A laser-based building construction ground leveling robot
By fusing data from the laser receiver array and inertial measurement unit and reconstructing it using Kalman filtering, the problem of elevation drift caused by laser signal distortion at the construction site was solved, enabling high-precision leveling operations in complex environments.
Patent Information
- Application Number
- CN202611145077.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-30
- Publication Date
- 2026-11-10
AI Technical Summary
Existing construction site leveling robots are prone to laser signal distortion or loss due to dust obstruction, multi-media reflection, and mechanical vibration interference, resulting in elevation benchmark drift and inaccurate elevation during leveling operations.
The system extracts a two-dimensional light intensity distribution feature vector using a laser receiving array and fuses it with data from an inertial measurement unit. The interference type is determined by feature fusion and distortion identification components. Kalman filtering is used to reconstruct the elevation benchmark, generate adaptive leveling control commands, and drive the leveling blade to move.
Under the interference of dust obstruction and multi-media reflection, it maintains the ability to lock the elevation, avoids the leveling elevation from being inaccurate, reduces response lag, and improves the stability of state estimation and the leveling effect.
Smart Images

Figure CN123064281A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of building construction ground leveling robots, specifically relating to a building construction ground leveling robot based on laser positioning. Background Technology
[0002] Existing construction site leveling robots typically use laser positioning to determine elevation. The working process is as follows: a rotating laser emitter is set up at the construction site to establish a horizontal reference plane. A single laser receiver is installed on the robot body to capture the laser signal. The received laser signal position is converted into an elevation deviation electrical signal. The control system drives the leveling blade to rise and fall according to the difference between the elevation deviation electrical signal and the preset elevation. At the same time, the robot moves along the preset trajectory to complete the leveling operation.
[0003] Construction sites are subject to dust obstruction, multi-media reflection, and mechanical vibration interference. The laser signal acquired by a single laser receiver is prone to distortion or loss, causing the elevation benchmark to drift. As a result, the leveling system cannot obtain the true elevation benchmark, leading to inaccurate elevation during leveling operations. Summary of the Invention
[0004] The purpose of this invention is to provide a laser-based building construction ground leveling robot, which can solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A laser-based ground leveling robot for building construction includes: a laser receiving array for capturing a rotating laser beam and extracting a two-dimensional light intensity distribution feature vector of the light spot; an inertial measurement unit for outputting the robot's pose data; a feature fusion and distortion recognition unit for fusing the two-dimensional light intensity distribution feature vector with the pose data in a feature space and determining the distortion type of the current laser signal through a pre-built interference recognition model, wherein the distortion type includes occlusion, scattering, or diffuse reflection; an elevation benchmark reconstruction unit for reconstructing the current elevation benchmark using Kalman filtering based on the distortion type and historical valid laser benchmark data; a leveling control command generation unit for calculating the cutting depth of the leveling blade based on the current elevation benchmark and a local ground elevation prediction model obtained by laser scanning, and generating an adaptive leveling control command including feedforward compensation; and a leveling actuator for driving the leveling blade to move according to the adaptive leveling control command.
[0006] Preferably, the laser receiving array is arranged at non-uniform intervals along the circumference of the leveling robot body to form a multi-layer ring topology. The two-dimensional light intensity distribution feature vector is extracted in the following way: each photoelectric detection node in the laser receiving array synchronously collects the light intensity amplitude of the rotating laser beam. Based on the spatial coordinate position of each photoelectric detection node in the multi-layer ring topology, the corresponding light intensity amplitude is spatially interpolated and mapped to generate a two-dimensional light intensity distribution matrix of the light spot. Principal component analysis is performed on the two-dimensional light intensity distribution matrix of the light spot to extract the principal component feature values. The principal component feature values are concatenated with the signal-to-noise ratio features of each photoelectric detection node to form the two-dimensional light intensity distribution feature vector.
[0007] Preferably, the feature fusion and distortion recognition component fuses the two-dimensional light intensity distribution feature vector with the pose data in the feature space as follows: the roll angle, pitch angle, and yaw angle parameters in the pose data are converted into quaternion attitude representation vectors; the quaternion attitude representation vectors and the two-dimensional light intensity distribution feature vectors are subjected to tensor product operation to construct a multimodal feature tensor; and the multimodal feature tensor is nonlinearly dimensionality reduced through manifold learning to obtain a low-dimensional fused feature representation. The pre-built interference recognition model is a support vector data description classifier built based on the low-dimensional fused feature representation. The support vector data description classifier determines the distortion type by calculating the Mahalanobis distance from the low-dimensional fused feature representation to the center of the hypersphere.
[0008] Preferably, the elevation benchmark reconstruction component reconstructs the current elevation benchmark using Kalman filtering based on the distortion type and historical valid laser reference data as follows: It queries a preset interference noise intensity mapping table according to the distortion type to obtain a dynamic observation noise covariance matrix matching the current interference environment; it injects the dynamic observation noise covariance matrix into the update equation of the Kalman filter; it constructs a state transition prediction equation using the historical valid laser reference data; the Kalman filter iteratively outputs an elevation state estimate and a covariance estimate based on the state transition prediction equation and the update equation containing the dynamic observation noise covariance matrix; and it uses the elevation state estimate as the current elevation benchmark.
[0009] Preferably, the method by which the leveling control command generation component calculates the cutting depth of the leveling blade and generates an adaptive leveling control command including feedforward compensation based on the current elevation benchmark and the ground local elevation prediction model obtained by laser scanning is as follows: A three-dimensional point cloud of the ground at a preset distance in front of the leveling blade is obtained through the laser scanning component; the three-dimensional point cloud is fitted with a local surface to generate the ground local elevation prediction model; the elevation change gradient along the leveling advance direction in the ground local elevation prediction model is extracted; the elevation change gradient is differentially calculated with the current elevation benchmark to obtain the cutting depth; the feedforward compensation is generated based on the derivative of the elevation change gradient; and the cutting depth and the feedforward compensation are superimposed to form the adaptive leveling control command.
[0010] Preferably, the leveling actuator drives the leveling blade according to the adaptive leveling control command in the following manner: the leveling actuator includes a left hydraulic cylinder and a right hydraulic cylinder that independently drive the left and right ends of the leveling blade to rise and fall; the cutting depth and feedforward compensation amount in the adaptive leveling control command are analyzed; the independent target displacements of the left and right ends of the leveling blade are calculated; the independent target displacements are input to the asymmetric proportional valve control model; the asymmetric proportional valve control model adjusts the flow rate and pressure of the rodless chamber of the left and right hydraulic cylinders respectively according to the independent target displacements, thereby realizing differential synchronous drive of the leveling blade's posture.
[0011] Preferably, during the process of extracting the two-dimensional light intensity distribution feature vector of the light spot by the laser receiving array, when the support vector data description classifier determines that the distortion type is scattering, the coordinates of the edge pixels whose light intensity amplitude crosses the preset attenuation threshold in the two-dimensional light intensity distribution matrix of the light spot are extracted, the edge pixel coordinates are fitted to obtain the edge contour of the light spot, the normal gradient change rate of each point on the edge contour of the light spot is calculated, and the spatial distribution density of the corresponding photodetector nodes in the multi-layer ring topology is combined to construct the refractive index inversion model of the scattering medium. The center offset compensation vector of the light spot is output according to the refractive index inversion model of the scattering medium, and the center offset compensation vector of the light spot is superimposed on the two-dimensional light intensity distribution feature vector.
[0012] Preferably, before constructing a multimodal feature tensor by performing a tensor product operation between the quaternion pose representation vector and the two-dimensional light intensity distribution feature vector, the angular velocity component in the pose data is extracted, the angular velocity component is input into the hyperbolic tangent activation function to calculate the attention weight factor, the attention weight factor is multiplied element-wise with the feature dimension corresponding to the high-frequency light intensity change in the two-dimensional light intensity distribution feature vector to obtain an adaptive weighted light intensity feature vector, and the adaptive weighted light intensity feature vector is used to replace the two-dimensional light intensity distribution feature vector and perform a tensor product operation with the quaternion pose representation vector.
[0013] Preferably, while injecting the dynamic observation noise covariance matrix into the Kalman filter update equation, the first-order difference of the elevation state estimate values at adjacent times in the historical effective laser reference data is calculated to obtain the elevation change rate, and an elevation change detection function is constructed. When the elevation change rate is greater than a preset change threshold, the Kalman filter gain adjustment mechanism is triggered, and the Kalman gain coefficient is forcibly set to zero to block the update of the state estimate by the current observation value. When the elevation change rate is less than or equal to the preset change threshold, the Kalman gain coefficient calculated based on the dynamic observation noise covariance matrix is restored.
[0014] Preferably, in the process of generating a local ground elevation prediction model by performing local surface fitting on the three-dimensional point cloud, the local normal vector of each data point in the three-dimensional point cloud is calculated, and the angular deviation between the local normal vectors of each data point in the neighborhood is statistically analyzed. When the angular deviation exceeds a preset consistency threshold, the corresponding data point is marked as an isolated feature point. The local area containing the isolated feature point is subdivided into grids, and the subdivided subgrids are assigned weight coefficients that are negatively correlated with the angular deviation. Based on the weighted subgrids, the least squares local surface fitting is re-executed to update the local ground elevation prediction model.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By extracting the two-dimensional light intensity distribution feature vector of the light spot and fusing it with the pose data in the feature space, the pre-constructed interference recognition model is used to determine the distortion type such as occlusion, scattering or diffuse reflection. Combined with historical effective laser reference data and Kalman filtering, the current elevation reference is reconstructed. This eliminates the reference distortion defect of a single laser signal source under dust occlusion and multi-media reflection interference, maintains the elevation locking capability of the leveling system under interference conditions, and avoids the leveling elevation inaccuracy caused by reference drift.
[0016] 2. Based on the local ground elevation prediction model, the elevation change gradient along the direction of travel is extracted to calculate the cutting depth and generate an adaptive leveling control command containing feedforward compensation, which reduces the response lag of the leveling actuator. According to the distortion type, the preset interference noise intensity mapping table is queried to obtain the dynamic observation noise covariance matrix and injected into the Kalman filter update equation, so that the elevation reconstruction process can adapt to noise fluctuations under different interference environments and improve the stability of state estimation. Attached Figure Description
[0017] Figure 1 This is a flowchart of the feature vector extraction process of the present invention; Figure 2 This is a flowchart of the feature fusion and distortion recognition process of the present invention; Figure 3This is a flowchart of the elevation benchmark reconstruction process of the present invention; Figure 4 A flowchart for generating leveling control commands according to the present invention; Figure 5 This is a flowchart of the driving process for the leveling actuator of the present invention; Figure 6 This is a flowchart of the scattering distortion spot compensation process of the present invention. Detailed Implementation
[0018] refer to Figure 1 In one embodiment, when a laser-positioned construction site leveling robot performs leveling operations, it first captures a rotating laser beam emitted by a rotating laser emitter using a laser receiving array. The laser receiving array is arranged at non-uniform intervals along the circumference of the leveling robot's body, forming a multi-layered ring-shaped topology. This multi-layered ring-shaped topology contains at least three layers of concentrically arranged photoelectric detection nodes. The number of photoelectric detection nodes in each ring increases with the ring radius, and the photoelectric detection nodes in adjacent layers are staggered in the circumferential direction. Each photoelectric detection node synchronously collects the light intensity amplitude of the rotating laser beam, and the collection frequency is an integer multiple of the rotation frequency of the rotating laser emitter, ensuring that each photoelectric detection node collects a complete laser spot signal at least once within each rotation cycle.
[0019] Specifically, based on the spatial coordinates of each photoelectric detection node within the multi-layered ring topology, the corresponding light intensity amplitude is spatially interpolated and mapped. A Cartesian coordinate system is established with the center of the leveling robot's body as the origin, and the spatial coordinates of each photoelectric detection node are represented as follows: ,in This is the numbering of the photoelectric detection node. and The coordinates are in the horizontal direction. represents the vertical coordinate. For any point to be interpolated... Its light intensity amplitude is calculated using the bilinear interpolation method, and the calculation formula is as follows:
[0020] in, These are the points to be interpolated. The light intensity amplitude of the four nearest neighbor photodetector nodes, These are the horizontal coordinates of the four photoelectric detection nodes. By performing the above bilinear interpolation operation on all points to be interpolated within the entire laser receiver array coverage area, a size of [value missing] is generated. Two-dimensional light intensity distribution matrix of light spot ,in and These represent the number of rows and columns of the matrix, respectively.
[0021] Two-dimensional light intensity distribution matrix of the light spot Principal component analysis was performed to extract principal component eigenvalues. First, the two-dimensional light intensity distribution matrix of the light spot was... Expand into a one-dimensional vector by columns Calculate the mean vector of this one-dimensional vector. Covariance Matrix :
[0022] in, A one-dimensional vector The One element, For the whole Vector. For the covariance matrix Perform eigenvalue decomposition to obtain eigenvalues. and the corresponding feature vector Before selection The largest eigenvalue is taken as the principal component eigenvalue, where The number of principal components is preset to satisfy the previous condition. The cumulative contribution rate of each feature value is greater than or equal to a preset threshold.
[0023] Calculate the signal-to-noise ratio characteristics of each photoelectric detection node, the first... Signal-to-noise ratio of each photoelectric detection node Defined as the ratio of the mean of the light intensity amplitude collected by this node to the standard deviation of the noise:
[0024] in, For the first The average light intensity amplitude collected by each photoelectric detection node within one rotation cycle. The corresponding noise standard deviation is given. The signal-to-noise ratio (SNR) characteristics of all photoelectric detection nodes are concatenated into an SNR feature vector. ,in This represents the total number of photoelectric detection nodes in the laser receiver array. The principal component eigenvalue vector... With signal-to-noise ratio eigenvectors The data is spliced together to form a two-dimensional light intensity distribution feature vector. .
[0025] The inertial measurement unit outputs the leveling robot's pose data in real time, including the roll angle. Pitch angle Yaw angle and angular velocity components The sampling frequency of the inertial measurement unit is synchronized with the acquisition frequency of the laser receiving array to ensure that the pose data and laser signal data are strictly aligned in time.
[0026] refer to Figure 2The feature fusion and distortion recognition component fuses the two-dimensional light intensity distribution feature vector with the pose data in the feature space. First, it fuses the roll angle from the pose data... Pitch angle and yaw angle Parameters converted into quaternion attitude representation vectors The conversion formula is:
[0027] in, Let be the real part of the quaternion. It represents the imaginary part of the quaternion.
[0028] Quaternion attitude representation vector With the two-dimensional light intensity distribution eigenvector Constructing multimodal feature tensors by performing tensor product operations :
[0029] in, Represents tensor product operation, multimodal feature tensor The dimension is .
[0030] Multimodal feature tensors are learned through manifold learning. Nonlinear dimensionality reduction mapping is performed to obtain a low-dimensional fused feature representation. In this embodiment, a local linear embedding algorithm is used for dimensionality reduction. First, the multimodal feature tensor is... Expand into a one-dimensional vector by rows Then for each sample point , search for it Calculate the reconstructed weight matrix using the nearest neighbor sample points. This minimizes the reconstruction error.
[0031] in, For the first The first sample point The nearest neighbor sample points, For the corresponding reconstruction weights, satisfying Based on the reconstructed weight matrix Compute low-dimensional embedding vectors This minimizes the reconstruction error in the low-dimensional space.
[0032] in, It is a low-dimensional fusion feature matrix. Let be the total number of samples. Using the aforementioned local linear embedding algorithm, the high-dimensional multimodal feature tensor is mapped to a tensor of dimension . Low-dimensional fusion feature representation ,in The preset low-dimensional space dimension.
[0033] The pre-built interference identification model is a Support Vector Data Description (SVM) classifier based on low-dimensional fused feature representations. The SVM classifier finds a minimum-volume hypersphere in the feature space, containing normal samples within the hypersphere and placing abnormal samples outside it. The center of the hypersphere is... , radius is This is obtained by solving the following optimization problem: in, For penalty parameters, These are slack variables.
[0034] Support Vector Data Description (SVM) classifiers compute low-dimensional fused feature representations. To the center of the super sphere Mahalanobis distance is used to determine the type of distortion. Mahalanobis distance The calculation formula is:
[0035] in, This is the covariance matrix of the normal samples. Three distance thresholds are preset. ,satisfy .when When the laser signal is determined to be distortion-free; when When the distortion type is determined to be occlusion; when When the distortion type is determined to be scattering; when At that time, the distortion type was determined to be diffuse reflection.
[0036] refer to Figure 3 The elevation datum reconstruction component reconstructs the current elevation datum based on distortion type and historical valid laser reference data, using Kalman filtering. First, it queries a preset interference noise intensity mapping table according to the distortion type to obtain a dynamic observation noise covariance matrix matching the current interference environment. The interference noise intensity mapping table is pre-constructed by collecting laser signal data under different interference environments and statistically analyzing its noise characteristics. The table includes the observation noise covariance matrix parameters corresponding to four distortion types: no distortion, occlusion, scattering, and diffuse reflection.
[0037] In this embodiment, the mapping relationship of interference noise intensity is shown in Table 1.
[0038] Table 1 Interference Noise Intensity Mapping Table
[0039] in, The noise variance of the elevation observations. The noise variance of the elevation change rate observation is represented by the subscripts 0, 1, 2, and 3, which correspond to four distortion types: no distortion, occlusion, scattering, and diffuse reflection, respectively.
[0040] The dynamic observation noise covariance matrix This is injected into the update equation of the Kalman filter. The state vector of the Kalman filter is defined as... ,in For the first The estimated elevation state at time [time]. For the first The rate of change of elevation at any given time. A state transition prediction equation is constructed using historical, valid laser reference data:
[0041] in, For the first The prior state estimate at time t. For the first The posterior state estimate at time t. Here is the state transition matrix:
[0042] is the sampling time interval for the Kalman filter.
[0043] The prediction equation for the prior covariance matrix is:
[0044] in, For the first The prior covariance matrix at time t, For the first The posterior covariance matrix at time t. Let be the process noise covariance matrix, and be a constant matrix.
[0045] The update equation for the Kalman filter is:
[0046] in, For the first Kalman gain coefficient at time 10:00 For the observation matrix:
[0047] For the first The observation vector at time moment includes the elevation observation value and elevation change rate observation value extracted from the laser signal. It is an identity matrix.
[0048] Kalman filtering iteratively outputs an elevation state estimate based on the state transition prediction equation and the update equation containing the covariance matrix of dynamic observation noise. and covariance estimates The estimated elevation status value As the current elevation benchmark.
[0049] refer to Figure 4 The leveling control command generation component calculates the cutting depth of the leveling blade based on the current elevation benchmark and the local ground elevation prediction model obtained by laser scanning, and generates adaptive leveling control commands including feedforward compensation. The laser scanning component acquires a 3D point cloud of the ground at a preset distance in front of the leveling blade. The scanning direction of the laser scanning component is consistent with the traveling direction of the leveling robot, and the scanning range covers the entire working width of the leveling blade.
[0050] A local surface fitting method is used to generate a local ground elevation prediction model from the 3D point cloud. In this embodiment, a quadratic surface is used for local fitting, and the equation of the quadratic surface is:
[0051] in, The coordinates of the data points in the 3D point cloud. Let be the coefficients of the quadratic surface. The coefficients of the quadratic surface are solved using the least squares method to minimize the sum of squared fitting errors:
[0052] in, For the coefficient vector, For the first The coordinates of a 3D point cloud data point This represents the total number of 3D point cloud data points involved in the fitting process. The coefficient vector is obtained by solving this problem. Then, a local ground elevation prediction model is obtained. .
[0053] Extract the elevation change gradient along the whole parallel advance direction from the local ground elevation prediction model. Let the whole parallel advance direction be... In the positive direction of the axis, the elevation change gradient is... For the local ground elevation prediction model Partial derivatives:
[0054] The cutting depth of the leveling blade is obtained by performing a difference calculation between the elevation change gradient and the current elevation benchmark. :
[0055] in, This refers to the horizontal coordinates of the current position of the leveling blade.
[0056] The feedforward compensation quantity is generated based on the derivative term of the elevation change gradient. The derivative term of the elevation change gradient is the ground local elevation prediction model's... Second-order partial derivatives:
[0057] Feedforward compensation It is proportional to the derivative of the elevation gradient:
[0058] in, This is the preset feedforward gain coefficient.
[0059] Cut depth With feedforward compensation Superimposed to form adaptive leveling control commands :
[0060] refer to Figure 5 The leveling actuator drives the leveling blade according to adaptive leveling control commands. The leveling actuator includes a left hydraulic cylinder and a right hydraulic cylinder that independently drive the left and right ends of the leveling blade to rise and fall. The cut depth and feedforward compensation in the adaptive leveling control commands are analyzed to calculate the independent target displacements at both ends of the leveling blade. The target attitude of the leveling blade is determined by both the cut depth and roll angle, and the target displacement of the left hydraulic cylinder is calculated. and the target displacement of the right hydraulic cylinder They are respectively:
[0061] in, To level the width of the scraper, This is the roll angle output by the inertial measurement unit.
[0062] An independent target displacement is input into an asymmetric proportional valve control model. Based on this independent target displacement, the asymmetric proportional valve control model adjusts the flow rate and pressure in the rodless chambers of the left and right hydraulic cylinders respectively. The asymmetric proportional valve control model considers the influence of the difference in area between the two chambers on the flow rate and pressure. By establishing the flow continuity equation and force balance equation of the hydraulic cylinder, precise control of the hydraulic cylinder displacement is achieved. The flow continuity equation of the asymmetric proportional valve control model is:
[0063] in, and These represent the flow rates of the rodless and rod-side chambers of the left hydraulic cylinder, respectively. and These are the effective areas of the rodless chamber and the rod chamber of the left hydraulic cylinder, respectively. This represents the displacement of the piston in the left hydraulic cylinder. and These are the internal leakage coefficients of the left hydraulic cylinder. and These are the external leakage coefficients of the left hydraulic cylinder. and These represent the pressures in the rodless and rod-side chambers of the left hydraulic cylinder, respectively. and These are the volumes of the rodless chamber and the rod chamber of the left hydraulic cylinder, respectively. Let be the bulk modulus of the hydraulic oil. The flow continuity equation for the right hydraulic cylinder is the same as that for the left hydraulic cylinder.
[0064] The force balance equations for the asymmetric proportional valve control model are:
[0065] in, The total mass of the left hydraulic cylinder piston and load. The viscous damping coefficient is... For the load spring stiffness, This is the external load force. The force balance equation for the right hydraulic cylinder is the same as that for the left hydraulic cylinder.
[0066] By solving the above flow continuity equation and force balance equation, the proportional valve opening commands for controlling the left and right hydraulic cylinders are obtained. The flow rate and pressure of the rodless chambers of the left and right hydraulic cylinders are adjusted respectively to achieve differential synchronous drive of the leveling blade position.
[0067] In this embodiment, a rotating laser beam is captured by a multi-layered ring-shaped laser receiving array, and a two-dimensional light intensity distribution feature vector is extracted. Multimodal feature fusion is performed using pose data output from an inertial measurement unit. A support vector data description classifier is used to determine the distortion type of the laser signal. Based on the distortion type, a dynamic observation noise covariance matrix is obtained and injected into a Kalman filter to reconstruct the current elevation benchmark. The cutting depth is calculated based on a local ground elevation prediction model, and an adaptive leveling control command including feedforward compensation is generated. A differential synchronous drive of the leveling blades is achieved by driving the left and right hydraulic cylinders through an asymmetric proportional valve control model. The above technical solution fully realizes the basic functions of a laser-positioned construction ground leveling robot, enabling ground leveling operations in complex construction site environments.
[0068] refer to Figure 6In one optional embodiment, during the extraction of the two-dimensional intensity distribution feature vector of the laser spot from the laser receiving array, when the support vector data description classifier determines the distortion type to be scattering, the coordinates of edge pixels whose intensity amplitude crosses a preset attenuation threshold in the two-dimensional intensity distribution matrix of the laser spot are extracted. The preset attenuation threshold is a preset proportion of the intensity amplitude at the center of the normal laser spot. All pixels in the two-dimensional intensity distribution matrix of the laser spot are traversed. When the intensity amplitude of a pixel is greater than or equal to the preset attenuation threshold and the intensity amplitude of its adjacent pixels is less than the preset attenuation threshold, the pixel is marked as an edge pixel, and its coordinates are recorded. .
[0069] The edge contour of the light spot is obtained by fitting the edge pixel coordinates. In this embodiment, the least squares ellipse fitting method is used to fit the edge contour of the light spot. The general equation of an ellipse is:
[0070] in, Let be the coefficients of the ellipse, satisfying The coefficients of the ellipse equation are solved using the least squares method to minimize the sum of the squared distances from all edge pixels to the ellipse.
[0071] in, Let be the total number of edge pixels. After solving for the elliptic coefficients, the elliptic equation of the light spot edge contour is obtained.
[0072] Calculate the rate of change of the normal gradient at each point on the edge contour of the light spot. For any point on the ellipse... Its normal vector is The intensity gradient in the normal direction is the directional derivative of the intensity distribution function in the normal direction:
[0073] in, Let be the gradient vector of the light intensity distribution function. Let be the normal vector. The rate of change of the normal gradient is the derivative of the normal light intensity gradient along the elliptical contour:
[0074] in, This is the arc length parameter of the elliptical profile.
[0075] By combining the spatial distribution density of corresponding photodetector nodes in a multi-layered ring topology, an inversion model for the refractive index of the scattering medium is constructed. The refractive index of the scattering medium... The following relationship exists between the rate of change of the normal gradient of the light spot edge contour and the spatial distribution density of the photodetector nodes:
[0076] in, The refractive index of air, This is a preset proportional coefficient. The average spatial distribution density of photoelectric detection nodes within the area covered by the light spot. For the first The absolute value of the rate of change of the normal gradient at each edge pixel.
[0077] The beam spot center offset compensation vector is output based on the refractive index inversion model of the scattering medium. The beam spot center offset is proportional to the difference in refractive index of the scattering medium, and the offset direction is related to the laser incident direction and the distribution direction of the scattering medium. (Beam spot center offset compensation vector) The calculation formula is:
[0078] in, The preset offset compensation coefficient is used. is the unit vector of the laser incident direction.
[0079] The spot center offset compensation vector is superimposed onto the two-dimensional light intensity distribution feature vector. Specifically, the two components of the spot center offset compensation vector are added as additional features to the end of the two-dimensional light intensity distribution feature vector to obtain the compensated two-dimensional light intensity distribution feature vector. .
[0080] In this embodiment, the relationship between the refractive index of the scattering medium and the offset of the light spot center is shown in Table 2.
[0081] Table 2. Relationship between refractive index of scattering medium and offset of light spot center
[0082] Furthermore, before constructing a multimodal feature tensor by performing a tensor product operation between the quaternion attitude representation vector and the two-dimensional light intensity distribution feature vector, the angular velocity component in the pose data is extracted. The angular velocity component is input into the hyperbolic tangent activation function to calculate the attention weight factor. Attention weight factor The calculation formula is:
[0083] in, This is the preset angular velocity gain coefficient.
[0084] Identify the feature dimensions corresponding to high-frequency light intensity changes in the two-dimensional light intensity distribution feature vector. Perform a Fast Fourier Transform on the two-dimensional light intensity distribution feature vector to obtain its frequency domain representation. Select the feature dimensions corresponding to the high-frequency components in the frequency domain as the feature dimensions of high-frequency light intensity changes. Let the set of feature dimensions of high-frequency light intensity changes be denoted as... Then for the two-dimensional light intensity distribution feature vector The first in each feature component ,like Then, this characteristic component corresponds to the high-frequency light intensity change.
[0085] Attention weight factor By multiplying element-wise the feature dimension corresponding to the high-frequency light intensity change in the two-dimensional light intensity distribution feature vector, an adaptive weighted light intensity feature vector is obtained. :
[0086] in, The first of the adaptive weighted light intensity feature vectors Each feature component.
[0087] The adaptive weighted light intensity feature vector Replacement of two-dimensional light intensity distribution feature vector With quaternion attitude representation vector Perform tensor product operations to construct multimodal feature tensors. The subsequent processes, such as manifold learning dimensionality reduction, support vector data description classifier for determining distortion type, and elevation benchmark reconstruction, are the same as those in the aforementioned embodiments.
[0088] In this embodiment, to address the issue of spot center offset caused by scattering distortion, a refractive index inversion model of the scattering medium is constructed by extracting the spot edge contour and calculating the normal gradient change rate. The resulting spot center offset compensation vector is then superimposed onto the two-dimensional light intensity distribution feature vector, improving the accuracy of laser signal feature extraction. Simultaneously, an attention weighting factor is calculated using the angular velocity component to adaptively weight the high-frequency light intensity variation feature dimension in the two-dimensional light intensity distribution feature vector. This enhances the adaptability to changes in robot motion state during feature fusion, further improving the accuracy of distortion recognition.
[0089] In another alternative embodiment, while injecting the dynamic observation noise covariance matrix into the Kalman filter update equation, the first-order difference of the elevation state estimates at adjacent times in the historical valid laser reference data is calculated to obtain the elevation change rate. Rate of change of elevation at time for:
[0090] in, For the first The estimated elevation state value at time [time]. For the first Estimated elevation state at time t.
[0091] Constructing a height mutation detection function :
[0092] in, For the first State estimate of the rate of change of elevation at time t. This represents the time interval between adjacent moments.
[0093] When the elevation change detection function Greater than the preset mutation threshold When this occurs, the Kalman filter's gain adjustment mechanism is triggered, adjusting the Kalman gain coefficient. By forcing the matrix to zero, the update of the state estimate based on the current observation is blocked. At this point, the posterior state estimate and posterior covariance matrix of the Kalman filter are as follows:
[0094] When the elevation change detection function Less than or equal to the preset mutation threshold At that time, the Kalman gain coefficients calculated based on the dynamic observation noise covariance matrix are restored, and the normal Kalman filter update process is performed.
[0095] Furthermore, in the process of generating a local ground elevation prediction model by fitting a local surface to the 3D point cloud, the local normal vector of each data point in the 3D point cloud is calculated. For the first... 3D point cloud data points Select its Construct the covariance matrix using the nearest neighbor data points. :
[0096] in, For the first The coordinate vector of the nearest neighbor data points, This is the mean vector of the nearest neighbor data points. For the covariance matrix... Eigenvalue decomposition yields three eigenvalues. and the corresponding feature vector Minimum eigenvalue corresponding feature vector That is, the first Local normal vectors of data points.
[0097] The angular deviation between the local normal vectors of each data point within the statistical neighborhood is calculated. For the ... The data point, its neighborhood number is the _th data point The local normal vector of each data point is The angle between the two normal vectors for:
[0098] Calculate the first The average of the angles between the normal vectors of each data point and the normal vectors of all data points in its neighborhood. , which is the deviation of the angle between the normal vectors of the data point.
[0099] When the angle between the normal vectors deviates Exceeding the preset consistency threshold At this point, the corresponding data point is marked as an isolated feature point. The local region containing the isolated feature point is then subdivided into four equal-sized sub-grids. Each sub-grid is assigned a weight coefficient negatively correlated with the included angle deviation. :
[0100] in, This is the preset weight decay coefficient.
[0101] In this embodiment, the relationship between the local surface fitting weight coefficient and the deviation of the angle between the normal vectors is shown in Table 3.
[0102] Table 3. Relationship between local surface fitting weight coefficients and deviation of the angle between the normal vectors.
[0103] The local ground elevation prediction model is updated by re-performing the least squares local surface fitting based on a weighted subgrid. The objective function of the weighted least squares fitting is:
[0104] Solving the weighted least squares problem above yields the updated quadratic surface coefficient vector. This yields the updated local ground elevation prediction model. The subsequent processes for extracting elevation change gradients, calculating cutting depth, generating feedforward compensation, and generating leveling control commands are the same as in the aforementioned embodiments.
[0105] In this embodiment, a gain adjustment mechanism for the Kalman filter is triggered by constructing an elevation change detection function. When an elevation change is detected, the update of the state estimate based on the current observation value is blocked, thus avoiding elevation benchmark jumps caused by sudden strong interference. Simultaneously, during the local surface fitting process, isolated feature points are marked by calculating the deviation of the local normal vector angle. The local area containing isolated feature points is then subdivided into grids and assigned negative correlation weight coefficients, improving the fitting accuracy of the local ground elevation prediction model for complex terrain, thereby enhancing the accuracy of the leveling blade cutting depth calculation.
Claims
1. A laser-based ground leveling robot for building construction, characterized in that, Includes: a laser receiving array that captures the rotating laser beam and extracts the two-dimensional intensity distribution feature vector of the light spot; Inertial measurement unit, outputting pose data of the leveling robot; The feature fusion and distortion recognition component fuses the two-dimensional light intensity distribution feature vector with the pose data in the feature space, and determines the distortion type of the current laser signal through a pre-built interference recognition model. The distortion type includes occlusion, scattering or diffuse reflection. The elevation benchmark reconstruction component reconstructs the current elevation benchmark based on the aforementioned distortion type and historical valid laser benchmark data, using Kalman filtering. The leveling control command generation component calculates the cutting depth of the leveling blade based on the current elevation benchmark and the local ground elevation prediction model obtained by laser scanning, and generates an adaptive leveling control command that includes feedforward compensation. The leveling actuator drives the leveling blade to move according to the adaptive leveling control command.
2. The laser-positioning-based construction ground leveling robot according to claim 1, characterized in that, The laser receiving array is arranged at non-uniform intervals along the circumference of the leveling robot body to form a multi-layer ring topology. The two-dimensional light intensity distribution feature vector is extracted in the following way: each photoelectric detection node in the laser receiving array synchronously collects the light intensity amplitude of the rotating laser beam. Based on the spatial coordinate position of each photoelectric detection node in the multi-layer ring topology, the corresponding light intensity amplitude is spatially interpolated and mapped to generate a two-dimensional light intensity distribution matrix of the light spot. Principal component analysis is performed on the two-dimensional light intensity distribution matrix of the light spot to extract the principal component feature values. The principal component feature values are concatenated with the signal-to-noise ratio features of each photoelectric detection node to form the two-dimensional light intensity distribution feature vector.
3. The laser-positioning-based building construction ground leveling robot according to claim 2, characterized in that, The feature fusion and distortion recognition component fuses the two-dimensional light intensity distribution feature vector with the pose data in the feature space as follows: the roll angle, pitch angle and yaw angle parameters in the pose data are converted into quaternion pose representation vectors; the quaternion pose representation vectors and the two-dimensional light intensity distribution feature vectors are subjected to tensor product operation to construct a multimodal feature tensor; and the multimodal feature tensor is subjected to nonlinear dimensionality reduction mapping through manifold learning to obtain a low-dimensional fused feature representation. The pre-built interference identification model is a support vector data description classifier built based on the low-dimensional fusion feature representation. The support vector data description classifier determines the distortion type by calculating the Mahalanobis distance from the low-dimensional fusion feature representation to the center of the hypersphere.
4. The laser-positioning-based building construction ground leveling robot according to claim 1, characterized in that, The elevation benchmark reconstruction component reconstructs the current elevation benchmark based on the distortion type and historical valid laser reference data using Kalman filtering as follows: It queries a preset interference noise intensity mapping table according to the distortion type to obtain a dynamic observation noise covariance matrix matching the current interference environment; it injects the dynamic observation noise covariance matrix into the update equation of the Kalman filter; it constructs a state transition prediction equation using the historical valid laser reference data; and iteratively outputs an elevation state estimate and a covariance estimate based on the state transition prediction equation and the update equation containing the dynamic observation noise covariance matrix. The elevation state estimate is then used as the current elevation benchmark.
5. The laser-positioning-based building construction ground leveling robot according to claim 1, characterized in that, The method by which the leveling control command generation component calculates the cutting depth of the leveling blade and generates an adaptive leveling control command including feedforward compensation based on the current elevation benchmark and the local ground elevation prediction model obtained by laser scanning is as follows: the laser scanning component obtains a three-dimensional point cloud of the ground at a preset distance in front of the leveling blade; the three-dimensional point cloud is fitted with a local surface to generate the local ground elevation prediction model; the elevation change gradient along the leveling advance direction in the local ground elevation prediction model is extracted; the elevation change gradient is differentially calculated with the current elevation benchmark to obtain the cutting depth; the feedforward compensation is generated based on the derivative of the elevation change gradient; and the cutting depth and the feedforward compensation are superimposed to form the adaptive leveling control command.
6. The laser-positioning-based construction ground leveling robot according to claim 5, characterized in that, The leveling actuator drives the leveling blade according to the adaptive leveling control command in the following way: The leveling actuator includes a left hydraulic cylinder and a right hydraulic cylinder that independently drive the left and right ends of the leveling blade to rise and fall. It analyzes the cutting depth and feedforward compensation amount in the adaptive leveling control command, calculates the independent target displacement of the left and right ends of the leveling blade, and inputs the independent target displacement to the asymmetric proportional valve control model. The asymmetric proportional valve control model adjusts the flow rate and pressure of the rodless chamber of the left hydraulic cylinder and the right hydraulic cylinder respectively according to the independent target displacement, so as to realize the differential synchronous drive of the leveling blade posture.
7. The laser-positioning-based building construction ground leveling robot according to claim 2, characterized in that, During the process of extracting the two-dimensional light intensity distribution feature vector of the laser receiving array, when the support vector data description classifier determines that the distortion type is scattering, the coordinates of the edge pixels whose light intensity amplitude crosses the preset attenuation threshold in the two-dimensional light intensity distribution matrix of the light spot are extracted, the edge pixel coordinates are fitted to obtain the edge contour of the light spot, the normal gradient change rate of each point on the edge contour of the light spot is calculated, and the spatial distribution density of the corresponding photoelectric detection nodes in the multi-layer ring topology is combined to construct the refractive index inversion model of the scattering medium. The center offset compensation vector of the light spot is output according to the refractive index inversion model of the scattering medium, and the center offset compensation vector of the light spot is superimposed on the two-dimensional light intensity distribution feature vector.
8. The laser-positioning-based building construction ground leveling robot according to claim 3, characterized in that, Before constructing a multimodal feature tensor by performing a tensor product operation between the quaternion pose representation vector and the two-dimensional light intensity distribution feature vector, the angular velocity component in the pose data is extracted. The angular velocity component is input into the hyperbolic tangent activation function to calculate the attention weight factor. The attention weight factor is then multiplied element-wise with the feature dimension corresponding to the high-frequency light intensity change in the two-dimensional light intensity distribution feature vector to obtain an adaptive weighted light intensity feature vector. The adaptive weighted light intensity feature vector is then used to replace the two-dimensional light intensity distribution feature vector and perform a tensor product operation with the quaternion pose representation vector.
9. The laser-positioning-based building construction ground leveling robot according to claim 4, characterized in that, While injecting the dynamic observation noise covariance matrix into the Kalman filter update equation, the first-order difference of the elevation state estimate values at adjacent times in the historical effective laser reference data is calculated to obtain the elevation change rate. An elevation change detection function is constructed. When the elevation change rate is greater than a preset change threshold, the Kalman filter gain adjustment mechanism is triggered, and the Kalman gain coefficient is forcibly set to zero to block the update of the state estimate by the current observation value. When the elevation change rate is less than or equal to the preset change threshold, the Kalman gain coefficient calculated based on the dynamic observation noise covariance matrix is restored.
10. The laser-positioning-based construction ground leveling robot according to claim 5, characterized in that, In the process of generating a local ground elevation prediction model by performing local surface fitting on the three-dimensional point cloud, the local normal vector of each data point in the three-dimensional point cloud is calculated, and the angular deviation between the local normal vectors of each data point in the neighborhood is statistically analyzed. When the angular deviation exceeds a preset consistency threshold, the corresponding data point is marked as an isolated feature point. The local area containing the isolated feature point is subdivided into grids, and the subdivided subgrids are assigned weight coefficients that are negatively correlated with the angular deviation. Based on the weighted subgrids, the least squares local surface fitting is re-executed to update the local ground elevation prediction model.
