A method for ground feature recognition of tracked vehicles

Abstract

The present invention relates to a method for identifying ground features of a tracked vehicle, which belongs to the field of vehicle control technology and solves the problem of inaccurate ground feature identification in the prior art. The method includes synchronizing ground coverage features and ground state features in time and space to obtain synchronization features; constructing a vehicle kinematic equation based on a vehicle kinematic model of an instantaneous steering center, and using the synchronization features to solve the equation to estimate the vehicle sliding parameters at each moment; performing principal component analysis on the ground elevation information in the synchronization features to obtain the ground unit normal vector at each moment; estimating the vehicle posture sequence based on the ground unit normal vector at each moment; obtaining the ground texture type based on the deviation between the estimated vehicle posture sequence and the vehicle posture sequence in the synchronization features; and obtaining the ground adhesion coefficient interval based on the vehicle speed, vehicle sliding parameter, surface material, and ground texture type using a clustering algorithm. Accurate identification of ground texture type and ground adhesion coefficient interval is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of feature recognition, and in particular to a ground feature recognition method for a tracked vehicle. Background Art

[0002] The road environment is complex and changeable, the terrain features are complex, and the ground features are strongly correlated with the vehicle's driving status, which will directly affect the vehicle's safety, passability and mission efficiency.

[0003] Currently, direct data measurement methods are commonly used, where additional sensors are installed on the vehicle body and suspension to measure vehicle motion, and a vehicle kinematic model is constructed based on the interaction between the vehicle and the ground to calculate the characteristics of the ground on which the vehicle is traveling.

[0004] This existing method requires the addition of additional sensors to obtain data, and it is not easy to deploy sensors for measurement in some extreme scenarios. Moreover, the sensors installed on the vehicle body and suspension are not accurate enough, and are affected by noise interference, making it more difficult to predict ground features. Summary of the Invention

[0005] In view of the above analysis, an embodiment of the present invention aims to provide a method for identifying ground features of a tracked vehicle, so as to solve the problem of inaccurate ground feature identification in the prior art.

[0006] An embodiment of the present invention provides a method for identifying ground features of a tracked vehicle, comprising the following steps:

[0007] The collected ground cover features and ground state features are synchronized in time and space to obtain synchronized features; the ground cover features include surface material and ground elevation information, and the ground state features include vehicle speed and vehicle posture sequence;

[0008] Constructing a vehicle kinematics equation based on a vehicle kinematics model of an instantaneous steering center, solving the vehicle kinematics equation using the synchronization feature and estimating the vehicle slip parameters at each moment;

[0009] performing principal component analysis on the ground elevation information in the synchronization feature to obtain a ground unit normal vector at each moment; estimating a vehicle pose sequence based on the ground unit normal vector at each moment; and obtaining a ground texture type at each moment based on a deviation between the estimated vehicle pose sequence and the vehicle pose sequence in the synchronization feature;

[0010] According to the vehicle speed, vehicle sliding parameters, surface material and ground texture type at each moment, a clustering algorithm is used to obtain the ground adhesion coefficient interval at each moment; the ground texture type and ground adhesion coefficient interval at each moment are used as the identified ground features.

[0011] Based on a further improvement of the above method, the collected ground cover features and ground state features are synchronized in time and space, including: first using the least squares method to fit each data in the ground state feature to a respective time polynomial, and calculating the ground state features of each moment corresponding to the ground cover feature based on the fitted polynomial to complete the time synchronization of the ground state feature and the ground cover feature; then completing the spatial synchronization of the ground cover feature and the ground state feature by calculating the transformation matrix between different coordinate systems.

[0012] Based on a further improvement of the above method, calculating the transformation matrix between different coordinate systems includes: calculating the transformation matrix between the camera and the IMU coordinate system of the integrated navigation through the following steps:

[0013] Constructing a rotation change matrix between multiple pairs of adjacent and continuous camera images, establishing an overdetermined equation based on the multiplication of the rotation change matrix and the rotation matrix in the transformation matrix to be solved being 0, and calculating the rotation matrix by solving the overdetermined equation;

[0014] Constructing a translation change matrix based on a sliding window, the translation change matrix includes: a translation matrix in the transformation matrix to be solved, multiple ground state features, and observation depths of multiple features on the camera image, and solving the translation matrix using a non-iterative linear method with the goal of minimizing the sum of the Mahalanobis norms of various errors in the translation change matrix within the sliding window;

[0015] According to the rotation matrix and the translation matrix, a transformation matrix between the camera and the IMU coordinate system of the integrated navigation is obtained.

[0016] Based on further improvements of the above method, a vehicle kinematic model based on the instantaneous turning center includes: when the tracked vehicle moves, the vehicle body rotates around the first instantaneous turning center, and the line connecting the vehicle body collective center and the first instantaneous turning center is perpendicular to the actual speed direction of the vehicle; the left and right side track grounding sections rotate around the second instantaneous turning center and the third instantaneous turning center respectively; the slip coefficient is considered based on the longitudinal linear velocity of the left track, and the slip coefficient is considered based on the longitudinal linear velocity of the right track.

[0017] Based on the further improvement of the above method, the vehicle kinematic equation is expressed by the following formula:

[0018]

[0019] Among them, v x Indicates the longitudinal velocity of the tracked vehicle body’s geometric center in the IMU coordinate system, v y represents the lateral velocity of the tracked vehicle body’s geometric center in the IMU coordinate system, ω z represents the vehicle's yaw rate; v sl and vsr Indicates the longitudinal linear speed of the tracks on both sides; x c The x-coordinate value of the first instantaneous turning center; y l and y r They represent the y coordinate values ​​of the second instantaneous turning center and the third instantaneous turning center respectively; f l Indicates the slip coefficient of the left track, f r It represents the slip coefficient of the right track; D represents the center distance of the tracks on both sides.

[0020] Based on the further improvement of the above method, the vehicle sliding parameters are obtained in the following way:

[0021] The slip coefficient and the slip coefficient in the vehicle kinematic equation are solved, and after adding the positive and negative directions according to their consistency with the vehicle's traveling direction, the average value of the slip coefficient and the slip coefficient in the added positive and negative directions is calculated to obtain the value.

[0022] Based on a further improvement of the above method, principal component analysis is performed on the ground elevation information in the synchronization feature to obtain the ground unit normal vector at each moment. The point cloud corresponding to the vehicle trajectory point at each moment is used as the center point cloud, and the area to be analyzed of each center point cloud is obtained according to the ground elevation information. The spatial coordinates of all point clouds in the area to be analyzed of each center point cloud are formed into a matrix and then principal component analysis is performed to obtain the first principal component vector. The first principal component vector is normalized to obtain the corresponding ground unit normal vector.

[0023] Based on a further improvement of the above method, the ground texture type is obtained according to the deviation between the estimated vehicle posture sequence and the vehicle posture sequence in the synchronization feature, including: constructing multiple independent classification models based on the random forest algorithm, and then making the final classification decision by voting on the principle of minority obeys majority; wherein, multiple independent classification models all adopt a support vector machine model, and based on the constructed ground texture truth table, the ground texture type is obtained according to the amplitude spectrum deviation, power spectrum density deviation and cross-correlation vector deviation; the ground texture types include: solid and soft.

[0024] Based on the further improvement of the above method, the amplitude spectrum deviation, power spectrum density deviation and cross-correlation vector deviation are calculated by Fourier transforming the estimated vehicle posture sequence and the vehicle posture sequence in the synchronization feature as a time series, and then calculating the amplitude spectrum, power spectrum density and cross-correlation vector respectively; the amplitude spectrum deviation is calculated using the mean square error, the power spectrum density deviation is calculated using the cosine similarity, and the cross-correlation vector deviation is calculated using the Euclidean distance.

[0025] Based on the further improvement of the above method, according to the vehicle speed, vehicle sliding parameters, surface material and ground texture type at each moment, a clustering algorithm is used to obtain the ground adhesion coefficient interval at each moment, including:

[0026] Based on the ground adhesion coefficient truth table, a record is extracted from each ground adhesion coefficient interval and encoded as the initial cluster center in the clustering algorithm. The vehicle speed, vehicle sliding parameters, surface material and ground texture type at each moment are encoded as the encoding vector of a sample. The clustering algorithm is used to cluster the same number of samples in a preset time period into multiple categories. Based on the ground adhesion coefficient interval corresponding to the initial cluster center of each category, the ground adhesion coefficient interval of each sample in the category is obtained.

[0027] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0028] 1. Collect multimodal data features from different perception areas of tracked vehicles, and improve the accuracy of environmental perception through temporal and spatial synchronization, enhance real-time perception capabilities, and improve the adaptability and accuracy of tracked vehicles in identifying ground features. This assists the tracked vehicle planning module under UAV maneuvers in selecting more efficient and cost-effective paths, enabling the control module to better ensure the safety of tracked vehicles in different driving scenarios.

[0029] 2. Using environmental perception to obtain ground elevation information, the vehicle pose sequence is estimated and compared with the actual vehicle pose sequence. This increases the number of information modes and improves the accuracy of identifying the firmness / softness of the ground.

[0030] 3. A vehicle kinematic model based on the instantaneous steering center is constructed, which is closer to actual operating conditions and improves the calculation accuracy of sliding parameters. At the same time, a ground adhesion coefficient truth table is constructed considering different vehicle speeds, vehicle sliding parameters, surface materials, and ground texture types. This improves the rationality of the initial cluster center selection of the clustering algorithm, makes the clustering algorithm more generalizable, and improves the prediction accuracy of the ground adhesion coefficient range in actual scenarios.

[0031] In the present invention, the above-mentioned technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of the present invention will be described in the following description, and some advantages will become apparent from the description or be learned through practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0033] Figure 1 This is a flow chart of a method for identifying ground features of a tracked vehicle in an embodiment of the present invention;

[0034] Figure 2 Schematic diagram of a vehicle kinematic model based on the instantaneous turning center in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, and are not used to limit the scope of the present invention.

[0036] A specific embodiment of the present invention discloses a method for identifying ground features of a tracked vehicle, such as Figure 1 As shown, the following steps are included:

[0037] S1. Synchronize the collected ground cover features and ground state features in time and space to obtain synchronized features; the ground cover features include surface material and ground elevation information, and the ground state features include vehicle speed and vehicle posture sequence.

[0038] It should be noted that the tracked vehicle in this embodiment is a ground-based unmanned mobile platform equipped with a camera, an environmental perception lidar, and a body-perception integrated navigation system. It perceives different areas at the same time. The multi-environmental perception area of ​​the camera and environmental perception lidar covers a concentric circle range of 0.6m to 15m with the center of the vehicle as the center, and is used to obtain ground coverage features, including: surface materials identified from images captured by the camera and ground elevation information obtained from point cloud data collected by the lidar. Surface materials are categorized as paved roads, dirt roads, gravel roads, icy roads, and grass. The ground elevation information is a 0.2m resolution elevation grid map. The body-perception integrated navigation system's body-perception perception area covers a range of 0.6m with the center of the vehicle as the center, and is used to obtain ground state features from the data collected by the integrated navigation system, including: vehicle speed, drive wheel speed, and vehicle posture sequence, where the vehicle posture includes the vehicle's pitch angle, roll angle, and roll angle.

[0039] Since the ground cover characteristics and ground state characteristics are the perception results of different areas and the sampling frequencies are also different, the results of these two areas are synchronized in time and space, and the aligned results are used to estimate the terrain characteristics of the same location.

[0040] Specifically, when performing time synchronization, each data in the ground state feature is fitted using the least squares method to obtain a time t b The body perception information curve S(t bConsidering that the change of ground state characteristics is not linear, a polynomial is used for fitting. At the same time, considering the influence of noise, the degree of the polynomial cannot be too high. Therefore, a cubic polynomial is used for fitting, which not only takes into account the nonlinear characteristics but also eliminates the influence of measurement noise on the least squares method. The formula is as follows:

[0041] S(t b )=y b =a0+a1t b +a2t b 2 +a3t b 3 (1)

[0042] Using the least square method, with the goal of minimizing the sum of squares of errors, we fit the coefficients a0, a1, a2, and a3, and construct the curve S(t b ).

[0043] Furthermore, according to the acquisition time t of the environment perception lidar c , find the ground state characteristics y corresponding to the time on each fitted curve c =S(t c ), which corresponded to the ground coverage characteristics at the same time and achieved time synchronization.

[0044] It should be noted that spatial synchronization is a calibration method. Tracked vehicles usually involve multiple coordinate systems. By calculating the transformation matrix between different coordinate systems, the spatial synchronization of ground coverage features and ground state features is completed.

[0045] Among them, the transformation matrix between different coordinate systems includes but is not limited to: using a camera with environmental sensing capabilities and a combined navigation system with body perception capabilities as calibration objects for two environments to perform external parameter calibration. That is, at the same time, based on the pose estimated by the camera image and the actual pose obtained by the combined navigation system with body perception, the transformation matrix T between the camera and IMU coordinate systems is obtained. bc (including the rotation matrix q bc and the translation matrix p bc ); The ground coverage features at the same time are converted into data in the IMU coordinate system according to the transformation matrix.

[0046] Specifically, the optical flow method is used to extract the adjacent frames c of the camera image i ,c i+1 The relative motion between At the same time, two adjacent frames b under IMU i ,b i+1 The pose change is expressed as Construct the overdetermined matrix equation shown below:

[0047]

[0048] Since the equation contains quaternions and spatial transformations, it is not full rank and cannot be solved directly. The following approximate solutions are given for the rotation and translation matrices respectively.

[0049] First, by decomposing the posture changes, the rotation changes of the camera and IMU in two adjacent frames are obtained as follows: and Construct the change matrix Q of adjacent frames as shown below i,i+1 :

[0050]

[0051] Based on the single-frame changes, the rotation change matrix between multiple pairs of adjacent and continuous camera images is constructed. The overdetermined equation is established by multiplying the rotation change matrix with the rotation matrix in the transformation matrix to be solved to 0, avoiding the problem of being unable to solve due to insufficient rank:

[0052]

[0053] Among them, N represents the number of frames, w i,i+1 Represents the change weight between two adjacent images, Q N Represents the selection change matrix for multiple pairs of adjacent frames.

[0054] It should be noted that if the rotation matrix q is estimated based on the initial images bc Initial value of Then first calculate the threshold r between two adjacent images according to the following formula: i,i+1 :

[0055]

[0056] Where a represents the coefficient and tr(·) represents the trace of the matrix.

[0057] Then calculate the corresponding change weight using the following formula to calculate the rotation matrix q bc :

[0058]

[0059] Preferably, the preset threshold value is set to 0.02.

[0060] If the initial value of the rotation matrix cannot be estimated Then Q N After filling the full rank, by calculating Q N The rotation matrix q is calculated by the right unit singular vector of the minimum singular value of bc .

[0061] After completing the calculation of the rotation matrix, a tightly coupled sliding window is used to calculate the position changes of multiple ground state features and image estimates in the IMU coordinate system to initialize the translation matrix.

[0062] Specifically, a translation change matrix χ is constructed based on the sliding window, which includes: the translation matrix in the transformation matrix to be solved, multiple ground state features and the observation depth of multiple features on the camera image, which is expressed by the following formula:

[0063]

[0064] Among them, x j represents the jth ground state feature, Represents the gravity vector of the jth ground state feature, N n represents the number of ground state features in the sliding window, M m represents the number of features with sufficient parallax (significant displacement) in the sliding window; n and m are the starting indices in the sliding window; λ m is the depth of the mth feature on the image from the first observation; It is the initial translation matrix used to calculate the spatial motion result of the translation calculation; Represents the vehicle's pitch angle, roll angle, and ground state characteristics.

[0065] After initialization, the translation matrix p is completed by performing maximum likelihood estimation on the matrix obtained by the sliding window bc Estimation, where the maximum likelihood estimation is to minimize the sum of the Mahalanobis norms of all measurement errors from the IMU and camera within the sliding window, as shown below:

[0066]

[0067] Wherein, subscript B represents feature b k The set of all IMU measurements, subscript C represents feature c l and camera pose c h The set of all observations between p Represents the changing model of camera and IMU positioning under time alignment; Indicates the change in the camera's estimated posture. Represents the linear camera measurement model, which is a 3rd-order square matrix containing the camera intrinsic parameters; Represents the linear IMU measurement model, which is a 4-order square matrix including the rotation matrix and translation matrix; Indicates the change in IMU measurement, Indicates the translation change between two adjacent frames, Indicates the rotation change between two adjacent frames.

[0068] The translation matrix is ​​solved using a non-iterative linear method, and the minimum Mahalanobis distance of the minimum sliding window is used as the optimization target. The translation matrix p is solved using the Newton method. bc Optimize.

[0069] According to the above obtained rotation matrix q from the camera to the IMU bc And the translation matrix p bc , forming the transformation matrix T bc , achieving a spatial alignment.

[0070] In this embodiment, the ground coverage feature reflects the visual characteristics of the environment's appearance, and the ground state feature reflects the "invisible" "tactile" characteristics of the surface. By aligning the two features, the accuracy of environmental perception is improved, the real-time perception capability is enhanced, and the adaptability and accuracy of tracked vehicle ground feature recognition are improved.

[0071] S2. Construct a vehicle kinematic model based on the instantaneous steering center and use the synchronization feature to estimate the vehicle sliding parameters at each moment.

[0072] It's important to note that the interaction between the vehicle and the ground can be captured by the sliding parameter, which represents the interaction between the tracks and the ground. Tracked vehicles do not travel in a straight line, but rather are abstracted into a combination of steering primitives with multiple curvatures. Because the track contact area of ​​a tracked vehicle is large, slippage and rotation occur during steering. Therefore, the track contact area of ​​a tracked vehicle cannot be simplified to a single point. Therefore, the trajectory of a tracked vehicle depends on numerous factors and cannot be directly predicted simply by the angular velocity of the tracks on both sides.

[0073] The actual sliding parameters are derived from the instantaneous turning situation of the sliding. This embodiment abstracts the instantaneous turning situation into Figure 2 The vehicle kinematic model based on the instantaneous turning center shown in the figure shows that when the vehicle is in motion, the vehicle body can be considered to rotate about the first instantaneous turning center, while the two tracks not only follow the vehicle body at the drag velocity but also move relative to the vehicle body at a longitudinal linear velocity. Therefore, the absolute velocity of a point on the track is the vector sum of the drag velocity and the longitudinal linear velocity at that point. The two track ground sections are considered as rigid bodies, and the left and right track ground sections rotate about the second and third instantaneous turning centers, respectively.

[0074] exist Figure 2 In the figure, the tracked vehicle moves on the horizontal ground and the center of mass of the tracked vehicle coincides with the geometric center. The height information is ignored, OXY is used to represent the earth coordinate system, and oxy is used to represent the vehicle body IMU coordinate system. xis the longitudinal velocity of the tracked vehicle body’s geometric center o in the IMU coordinate system, v y is the lateral velocity of the geometric center of the tracked vehicle body in the IMU coordinate system, and the absolute velocity v of the geometric center of the vehicle body o It is v x and v y Vector sum. ω z is the vehicle yaw rate, counterclockwise is positive, ω v is the driving wheel speed of the tracked vehicle, which is collected by the vehicle wheel speed sensor; r v is the radius of the tracked vehicle's driving wheel. θ represents the vehicle's heading angle, O c (x c ,y c ) is the first instantaneous turning center of the vehicle; l (x l ,y l ) and O r (x r ,y r ) represent the second and third instantaneous turning centers corresponding to the left and right tracks, respectively; v sl and v sr Indicates the longitudinal linear speed of the tracks on both sides; O l and O r The line between them constructs points M and N on the left and right tracks, and the v of points M and N on the tracks qM 、v qN Indicates the drag speed of the left and right tracks of the tracked vehicle; D is the center distance of the tracks on both sides, where O c The line connecting the geometric center o of the vehicle body is perpendicular to the actual velocity direction v of the vehicle o .

[0075] Furthermore, using the coordinates of the tracked vehicle's center of mass o = [X, Y, θ] T To express the position of the tracked vehicle in the earth coordinate system, the formula is as follows:

[0076]

[0077] The vehicle speed in the traditional vehicle coordinate system is used in the kinematic model, and the slip coefficient f of the left track is taken into account based on the longitudinal linear speed of the left and right tracks. l and the slip coefficient f of the right track r , the vehicle longitudinal velocity, lateral velocity and yaw rate are obtained according to the following formula:

[0078]

[0079] Furthermore, by analyzing the track contact section of the tracked vehicle, the coordinate values ​​of each instantaneous turning center of the tracked vehicle are calculated using the following formula, which describes the relationship between the instantaneous turning center of the tracks and the vehicle body on both sides and other kinematic parameters when the vehicle turns:

[0080]

[0081] According to the above formula, the vehicle longitudinal velocity, lateral velocity and steering angular velocity are as follows:

[0082]

[0083] Among them, the longitudinal linear speed of the crawlers on both sides is calculated by dividing the speed of the crawler driving wheels on both sides by ω sl 、ω sr and the radius r of the track drive wheel r Multiplying them together, the formula is as follows:

[0084]

[0085] According to formula (10) and formula (12), the vehicle kinematic equation is as follows:

[0086]

[0087] Get the vehicle longitudinal velocity v from the synchronization feature x and the vehicle lateral velocity v y , the longitudinal linear velocity v of the left and right tracks is obtained by the speed of the track drive wheel sl 、v sr and the actual instantaneous steering center coordinates, and finally calculate the current vehicle slip coefficient f l and slip coefficient f r .

[0088] It should be noted that the slip coefficient and the slip coefficient have positive and negative directions. A slip coefficient direction that aligns with the vehicle's travel direction is recorded as positive, and vice versa. A slip coefficient direction that aligns with the vehicle's travel direction is recorded as negative, and vice versa is recorded as positive. The slip coefficient direction and the slip coefficient direction are each calculated by calculating the difference between the drag velocity and the longitudinal linear velocity of the corresponding side track.

[0089] Give slip coefficient f l and slip coefficient f l After adding the positive and negative directions, the vehicle's sliding parameter f is obtained according to the following formula:

[0090]

[0091] S3. Perform principal component analysis on the ground elevation information in the synchronization features to obtain the ground unit normal vector at each moment; estimate the vehicle posture sequence based on the ground unit normal vector at each moment; and obtain the ground texture type at each moment based on the deviation between the estimated vehicle posture sequence and the vehicle posture sequence in the synchronization features.

[0092] It should be noted that the ground texture types include solid and soft. This feature is part of the ground features to be identified in this embodiment. The principal component analysis method is used on the ground elevation information synchronized in step S1 to calculate the normal vector of the discrete surface and estimate the possible posture information of the vehicle; the deviation between the estimated posture information and the posture information synchronized in step S1 is compared to determine the solid or soft characteristics of the ground.

[0093] The principal component analysis method is calculated on the elevation information to obtain the ground unit normal vector at each moment, including: taking the point cloud corresponding to the vehicle trajectory point at each moment as the central point cloud, obtaining the area to be analyzed composed of each central point cloud and its surrounding neighboring point clouds based on the ground elevation information, projecting the area to be analyzed, fitting the elevation information of the area to be analyzed to find a direction vector that makes the distribution of the projection points of the point cloud in this area on the direction vector the most concentrated. This direction vector is the ground unit normal vector to be calculated.

[0094] Specifically, in the ground elevation information, a grid area with a distance of 2 units from the central point cloud in 8 directions (up, down, left, right, and four diagonals) is taken as the area to be analyzed, that is, the area to be analyzed is a grid area of ​​25 point clouds including the central point cloud; the three-dimensional spatial coordinates of each point cloud in the area to be analyzed are organized into a 25×3 matrix and then principal component analysis is performed to obtain the first principal component vector, that is, the maximum eigenvalue vector.

[0095] The first principal component vector λ max The components λ in the x, y, and z directions x ,λ y ,λ z After normalization, the corresponding ground unit normal vector γ is obtained, and the formula is as follows:

[0096]

[0097] According to the components of the ground unit normal vector in each coordinate axis of the geodetic coordinate system, the angle between the ground unit normal vector and the z-axis is used as the z-axis rotation angle of the IMU coordinate system relative to the geodetic coordinate system, and the rotation matrix of the geodetic coordinate system relative to the IMU coordinate system is calculated; according to the rotation matrix and the coordinates of the vehicle trajectory points at each moment, the vehicle posture sequence is estimated. in represents the pitch angle in the geodetic coordinate system at time t, is the roll angle in the geodetic coordinate system at time t, is the roll angle in the geodetic coordinate system at time t.

[0098] Furthermore, the ground texture type is obtained based on the deviation between the estimated vehicle pose sequence and the vehicle pose sequence in the synchronization features, including: constructing multiple independent classification models based on the random forest algorithm, and then making the final classification decision by voting on the principle of minority obeys majority; among them, multiple independent classification models all adopt the support vector machine model, and obtain classification results based on the constructed ground texture truth table according to the amplitude spectrum deviation, power spectrum density deviation and cross-correlation vector deviation.

[0099] Specifically, the estimated vehicle posture sequence and the vehicle posture sequence synchronized in step S1 are Fourier transformed as time series, and the amplitude spectrum, power spectrum density and cross-correlation vector are calculated respectively; the amplitude spectrum deviation is calculated using the mean square error, the power spectrum density deviation is calculated using the cosine similarity, and the cross-correlation vector deviation is calculated using the Euclidean distance.

[0100] During implementation, the system drives on roads with various surface materials. By calculating the height difference between the ruts left by the tracked vehicle and the actual ground, the solid or soft classification is directly obtained as the most intuitive label. Then, by analyzing the estimated vehicle pose sequences and the collected vehicle pose sequences on different surface materials, the thresholds of the amplitude spectrum deviation, power spectrum density deviation and cross-correlation vector deviation are determined to construct a ground texture truth table.

[0101] As shown in Table 1, the table exemplarily shows a rutting depth, thresholds of various deviations, and corresponding classifications for a tracked vehicle on a ground with different surface materials.

[0102] Table 1 Truth value table of ground firmness / softness for tracked vehicles

[0103]

[0104] S4. Based on the vehicle speed, vehicle sliding parameter, surface material and ground texture type at each moment, a clustering algorithm is used to obtain the ground adhesion coefficient interval at each moment; the ground texture type and ground adhesion coefficient interval at each moment are used as the identified ground features.

[0105] It should be noted that the ground adhesion coefficient, as an important component of ground characteristics, provides more accurate constraints for vehicle motion planning and control. The ground adhesion coefficient is in the range (0, 1), and in this embodiment, it is divided into 10 intervals: (0-0.10, 0.10-0.20, ..., 0.90-1.00).

[0106] Considering that clustering algorithms, especially K-means algorithm, seek local optimal solutions through iterative optimization, the initial cluster centers are very sensitive. The selection of the initial cluster centers has a great influence on the final clustering results. If the initial cluster centers are not selected properly, it may lead to low clustering quality or slow algorithm convergence.

[0107] In this embodiment, the ground adhesion coefficient truth table of the tracked vehicle is obtained by repeated testing on roads of various surface materials, and the initial cluster center is selected from the truth table.

[0108] Specifically, a suspension displacement sensor, a vehicle engine power sensor, and a body acceleration sensor are installed on a tracked vehicle. Under a test environment consisting of various surface materials, vehicle speeds, ground textures, and sliding parameters, the vehicle engine power sensor is used to determine the power and speed of the vehicle engine, and the traction force f provided by the vehicle engine in real time is obtained. c ; Then through the known vehicle weight m c and vehicle acceleration a c , the ground adhesion coefficient under each test environment is obtained by the following formula:

[0109]

[0110] After repeated testing in each test environment, the true ground adhesion coefficient value for that test environment was obtained based on the range of the average ground adhesion coefficient. Ultimately, the ground adhesion coefficient true values ​​for multiple test environments formed a ground adhesion coefficient truth table for tracked vehicles, as shown in Table 2.

[0111] Table 2 Truth table of ground adhesion coefficient of tracked vehicles

[0112]

[0113] Because the test environment cannot cover all the complex environments in reality, the truth table is only used to select the initial cluster centers. Specifically, a record is randomly selected from the tracked vehicle ground adhesion coefficient truth table for each of the 10 ground adhesion coefficient intervals. The corresponding vehicle speed interval, vehicle slip parameter interval, ground surface material, and ground texture type are encoded separately. The resulting encoding vectors are then concatenated and used as the encoding vectors for the 10 initial cluster centers.

[0114] It should be noted that the surface material and ground texture type are text features, and the word2vec model is used to obtain their respective text feature codes; the vehicle speed and vehicle sliding parameter are intervals, and their corresponding interval feature codes are obtained; the text feature codes and interval feature codes are concatenated to obtain the encoding vectors of each record in the truth table.

[0115] Furthermore, based on the vehicle speed, vehicle sliding parameters, surface material, and ground texture type at each moment, a clustering algorithm is used to obtain the ground adhesion coefficient interval at each moment, including:

[0116] The vehicle speed, vehicle sliding parameter, surface material and ground texture type at each moment are regarded as the four-dimensional features of a sample. The encoding vector of each sample is obtained in the same way as the encoding vector of the initial cluster center. The k-means algorithm is used to cluster the same number of samples in a preset time period into multiple categories. The ground adhesion coefficient interval corresponding to the initial cluster center of each category is obtained according to the ground adhesion coefficient interval of each sample in the category.

[0117] Preferably, in the k-means algorithm, the distance between samples is measured by calculating the Minkowski distance, and the Minkowski distance parameter p is set to 0.5.

[0118] Ultimately, the ground texture type and ground adhesion coefficient range at each moment are used as the ground feature identification results to assist the tracked vehicle planning module under unmanned aerial vehicle maneuvers in selecting more efficient and cost-effective paths, enabling the control module to better ensure the safety of tracked vehicles in different driving scenarios.

[0119] Compared with the prior art, the method for identifying ground features of tracked vehicles provided in this embodiment collects multimodal data features of different perception areas of tracked vehicles, and improves the accuracy of environmental perception through synchronization of time and space, enhances real-time perception capabilities, and improves the adaptability and accuracy of ground feature recognition of tracked vehicles. The ground elevation information obtained by environmental perception is used to estimate the vehicle posture sequence, which is compared with the actual vehicle posture sequence, increasing the number of modes of information and improving the accuracy of identifying the firmness / softness of the ground. A vehicle kinematic model based on the instantaneous steering center is constructed, which is closer to the actual operating conditions and improves the calculation accuracy of the sliding parameters; at the same time, a ground adhesion coefficient truth table is constructed considering different vehicle speeds, vehicle sliding parameters, surface materials and ground texture types, which improves the rationality of the initial cluster center selection of the clustering algorithm, makes the clustering algorithm have better generalization capabilities, and improves the prediction accuracy of the ground adhesion coefficient range in actual scenarios.

[0120] Those skilled in the art will appreciate that all or part of the process steps of the above-described embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, such as a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0121] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention.

Claims

1. A method for identifying ground features of a tracked vehicle, characterized in that: The following steps are involved: The collected ground cover features and ground state features are synchronized in time and space to obtain synchronized features; the ground cover features include surface material and ground elevation information, and the ground state features include vehicle speed and vehicle posture sequence; A vehicle kinematic model based on an instantaneous turning center is used to construct a vehicle kinematic equation, and the synchronization feature is used to solve the vehicle kinematic equation and estimate the vehicle slip parameter at each moment. The vehicle kinematic model based on the instantaneous turning center includes: when the tracked vehicle moves, the vehicle body rotates around a first instantaneous turning center, and the line connecting the vehicle body geometric center and the first instantaneous turning center is perpendicular to the actual speed direction of the vehicle; the left and right track ground contact sections rotate around a second instantaneous turning center and a third instantaneous turning center, respectively; the slip coefficient is considered based on the longitudinal linear velocity of the left track, and the slip coefficient is considered based on the longitudinal linear velocity of the right track; the vehicle slip parameter is obtained by solving the slip coefficient and the slip coefficient in the vehicle kinematic equation, and after adding the positive and negative directions according to their consistency with the vehicle's travel direction, calculating the average of the added slip coefficient and the slip coefficient in the positive and negative directions; performing principal component analysis on the ground elevation information in the synchronization feature to obtain a ground unit normal vector at each moment; estimating a vehicle pose sequence based on the ground unit normal vector at each moment; and obtaining a ground texture type at each moment based on a deviation between the estimated vehicle pose sequence and the vehicle pose sequence in the synchronization feature; According to the vehicle speed, vehicle sliding parameters, surface material and ground texture type at each moment, a clustering algorithm is used to obtain the ground adhesion coefficient interval at each moment; the ground texture type and ground adhesion coefficient interval at each moment are used as the identified ground features.

2. The method for recognizing ground features of a tracked vehicle according to claim 1, wherein: The collected ground cover features and ground state features are synchronized in time and space, including: fitting each data in the ground state feature with a respective time polynomial using the least squares method, calculating each ground state feature at the corresponding moment of the ground cover feature based on the fitted polynomial, and completing the time synchronization of the ground state feature and the ground cover feature; and then completing the spatial synchronization of the ground cover feature and the ground state feature by calculating the transformation matrix between different coordinate systems.

3. The method for recognizing ground features of a tracked vehicle according to claim 2, wherein: Calculating the transformation matrix between different coordinate systems includes: calculating the transformation matrix between the camera and the IMU coordinate system of the integrated navigation through the following steps: Constructing a rotation change matrix between multiple pairs of adjacent and continuous camera images, establishing an overdetermined equation based on the multiplication of the rotation change matrix and the rotation matrix in the transformation matrix to be solved being 0, and calculating the rotation matrix by solving the overdetermined equation; Constructing a translation change matrix based on a sliding window, the translation change matrix includes: a translation matrix in the transformation matrix to be solved, multiple ground state features, and observation depths of multiple features on the camera image, and solving the translation matrix using a non-iterative linear method with the goal of minimizing the sum of the Mahalanobis norms of various errors in the translation change matrix within the sliding window; According to the rotation matrix and the translation matrix, a transformation matrix between the camera and the IMU coordinate system of the integrated navigation is obtained.

4. The method for recognizing ground features of a tracked vehicle according to claim 1, wherein: The vehicle kinematic equation is expressed by the following formula: Among them, v x Indicates the longitudinal velocity of the tracked vehicle body’s geometric center in the IMU coordinate system, v y represents the lateral velocity of the tracked vehicle body’s geometric center in the IMU coordinate system, ω z represents the vehicle's yaw rate; v sl and v sr Indicates the longitudinal linear speed of the tracks on both sides; x c The x-coordinate value of the first instantaneous turning center; y l and y r They represent the y coordinate values ​​of the second instantaneous turning center and the third instantaneous turning center respectively; f l Indicates the slip coefficient of the left track, f r It represents the slip coefficient of the right track; D represents the center distance of the tracks on both sides.

5. The method for recognizing ground features of a tracked vehicle according to claim 1, wherein: Principal component analysis is performed on the ground elevation information in the synchronization feature to obtain the ground unit normal vector at each moment. The point cloud corresponding to the vehicle trajectory point at each moment is used as the center point cloud, and the area to be analyzed of each center point cloud is obtained according to the ground elevation information. The spatial coordinates of all point clouds in the area to be analyzed of each center point cloud are formed into a matrix and then principal component analysis is performed to obtain the first principal component vector. The first principal component vector is normalized to obtain the corresponding ground unit normal vector.

6. The method for recognizing ground features of a tracked vehicle according to claim 1, wherein: The method of obtaining the ground texture type based on the deviation between the estimated vehicle posture sequence and the vehicle posture sequence in the synchronization feature includes: constructing multiple independent classification models based on the random forest algorithm, and then making a final classification decision by voting on the principle of minority obeys majority; wherein, the multiple independent classification models all adopt a support vector machine model, and obtain the ground texture type based on the amplitude spectrum deviation, power spectrum density deviation and cross-correlation vector deviation based on the constructed ground texture truth table; the ground texture types include: solid and soft.

7. The method for recognizing ground features of a tracked vehicle according to claim 6, wherein: The amplitude spectrum deviation, power spectrum density deviation and cross-correlation vector deviation are calculated by Fourier transforming the estimated vehicle posture sequence and the vehicle posture sequence in the synchronization feature as a time series, and then calculating the amplitude spectrum, power spectrum density and cross-correlation vector respectively; the amplitude spectrum deviation is calculated using the mean square error, the power spectrum density deviation is calculated using the cosine similarity, and the cross-correlation vector deviation is calculated using the Euclidean distance.

8. The method for recognizing ground features of a tracked vehicle according to claim 1, wherein: The clustering algorithm is used to obtain the ground adhesion coefficient interval at each moment based on the vehicle speed, vehicle sliding parameters, surface material, and ground texture type at each moment, including: Based on the ground adhesion coefficient truth table, a record is extracted from each ground adhesion coefficient interval and encoded as the initial cluster center in the clustering algorithm. The vehicle speed, vehicle sliding parameters, surface material and ground texture type at each moment are encoded as the encoding vector of a sample. The clustering algorithm is used to cluster the same number of samples in a preset time period into multiple categories. Based on the ground adhesion coefficient interval corresponding to the initial cluster center of each category, the ground adhesion coefficient interval of each sample in the category is obtained.

Images