in real time. Step 2: Define the local coordinate system of the UGV (Unmanned Ground Vehicle) platform. The initial position of the UGV is selected as the origin of the coordinate system. , The axis points upward along the direction of gravity. plane perpendicular to Axial direction, and shaft and The axes are orthogonal to each other, let The axial direction is consistent with the initial motion direction of the UGV; and for Discretize the plane into a raster; define the feature matrix. The elements in the matrix are the eigenvalues ​​of each grid cell. ,in Indicates that the index is The roughness characteristic value of the raster. Indicates that the index is The characteristic value of the reflection intensity of the grid; Step 3: Define the radar coordinate system The location of the lidar is selected as the origin of the coordinate system. O The radar mounting plane is parallel to the UGV's motion plane, the x-axis points directly in front of the UGV and is parallel to the UGV's roll axis, the z-axis is vertically upward and parallel to the UGV's yaw axis, and the y-axis is determined by the right-hand rule. Step 4: After the lidar receives a new frame of point cloud, it performs a structured representation of the point cloud in the lidar coordinate system. Step 5: Extract features from the point cloud in the radar coordinate system, including the roughness and reflection intensity of the laser points; Step 6: Calculate the point cloud of the k-th frame based on the pose information. Coordinates in the local coordinate system are represented as ; Step 7: Based on coordinates Determine the index of each laser point in the discrete grid system; Step 8: Based on the roughness and reflection intensity extracted from the point cloud of the k-th frame in Step 4, process the feature matrix. Update; Step 9, according to the first t k Real-time position and attitude information of the ground-based unmanned platform (UGV) is used to determine the UGV's position and attitude. The grid index occupied within the plane; Step 10: Use the current driving speed and the feature value of the occupied grid of the ground unmanned platform UGV as input data for the self-supervised learning network dataset; Step 11: Use the passage cost of the ground unmanned platform UGV at each moment as the label of the self-supervised learning network dataset; Step 12: Repeat steps 4 to 11 to obtain a self-supervised learning dataset during the operation of the UGV (Unmanned Ground Vehicle) platform. Step 13: Train the self-supervised learning network using the dataset; Step 14: During the UGV's operation on the ground, the trained self-supervised learning network is used to predict the passage cost of its planned trajectory, thereby estimating the drivability of off-road terrain.

2. The method for estimating off-road terrain drivability under motion uncertainty conditions as described in claim 1, characterized in that, In step 2, a local coordinate system for the ground unmanned platform UGV is defined. The initial position of the ground unmanned platform UGV is selected as the origin of the coordinate system. , The axis points upward along the direction of gravity. plane perpendicular to Axial direction, and shaft and The axes are orthogonal to each other, let The axial direction is consistent with the initial motion direction of the ground unmanned platform UGV; in step 3, the radar coordinate system is defined. The location of the lidar is selected as the origin of the coordinate system. O The radar mounting plane is parallel to the motion plane of the ground unmanned platform UGV. The x-axis points directly in front of the UGV and is parallel to the UGV's roll axis. The z-axis is vertically upward and parallel to the yaw axis of the ground unmanned platform UGV. The y-axis is determined according to the right-hand rule.

3. The method for estimating off-road terrain drivability under motion uncertainty conditions as described in claim 2, characterized in that, Step 4 specifically includes: After the lidar receives a new frame of point cloud, let the... t k At any given time, the position coordinates of the collected point cloud in the radar coordinate system are represented as a matrix. ,in m This indicates the number of laser radar scan lines. n This indicates the maximum number of sampling points on a single scan line; In the radar coordinate system, the first k The coordinates of any laser point in the frame point cloud matrix are represented as follows: ,in Indicates the first i Scan line number, Indicates the first scan line. j One sampling point; If the lidar is located at the origin (0,0,0), then the distance from each point in the point cloud to the lidar is expressed as: In the formula The L2 norm of a vector.

4. The method for estimating off-road terrain drivability under motion uncertainty conditions as described in claim 3, characterized in that, In step 5, for a certain laser point Its roughness The calculation formula is as follows: in, Indicates the laser point The number of adjacent points on the same scan line.

5. The method for estimating off-road terrain drivability under motion uncertainty conditions as described in claim 4, characterized in that, The specific steps of step 8 are as follows: Step 8.1: Based on the index of each laser point in the point cloud within the discrete grid system, count the laser observation points corresponding to each grid in the discrete grid system; calculate the average roughness of these laser observation points. and average reflection intensity That is, the first cell within that grid. k Frame point cloud feature observations; Step 8.2: If the original raster feature value is 0, then update the raster feature value. Otherwise, use Kalman filtering to update the raster feature values.

6. The method for estimating off-road terrain drivability under motion uncertainty conditions as described in claim 5, characterized in that, The specific method for step 9 includes: Step 9.1: Assume the ground area occupied by the unmanned ground platform UGV is represented as a rectangular region S, with a length of... a Width is b Based on the measurement information from the inertial measurement unit, the first... t k The position of UGV in the local coordinate system at time t is the center of the rectangular region S. t k At any given moment, the roll axis direction of the UGV in the local coordinate system is the direction of the long side of the rectangular region S. Step 9.2: Divide the rectangular region S into a grid along its long and short sides, with a resolution of [resolution missing]. The coordinates of the center points of each grid obtained by dividing the rectangular region S in the local coordinate system are calculated. Based on these coordinates, the index of each grid center point in the discrete raster system of the local coordinate system is determined. The set of these indices is the... t k UGV at time in local coordinate system The grid index occupied in the plane.

7. The method for estimating off-road terrain drivability under motion uncertainty conditions as described in claim 6, characterized in that, Step 11 specifically includes: According to the settings M At each time point, the travel cost of the UGV (Unmanned Ground Vehicle) navigating the current terrain is calculated as a label for self-supervised learning. The calculation formula is as follows: in, , and These represent the strategies of the inertial measurement unit. t k The absolute values ​​of pitch angular velocity, roll angular velocity, and longitudinal acceleration of the ground unmanned platform UGV at all times; These are the normalized parameters.

8. The method for estimating off-road terrain drivability under motion uncertainty conditions as described in claim 7, characterized in that, When training the self-supervised learning network in step 13, the mean squared error (MSE) is used as the loss function for training. in, For the h-th feature data in the training set, The corresponding UGV (Unmanned Ground Vehicle) speed is input. This represents the output of a self-supervised learning network. H represents the corresponding passage cost, and H represents the number of data points in the training set.

9. The method for estimating off-road terrain drivability under motion uncertainty conditions as described in claim 8, characterized in that, The specific method for step 14 is as follows: Step 14.1: During the operation of the UGV (Unmanned Ground Vehicle) platform, repeat steps 4 to 8 to process the feature matrix within the discrete grid system. Update; Step 14.2: Based on the trajectory planned by the ground-based unmanned platform UGV, determine the future... Based on the position and orientation information at each moment, feature values ​​are extracted according to steps 9 and 10. And in combination with the planned driving speed Predict the passage cost at this location using a supervised learning network. By predicting the access cost at various locations in off-road terrain, the accessibility of off-road terrain can be estimated.

Citation Information

Patent Citations

  • 4D millimeter wave radar self-supervision odometer method for autonomous vehicle

    CN118397054A

  • LiDAR dynamic point cloud elimination method and device

    CN118533189A