Three-dimensional profile detection and deviation analysis method for target vehicle

Abstract

The application discloses a three-dimensional profile detection and deviation analysis method of a target vehicle, and belongs to the technical field of profile detection, and comprises the following steps: synchronously collecting initial point cloud data and motion data, mapping to a constructed vehicle body dynamic coordinate system, and performing data alignment verification; calculating an instantaneous motion intensity index, generating a theoretical deviation value by using a dynamic coupling error model, and performing reverse deviation correction; identifying an instantaneous interference event, performing adaptive statistical filtering on the vehicle body point cloud data in the interference period; assigning real-time confidence weights, performing weighted point cloud registration and pose solution to generate initial pose data; obtaining an inertial predicted pose, calculating a pose estimation residual error, using a Kalman filter to update and feed back in real time, and generating a pose sequence; and according to the pose sequence, adopting a point cloud fusion algorithm to perform three-dimensional profile reconstruction and reconstruction deviation analysis, effectively eliminating the influence of dynamic coupling and instantaneous interference, and avoiding pose deviation accumulation.

Description

Technical Field

[0001] This invention relates to a method for three-dimensional contour detection and deviation analysis of a target vehicle, belonging to the field of contour detection technology. Background Technology

[0002] In the fields of industrial manufacturing and intelligent equipment, the detection of the 3D contour of a target vehicle relies on accurate pose estimation to achieve data alignment. Traditional detection methods mostly depend on static or offline measurements, which are not suitable for moving vehicles on the production line or vehicles in motion, such as in-transit detection. With the development of multi-sensor fusion and intelligent algorithms, pose estimation technology is gradually expanding into dynamic scenes. Combining multi-source data such as vision and inertia improves response speed and enhances dynamic adaptability to a certain extent, providing technical support for real-time detection of moving vehicles.

[0003] However, existing technologies do not consider the dynamic coupling relationship between vehicle motion state and sensor data in dynamic scenarios, as well as the cumulative impact of instantaneous disturbances on estimation results. Specifically, although existing solutions optimize data fusion efficiency, they do not fully correlate vehicle motion acceleration, angular velocity and sensor data deviations, and also ignore the impact of sudden disturbances, such as instantaneous vibrations, on parameter calibration during dynamic processes. This leads to the accumulation of deviations in pose estimation, making it impossible to accurately match the real-time motion state of the vehicle, ultimately affecting the overall accuracy of 3D contour detection and making it difficult to meet the detection requirements of high-precision moving vehicles. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a method for three-dimensional contour detection and deviation analysis of target vehicles. By compensating for coupling deviations through motion intensity, and based on adaptive filtering of interference energy, weighted registration and pose calculation are performed. Parameters are updated in real time to prevent error accumulation, and finally the contour is reconstructed, deviations and defects are analyzed, thus solving the problem of pose deviation accumulation caused by dynamic coupling and instantaneous interference.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The method for three-dimensional contour detection and deviation analysis of the target vehicle includes:

[0007] Initial point cloud data and motion data are collected synchronously, mapped to the constructed vehicle dynamic coordinate system, and data alignment is verified to generate vehicle point cloud data and inertial motion data.

[0008] The instantaneous motion intensity index is calculated, and the theoretical deviation value of the vehicle body point cloud data at the current moment is generated using the dynamic coupling error model, and the reverse deviation correction is performed.

[0009] Identify instantaneous interference events, record the interference period, calculate the interference energy, and perform adaptive statistical filtering on the vehicle body point cloud data during the interference period.

[0010] Assign real-time confidence weights and perform weighted point cloud registration and pose calculation to generate initial pose data;

[0011] The inertial predicted pose is obtained, the pose estimation residual is calculated, and the parameters and extrinsic parameter matrix of the dynamic coupling error model are updated and fed back using a Kalman filter to generate a pose sequence.

[0012] Based on the pose sequence, a point cloud fusion algorithm is used to reconstruct the three-dimensional contour, generate a three-dimensional contour model of the vehicle body, and perform reconstruction deviation analysis to generate a comprehensive vehicle body inspection report.

[0013] Specifically, the steps for data alignment verification include:

[0014] The initial point cloud data is segmented using feature points to obtain the initial centroid;

[0015] Obtain the theoretical centroid of the target vehicle, perform dynamic centroid correction on the initial centroid, generate the centroid correction amount, and calculate the actual centroid coordinates;

[0016] A dynamic coordinate system for the vehicle body is established with the actual centroid coordinates as the origin;

[0017] Acquire a dynamic calibration target and perform feature point matching with the calibration target point cloud data in the initial point cloud data to generate an initial calibration point pair;

[0018] By using the sensor extrinsic parameter matrix, the calibration target point cloud data is mapped to the vehicle body dynamic coordinate system to obtain the mapped point cloud coordinates;

[0019] Based on the initial calibration point pair, the corrected extrinsic parameter matrix is ​​generated using the extrinsic parameter optimization model;

[0020] The initial point cloud data is mapped to the vehicle dynamic coordinate system using the corrected extrinsic parameter matrix to obtain vehicle point cloud data. The positional deviation between the motion data and the actual centroid coordinates is corrected to generate inertial motion data.

[0021] Specifically, the data alignment verification steps also include:

[0022] Extract vehicle body feature points, obtain the coordinates of vehicle body feature points in adjacent frames, and calculate the difference in coordinate changes;

[0023] Based on the motion direction of the inertial motion data, the spatial correlation error is calculated; if the spatial correlation error is not greater than the spatial error threshold, the spatial verification is deemed to have passed; otherwise, a data resampling command is triggered.

[0024] Obtain the timestamp interval. If the timestamp interval is not greater than the verification multiple of the sampling period, the time verification is deemed to have passed; otherwise, a data resampling instruction is triggered.

[0025] The number of point clouds in a single frame of vehicle body point cloud data and the number of complete parameters in a single set of inertial motion data are counted. The number of missing parameters is calculated, the theoretical number of parameters is obtained, and the data missing rate is calculated.

[0026] If the data missing rate is not greater than the data missing threshold, the data verification is deemed successful; otherwise, a data re-sampling instruction is triggered.

[0027] Specifically, the steps for correcting reverse bias include:

[0028] Extract the acceleration and angular velocity components at the current sampling moment, remove the static components, compare them one by one with the classification criteria of motion state, and record the current motion state label.

[0029] Obtain the acceleration and angular velocity components corresponding to the current motion state, and call the corresponding component weights to calculate the basic intensity of the current motion state;

[0030] Calculate the intensity of dynamic change based on the instantaneous rate of change of the current motion state;

[0031] The instantaneous motion intensity is obtained by weighted summation of the base intensity and the dynamically changing intensity.

[0032] Based on the motion state labels, the corresponding sub-model is called from the pre-trained dynamic coupling error model to generate theoretical deviation values;

[0033] In the vehicle body point cloud data, the correspondence in the initial calibration point pair is used to obtain the vehicle body calibration point cloud corresponding to the prime number dynamic calibration target, and the actual deviation is calculated.

[0034] The model deviation is calculated by the difference between the theoretical deviation value and the actual deviation value, and the sub-model is adjusted to generate a coupling deviation value predicted by the fine-tuned sub-model.

[0035] Specifically, the steps for correcting reverse bias also include:

[0036] The point cloud region is divided and a region compensation coefficient is set to correct the coupling deviation value and obtain the region compensation value. For the point cloud points in each point cloud region, the region compensation value is used to perform reverse compensation to generate a compensated vehicle body point cloud.

[0037] Feature points corresponding to the dynamic calibration target are extracted from the point cloud of the compensated vehicle body, and the compensation error is calculated.

[0038] If the compensation error in all directions is not greater than the compensation error threshold, the compensation is deemed qualified.

[0039] If the compensation error in any direction exceeds the compensation error threshold, a secondary correction process is initiated to calculate the secondary correction value and generate a new compensated vehicle point cloud, until the compensation error in all directions meets the threshold requirement.

[0040] Specifically, the steps for identifying transient disturbance events include:

[0041] By using wavelet transform for denoising, the inertial motion data is split into low-frequency approximate components and high-frequency detail components.

[0042] Set a soft threshold to quantize high-frequency detail components and generate transient interference characteristic signals;

[0043] Traverse the instantaneous interference feature signals, extract the acceleration interference amplitude and angular velocity interference amplitude at each moment, and use dual thresholds to initially screen suspected interference events;

[0044] For suspected interference events, calculate the interference energy, analyze the signal waveform of the suspected interference event to determine the type of interference, and generate interference event markers;

[0045] The corresponding filtering algorithm is invoked based on the type of interference, and the filtering parameters are calculated through parameter matching.

[0046] By combining point cloud region labels, the compensation vehicle body point cloud during the interference period is divided into corresponding regions. At the same time, based on the point cloud curvature threshold, detail sub-regions and planar sub-regions are marked in each region.

[0047] Differentiated filtering is performed for different types of interference. For point clouds that are not associated with interference events, only a first-level smoothing filter is performed to obtain a denoised vehicle body point cloud.

[0048] Specifically, the steps for weighted point cloud registration and pose calculation include:

[0049] The instantaneous motion intensity and interference energy from different perspectives are normalized to obtain normalized motion intensity and normalized interference energy, and the basic weight of each perspective is calculated.

[0050] The basic weights are corrected by combining the regional compensation coefficients corresponding to the viewpoints to obtain confidence weights, and then normalized to generate a set of confidence weights for each viewpoint.

[0051] Divide the reference point cloud and the point cloud to be registered, and use the weighted iterative nearest point algorithm to generate a fused point cloud;

[0052] Extract vehicle body feature points and calculate curvature, filter candidate feature points, and combine the relative positional relationship of candidate feature points on the vehicle body to match and obtain actual feature points;

[0053] Obtain the vehicle's position coordinates, and calculate the attitude angles by using the coordinate differences between the wheel center and the corner of the roof.

[0054] Extract the dynamic calibration target feature points from the fused point cloud, map them to the global coordinate system through the solved pose parameters, and calculate the pose accuracy error;

[0055] Once the pose accuracy error is not less than the preset accuracy threshold, the vehicle body feature points are re-extracted and the calculation is repeated.

[0056] Specifically, the steps for calculating the pose estimation residuals include:

[0057] Perform a double integral on the acceleration data in the inertial motion data to obtain the inertial predicted position increment, and perform a single integral on the angular velocity data to obtain the inertial predicted attitude angle increment.

[0058] Using the initial pose data as the initial state vector, the extended Kalman filter is enabled to perform prediction, generating an inertial predicted pose, including the predicted position and the predicted attitude angle.

[0059] The initial pose data is compared with the inertial predicted pose, and the single-dimensional residuals are calculated, including position residuals and attitude residuals.

[0060] Configure position residual weights and attitude residual weights, and calculate the comprehensive residual of pose estimation;

[0061] Once the overall residual exceeds the pose residual threshold, it is determined that there is an error accumulation trend. Then, the parameters are updated, the source of deviation is located, and the corresponding parameters are updated based on the source of deviation. The pose is then recalculated based on the updated parameters, and the residual is verified.

[0062] Specifically, the steps for reconstructing the deviation analysis include:

[0063] For each timestamp of the denoised vehicle point cloud, combined with the corresponding pose sequence, a rigid body transformation mapping is performed to map it to the global coordinate system of the detection area, and the pose confidence at each time point is calculated.

[0064] The multi-moment point cloud in the global coordinate system is fused by pose confidence and contour completion is performed based on Poisson reconstruction to generate a three-dimensional contour model of the vehicle body.

[0065] Obtain the standard model of the target vehicle, and perform global ICP registration in the global coordinate system in conjunction with the three-dimensional contour model of the vehicle body;

[0066] For the registered 3D contour model of the vehicle body and the standard model, the registration deviation vector is calculated point by point, and the absolute deviation value of each point is obtained by using the Euclidean distance formula;

[0067] Points with absolute deviation values ​​exceeding a preset defect judgment threshold are selected, and deviation regions are divided based on Euclidean distance clustering. The average deviation of each deviation region is then calculated.

[0068] Specifically, the steps for reconstructing the deviation analysis also include:

[0069] For each deviation region, extract curvature features, rate of change of normal vector, region area, and average deviation to generate the deviation geometric features of the deviation region;

[0070] Based on the constructed defect classification model, the defect type of the current target vehicle is output, and the defect severity is calculated by weighting based on the average deviation and the area of ​​the region.

[0071] Extract the time series of roll angle, pitch angle and yaw angle from the pose sequence respectively. For each set of attitude angle time series, calculate the fluctuation amplitude, fluctuation frequency, standard deviation and peak factor.

[0072] The stationarity score is calculated by weighting multiple indicators, and the stationarity level is classified.

[0073] The beneficial effects of this invention are:

[0074] By constructing a unified spatiotemporal data benchmark through pulse synchronization and extrinsic parameter mapping, the problem of sensor sampling misalignment is solved, laying a data foundation free of basic bias for subsequent processing. Motion intensity is calculated based on inertial motion data and combined with a pre-trained dynamic coupling error model to establish a correlation between vehicle motion and sensing deviation. This reverses the systematic point cloud deviation caused by motion, overcoming the shortcomings of existing technologies that do not correlate motion state with sensing deviation. Instantaneous interference is monitored through a sliding window, combined with adaptive filtering of interference energy, preserving contour details while eliminating noise, thus mitigating the impact of sudden interference on data quality. Accurate registration of multi-view point clouds is achieved through dynamic weight allocation, allowing data less affected by motion and interference to dominate pose calculation, improving initial pose accuracy. Kalman filtering is then used to update model parameters and extrinsic parameters in a closed loop, blocking error accumulation paths and preventing pose accuracy from decreasing over time. Based on accurate pose, global contour reconstruction and multi-dimensional analysis are achieved, outputting a reliable detection report. This effectively eliminates the influence of dynamic coupling and instantaneous interference, avoids pose deviation accumulation, significantly improves 3D contour detection accuracy, and meets the detection requirements of high-precision moving vehicles. Attached Figure Description

[0075] Figure 1 A flowchart of a method for three-dimensional contour detection and deviation analysis of a target vehicle;

[0076] Figure 2 This is a flowchart of the reverse deviation correction in this invention;

[0077] Figure 3 This is a flowchart illustrating the identification of transient interference events in this invention;

[0078] Figure 4 This is a flowchart of the pose estimation residual in this invention. Detailed Implementation

[0079] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0080] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a method for three-dimensional contour detection and deviation analysis of a target vehicle, including:

[0081] Step S1: By integrating sensor networks, such as 3D vision sensors and inertial measurement units, when the target vehicle starts or enters the preset detection area, based on the pulse synchronization mechanism, all sensors are triggered to perform timestamp calibration to ensure that the sampling times of all sensors are strictly aligned, thereby acquiring initial point cloud data and motion data. With the centroid of the target vehicle as the origin, a vehicle dynamic coordinate system is established. Based on the pre-calibrated extrinsic parameter matrix, the initial point cloud data collected by the 3D vision sensors and the motion data collected by the inertial measurement units are uniformly mapped to the vehicle dynamic coordinate system through coordinate rotation and translation operations to achieve consistency of spatial position reference and perform data alignment verification to generate time-synchronized and spatially aligned vehicle point cloud data and inertial motion data.

[0082] Step S2: During the target vehicle's movement, acceleration and angular velocity cause slight deformations or vibrations in the optical components of the 3D vision sensor, such as the lens and laser emitter, leading to deviations in the vehicle point cloud. Based on inertial motion data, the acceleration and angular velocity amplitudes at each sampling moment are acquired, and the instantaneous motion intensity is calculated to quantify the severity of the target vehicle's current motion. Combined with the constructed dynamic coupling error model, the theoretical deviation values ​​of the vehicle point cloud data in the three directions of the vehicle's dynamic coordinate system at the current moment are generated, and reverse deviation correction is performed to eliminate the systematic deviation introduced by the vehicle's motion state, thereby generating a compensated vehicle point cloud. Specifically, based on historical data under different vehicle speeds, steering angles, and acceleration conditions, a dynamic coupling error model is constructed to establish a quantitative mapping relationship between the acceleration-angle group and the visual measurement deviation. The input is the motion intensity, and the output is the theoretical deviation value. The dynamic coupling error model includes input... The system consists of an input layer, a feature mapping layer, and a deviation output layer. The input layer receives instantaneous motion intensity and motion state labels. The feature mapping layer uses multinomial regression as the regression function to perform a nonlinear transformation on the instantaneous motion intensity and motion state labels in the input layer, mapping them to an intermediate deviation feature vector. The deviation output layer performs a linear transformation and dimension reconstruction on the intermediate deviation feature vector, outputting the theoretical deviation values ​​of the denoised vehicle point cloud in three directions in the vehicle's dynamic coordinate system. For training the dynamic coupling error model, vehicle motion data and corresponding point cloud measurement deviation data under different vehicle speeds, steering angles, and acceleration conditions are collected to construct a training dataset. The instantaneous motion intensity and motion state labels are used as inputs, and the actual measured point cloud deviations are used as labels to train the regression model. During training, mean squared error is used as the loss function, and the model parameters are optimized through gradient descent algorithm until the error between the model's predicted deviation and the actual deviation meets the preset accuracy requirements.

[0083] Step S3: Since transient interference can easily cause equipment to generate high-frequency, short-term abnormal signals, such as vibration caused by road bumps causing burrs in the point cloud on the top of the vehicle body, a monitoring mechanism based on a sliding time window is constructed to cover the duration of common transient interferences of vehicles. High-frequency components are extracted and differential operations are performed on the inertial motion data to identify transient interference events, record the interference period, and calculate the interference energy to quantify the severity of the interference. At the same time, adaptive statistical filtering is performed on the vehicle body point cloud data during the interference period. The filtering intensity is positively correlated with the interference energy. For each point cloud point in the compensated vehicle body point cloud, based on the distribution characteristics of its neighboring points, noise points that deviate from the normal distribution are removed while retaining the detailed features of the vehicle body outline to obtain the denoised vehicle body point cloud data.

[0084] Step S4: Since the point cloud data of the 3D vision sensor is multi-view, the degree of influence of the vehicle motion intensity and instantaneous interference on the sensor at different positions is different. Real-time confidence weights are assigned to the 3D vision sensor data of each view, and weighted point cloud registration and pose calculation are performed. For the point cloud pair to be registered, the contribution of the corresponding point is adjusted according to the confidence weight. Through iterative optimization, the vehicle point clouds of different views are accurately aligned in the vehicle dynamic coordinate system to achieve multi-view point cloud fusion. Based on the registered complete point cloud, the vehicle feature points, such as the wheel center and the roof corner, are extracted. Combined with the coordinates of the feature points in the vehicle dynamic coordinate system, the pose parameters of the target vehicle are calculated to generate the initial pose data of the target vehicle, including position and attitude angle.

[0085] Step S5: Integrate the inertial motion data through the kinematic model to obtain the inertial predicted pose. Combine the initial pose data to calculate the pose estimation residual. Use the pose estimation residual as the observation input. Use a Kalman filter to estimate the correction amount of the dynamic coupling error model parameters and the differential components of the extrinsic parameter matrices of each 3D vision sensor in real time. Update the model parameters and extrinsic parameter matrices through recursive operation and feed back the updated information to achieve continuous optimization of parameters. Finally, generate the pose sequence after feedback operation.

[0086] Step S6: Based on the final pose sequence, the denoised vehicle point cloud data at each moment is mapped to a global coordinate system with the origin of the detection area as the reference through coordinate transformation. A point cloud fusion algorithm is used for 3D contour reconstruction, stitching together the point cloud data from multiple views and moments into a complete 3D vehicle contour model. Reconstruction deviation analysis is then performed, comparing the reconstructed 3D vehicle contour model point-by-point with the standard model to calculate the 3D coordinate deviation of each point. Color coding is applied based on the deviation value to generate a color map of vehicle size deviation, visually displaying the deviation distribution area. Geometric features are extracted from the deviation areas, and a trained machine learning classification model is used to classify the deviation areas, automatically identifying specific types of vehicle defects such as dents, bulges, and twists, and recording the location, size, and severity of the defects. Simultaneously, the vehicle's operational stability is evaluated based on the pose data, calculating the fluctuation amplitude and frequency, comparing them with a preset stability threshold, and assessing the target vehicle's operational stability level during the detection process. The color map of size deviation, defect identification results, and operational stability are then integrated. The assessment process generates a comprehensive vehicle inspection report, including visual charts and quantitative data, providing a basis for vehicle quality evaluation. The machine learning classification model comprises an input layer, a feature encoding layer, and a classification output layer. The input layer receives multi-dimensional geometric features extracted from the deviation area. The feature encoding layer uses a fully connected neural network to encode the input geometric features, generating high-dimensional feature vectors to capture the inherent patterns of different defect types. The classification output layer outputs the probability distribution of defect types corresponding to the deviation area, such as dents, bulges, distortions, and normal shapes, and selects the category with the highest probability as the final classification result. For training the machine learning classification model, a large amount of known defect type 3D contour data of the vehicle body is collected. A defect sample library is constructed through manual annotation or simulation. For each sample, the aforementioned geometric features are extracted as model input, and the corresponding defect type is labeled. A suitable classification algorithm, such as support vector machine, is selected, using the cross-entropy loss function. The model parameters are optimized through backpropagation or gradient boosting algorithms until the classification accuracy meets a preset threshold.

[0087] Specifically, the steps for data alignment verification include:

[0088] Feature point segmentation is performed on the initial point cloud data of the target vehicle to extract the frame symmetry plane and the line connecting the centers of the front and rear wheels. Based on the intersection of the frame symmetry plane and the line connecting the centers of the wheels, the geometric center of the vehicle body is determined as the initial centroid.

[0089] The theoretical centroid of the target vehicle and the density distribution of the point cloud on the vehicle body shell are obtained. Combined with the initial centroid, dynamic centroid correction is performed. Based on the point cloud density difference between the front and rear of the target vehicle body, the correction amount in the X direction is calculated. Based on the difference between the theoretical centroid and the initial centroid in the Y direction, the correction amount in the Y direction is calculated. Based on the point cloud density difference between the roof and chassis, the Z-direction correction is calculated. This generates the centroid correction quantity. The actual centroid coordinates are calculated by summing the centroid correction amount and the initial centroid. A dynamic coordinate system for the vehicle body is then established with the actual centroid coordinates as the origin, as shown in the following expression:

[0090]

[0091]

[0092]

[0093] In the formula, This represents the point cloud density of the front part of the vehicle body (from the front axle to the midpoint of the door). This represents the point cloud density at the rear of the vehicle body (from the rear axle to the midpoint of the door). This is the horizontal correction factor. The Y-axis coordinate of the theoretical centroid. The initial centroid's Y-axis coordinate. Cloud density at the top of the vehicle. For chassis point cloud density, This is the vertical correction factor;

[0094] The coordinates of a preset dynamic calibration target within the detection area, including multiple reflective marker points, are obtained and matched with the calibration target point cloud data in the initial point cloud data. Based on the principle of minimum Euclidean distance, point-to-point matching is performed to generate initial calibration point pairs. The calibration target point cloud data is mapped to the vehicle dynamic coordinate system through the sensor extrinsic parameter matrix to obtain the mapped point cloud coordinates.

[0095] Combining the initial calibration point pairs, an extrinsic parameter optimization model is used to calculate the transformation error between the mapped point cloud coordinates and the dynamic calibration target using Euclidean distance. This minimizes the transformation error, optimizes the rotation matrix and translation vector, and yields a corrected extrinsic parameter matrix, including the corrected rotation matrix and translation vector. The extrinsic parameter optimization model comprises an input layer, an error calculation layer, an optimization solution layer, and an output layer. The input layer receives the initial calibration point pairs, the initial extrinsic parameter matrix, the mapped point cloud coordinates, and the physical coordinates of the dynamic calibration target. The initial calibration point pairs consist of the physical coordinates of the dynamic calibration target and the corresponding target point cloud coordinates in the initial point cloud. The initial extrinsic parameter matrix is ​​the rotation matrix and translation vector obtained from the sensor's factory calibration or preliminary calibration. The mapped point cloud coordinates are the coordinate set obtained by mapping the calibration target point cloud data to the vehicle's dynamic coordinate system using the initial extrinsic parameter matrix. The physical coordinates of the dynamic calibration target are the true coordinates of preset reflective marker points within the detection area. The error calculation layer is based on the principle of minimizing Euclidean distance. The transformation error between the mapped point cloud coordinates and the physical coordinates of the dynamic calibration target is calculated, an error function is constructed, and a nonlinear optimization algorithm, such as the Levenberg-Marquardt algorithm, is used to optimize the solution layer. The error function is minimized to find the optimal rotation matrix and translation vector. The output layer is used to output the corrected extrinsic parameter matrix. Regarding the training of the extrinsic parameter optimization model, dynamic calibration targets are deployed in the detection area, and the physical coordinates of their reflective markers are obtained. Initial point cloud data is collected, the calibration target point cloud coordinates are extracted, and initial calibration point pairs are generated. Using the initial extrinsic parameter matrix, the calibration target point cloud coordinates are mapped to the vehicle dynamic coordinate system to obtain the mapped point cloud coordinates. An error function is constructed, and the Levenberg-Marquardt algorithm is used to iteratively solve the error until it converges to a preset threshold, such as a root mean square error of less than 0.01 mm. The optimized rotation matrix and translation vector are used as the corrected extrinsic parameter matrix for subsequent coordinate system transformation of point cloud data.

[0096] By using the corrected extrinsic parameter matrix, the initial point cloud data is mapped to the vehicle dynamic coordinate system to obtain standardized vehicle point cloud data. At the same time, based on the corrected translation vector, the positional deviation between the motion data and the actual centroid coordinates is corrected to convert the motion data into standardized inertial motion data, ensuring that the spatial origin of the motion data and the point cloud data are consistent. Finally, the deviation of the pre-calibrated extrinsic parameter matrix caused by the small displacement of the sensor installation is corrected, reducing the spatial mapping error.

[0097] In dynamic scenarios, occlusion and loss of historical sensor data can occur when the target vehicle passes through the monitoring area, leading to data anomalies. Therefore, standardized data needs to be verified from three dimensions: spatial correlation, temporal continuity, and data integrity. Regarding spatial correlation, vehicle feature points, such as wheel centers and roof corners, are extracted from the vehicle point cloud data. The coordinates of these feature points in adjacent frames are obtained, and the difference in coordinate changes between adjacent frames is calculated. Combined with the motion direction from inertial motion data, the consistency between the feature point coordinate changes and the motion direction is verified. The spatial correlation error is calculated, and a spatial error threshold is set. If the spatial correlation error is not greater than the spatial error threshold, the spatial verification is considered successful; otherwise, the current frame is considered abnormal data, triggering a data resampling command. The expression is as follows:

[0098]

[0099] In the formula, For spatial correlation error, , The first Frame, First The X coordinates of the vehicle body feature points in the frame. For the first Frame X-axis acceleration, The time interval is two frames.

[0100] Regarding temporal continuity, the timestamp interval for acquiring adjacent frames of vehicle point cloud data and the timestamp interval for adjacent groups of inertial motion data are determined. If the timestamp interval is not greater than the verification multiple of the sampling period, the time verification is deemed to have passed; otherwise, the time continuity is deemed to have been interrupted, triggering a data resampling instruction.

[0101] Regarding data integrity, the number of point clouds in a single frame of vehicle body point cloud data and the number of complete parameters in a single set of inertial motion data are counted. At the same time, the theoretical number of parameters is obtained, which is the sum of the theoretical number of point clouds and the theoretical number of motions. By counting the number of point clouds and the number of complete parameters, the number of missing parameters is calculated. The data missing rate is calculated by the ratio of the number of missing parameters to the theoretical number of parameters, and a data missing threshold is set. If the data missing rate is not greater than the data missing threshold, the data verification is deemed to be successful; otherwise, the current frame is deemed to be abnormal data, and a data resampling instruction is triggered.

[0102] After verification, standardized data that is time-synchronized, spatially aligned, and free of anomalies is finally generated, including vehicle point cloud data and inertial motion data.

[0103] Specifically, the steps for correcting reverse bias include:

[0104] Based on inertial motion data, the acceleration and angular velocity components at the current sampling moment are extracted. A low-pass filtering algorithm is used to separate and remove static components in the data, such as gravitational acceleration, and only the components reflecting the dynamic motion of the vehicle are retained.

[0105] The motion state is defined by the direction and amplitude of acceleration and the direction and amplitude of angular velocity, avoiding misjudgment based on a single parameter. For example, relying solely on the positive direction of X-axis acceleration cannot distinguish between simple acceleration and acceleration during turning. It is necessary to combine the thresholds of Y-axis acceleration and Z-axis angular velocity to further eliminate interference, thereby generating a classification standard for motion state. Each state has a corresponding acceleration / angular velocity threshold. The filtered acceleration and angular velocity components are compared one by one with the multiple parameters in the classification standard, and the current motion state label is recorded and associated with the vehicle point cloud data at the corresponding time. This provides a state basis for subsequent error model adaptation and regional compensation, avoiding a one-size-fits-all approach to bias prediction that cannot match the error characteristics of different motions. The specific classification standard is set by those skilled in the art.

[0106] Based on the motion state label, the corresponding acceleration and angular velocity components are obtained, and the absolute values ​​of the components are taken respectively. The corresponding component weights are called, and the basic intensity of the current motion state is calculated by weighted summation to reflect the degree of influence of the motion amplitude on the error. For example, when the X-axis is accelerating in the positive direction, the corresponding components are the X-axis acceleration and the X-axis angular velocity.

[0107] The instantaneous rate of change of the acceleration and angular velocity components of the current motion state is calculated using the five-point numerical differentiation method. The absolute value of the instantaneous rate of change is taken and combined with the component weights to calculate the intensity of dynamic change, so as to supplement the additional influence of the speed of motion change on the error. For example, rapid acceleration will cause more significant instantaneous deformation of the sensor than gradual acceleration, which needs to be quantified separately.

[0108] The instantaneous motion intensity is obtained by weighted summation of the basic intensity and the dynamic intensity, with the weight of the basic intensity being higher than that of the dynamic intensity. This reflects the objective law that motion amplitude is the main influencing factor of error, while motion change rate is a secondary influencing factor. By stratifying and quantifying the dual-dimensional influence of static motion amplitude and dynamic abrupt changes, the correlation between motion intensity index and actual error is ensured to be closer.

[0109] Based on motion state labels, corresponding sub-models are called from the pre-trained dynamic coupling error model. For example, positive acceleration of X corresponds to the acceleration error sub-model. This avoids the mapping deviation caused by a single model adapting to all motion states. The instantaneous motion intensity is used as input to generate theoretical deviation values ​​to characterize the theoretical deviation caused by motion. The dynamic coupling error model is related to the motion state in the classification criteria and contains multiple sub-models. The number of sub-models is consistent with the types of motion states. All of them are trained through a large amount of experimental data to establish a linear mapping relationship between motion intensity and deviation value. That is, the deviation value in a certain direction is equal to the mapping coefficient in the corresponding direction multiplied by the motion intensity, plus the static deviation intercept in the corresponding direction. The mapping coefficient and static deviation intercept of different sub-models are different due to the differences in motion states.

[0110] To avoid inaccurate predictions due to model parameter aging over time, the vehicle body point cloud data is used to obtain the vehicle body calibration point cloud corresponding to the dynamic calibration target by utilizing the correspondence between the initial calibration point pairs. The actual deviation is calculated by the difference between the vehicle body calibration point cloud and the dynamic calibration target. The model deviation is calculated by the difference between the theoretical deviation value and the actual deviation. The mapping coefficients and static deviation intercept of the sub-model are adjusted by the least squares correction algorithm until the model deviation is no greater than the schedule requirement error. Finally, the coupling deviation value predicted by the fine-tuned sub-model is output.

[0111] Based on the X-axis coordinates of the vehicle body point cloud data in the vehicle dynamic coordinate system, the point cloud regions are divided and corresponding point cloud region labels are configured, such as the front of the vehicle, the body of the vehicle, and the rear of the vehicle. Different regions are defined by the range of X-axis coordinates to ensure that each region corresponds to the actual structural part of the vehicle body.

[0112] For different point cloud regions, based on the influence of sensor viewing angle differences on deviation response, different regional compensation coefficients are predefined. Through multiplication operations, the coupling deviation value is corrected to obtain the regional compensation value. For point cloud points in each point cloud region, reverse compensation is performed according to the regional compensation value. The standardized coordinates of the point cloud points in each direction are subtracted from the regional compensation value in the corresponding direction to obtain the compensated point cloud coordinates. This offsets the systematic deviation caused by motion coupling in that region. All regionally compensated point cloud points are integrated to generate a compensated vehicle body point cloud. The regional compensation coefficient can be set by those skilled in the art or determined through sensor sensitivity experiments. For example, the front area of ​​the vehicle is collected by the front sensor and is more sensitive to motion deviation response, so the regional compensation coefficient is 1.1. The rear area of ​​the vehicle is collected by the side sensor and the deviation response is relatively weak, so the regional compensation coefficient is 0.9. The body area is located in the middle of the sensor's viewing angle, so the regional compensation coefficient is 1.

[0113] Feature points corresponding to the dynamic calibration target are extracted from the compensated vehicle body point cloud. The compensation error of each feature point is calculated and compared with the compensation error threshold. If the compensation error in all directions is not greater than the compensation error threshold, the compensation is deemed qualified and the compensated vehicle body point cloud is directly output.

[0114] If the compensation error in any direction exceeds the compensation error threshold, a secondary correction process is initiated. The secondary correction value is calculated by multiplying the compensation error in the direction exceeding the threshold by the error convergence coefficient. Based on the difference between the compensated vehicle point cloud and the secondary correction value, a new compensated vehicle point cloud is generated. This process continues until the compensation error in all directions meets the threshold requirements. Finally, a qualified compensated vehicle point cloud is output, and key information of the entire compensation process, such as motion state label, total motion intensity, deviation value, and compensation error, is recorded to generate a compensation process log. The error convergence coefficient is the optimal convergence coefficient verified through multiple experiments. This avoids overcorrection in the first correction, which could lead to a reverse error, and ensures that the error can be reduced to within the threshold after the secondary correction.

[0115] Specifically, the steps for identifying transient disturbance events include:

[0116] For acceleration and angular velocity signals in inertial motion data, a wavelet transform denoising algorithm is used to perform multi-level decomposition of the signal, splitting the original signal into low-frequency approximate components and high-frequency detail components to avoid misjudgment caused by interference identification. A soft threshold is set, and the high-frequency detail components are quantized to distinguish between effective interference signals and residual low-frequency noise in the high-frequency detail components, suppressing noise while retaining pure interference features. The high-frequency detail components after thresholding are subjected to inverse wavelet transform to reconstruct the instantaneous interference feature signal, which contains only the high-frequency components of instantaneous interference for subsequent interference identification, while retaining the low-frequency approximate components for subsequent motion correlation verification of the filtering effect. The soft threshold is the product of the signal standard deviation and the natural logarithm of the square root of twice the number of sampling points.

[0117] Traverse the instantaneous interference feature signals, extract the acceleration interference amplitude and angular velocity interference amplitude at each moment, and use dual thresholds to initially screen suspected interference events; if the acceleration interference amplitude is not less than the acceleration interference amplitude threshold or the angular velocity interference amplitude is not less than the angular velocity interference amplitude threshold, and the duration is within the preset duration threshold range to cover common interference durations, mark it as a suspected interference event to distinguish interference from normal fluctuations;

[0118] For suspected interference events, the interference energy is calculated to reflect the intensity and duration of the interference, as shown in the following expression:

[0119]

[0120] In the formula, To interfere with energy, , These are the start and end times of the interference, respectively. For the duration of the interference, for The amplitude of acceleration disturbance at any moment, for The amplitude of the angular velocity disturbance at any given time. Angular velocity energy weight;

[0121] The signal waveforms of suspected interference events are analyzed to determine the type of interference. By combining the interference energy and the interference time period, interference event markers are generated and associated with the point cloud of the compensated vehicle body at the same time, so as to achieve accurate quantification of interference type and intensity and avoid the indiscriminate filtering in the subsequent process. Among them, if the waveform shows multiple reciprocating fluctuations, such as the up and down vibration caused by road bumps, it is judged as vibration interference; if the waveform is a single peak followed by rapid decay, such as the instantaneous impact caused by equipment collision, it is judged as impact interference.

[0122] Different regions of the vehicle body have different contour characteristics, requiring targeted adjustments to the filtering strategy. The corresponding filtering algorithm is called based on the type of interference. For vibration-related interference, an improved Gaussian filter is used to balance smoothness and detail preservation. For impact-related interference, an adaptive median filter is used to specifically eliminate isolated points. Filtering parameters are calculated through parameter matching for the filtering operation; the expression is shown below:

[0123]

[0124]

[0125] In the formula, The standard deviation of the filter for vibration-type interference. Based on the standard deviation, Energy coefficient This is the regional compensation coefficient. The size of the filtering window for impulse interference. Basic window, For the rounding function, ensure that the number of windows is odd;

[0126] By combining point cloud region labels, the compensation vehicle body point cloud during the interference period is divided into corresponding regions. At the same time, based on the point cloud curvature threshold, detail sub-regions and planar sub-regions are marked in each region. If the region curvature is greater than the point cloud curvature threshold, it is determined to be a detail sub-region; if the region curvature is not greater than the point cloud curvature threshold, it is determined to be a planar sub-region.

[0127] Differentiated filtering is performed for different types of interference. For point clouds associated with vibration-type interference, an improved Gaussian filter is enabled. Planar sub-regions are filtered according to the calculated standard deviation of the filter, while detailed sub-regions are filtered by multiplying the standard deviation of the filter by the detail protection coefficient to reduce the filtering intensity. At the same time, detail edges are preserved through an edge protection factor. The edge protection factor is calculated based on the difference in the angle between the point cloud normal vectors. When the angle exceeds the angle threshold, it is identified as an edge point and the filtering weight is reduced. The detail protection coefficient is used to reduce the filtering intensity and its value is less than 1.

[0128] For point clouds associated with impact-type interference, adaptive median filtering is enabled. Each point cloud point is traversed. If the distance between a point cloud point and its neighboring points is not less than a multiple of the neighborhood standard deviation, it is determined to be an isolated noise point. The coordinates of the corresponding point cloud point are replaced with the median value within the filtering window. If it is a detail sub-region, the original filtering window is replaced with the basic window to avoid covering too many detail points. At the same time, a neighboring point count check is added. If there are too few neighboring points, no filtering is performed to avoid detail points being misjudged as noise points.

[0129] For point clouds without associated interference events, only one level of smoothing filter is performed to avoid contour distortion in noise-free areas, and finally a denoised vehicle body point cloud is obtained.

[0130] Feature points of the dynamically calibrated target are extracted from the denoised vehicle point cloud. The ratio of the number of retained feature points to the original number is calculated to obtain the denoising integrity rate. Once the denoising integrity rate is lower than the preset denoising integrity rate threshold, it is determined that details are lost. Then, the process returns to the adjustment of the filtering strategy, the filtering intensity of the corresponding area is increased by a preset ratio, and the filtering is re-executed.

[0131] Extract key contour lines of the vehicle body, such as the edge of the roof and the side lines of the body. Calculate the curvature change of adjacent points on the contour lines using the three-point curvature calculation method. If the curvature change exceeds the curvature change threshold, it is determined to be noise residue. Then return to the adjustment filtering strategy, increase the filtering intensity of the corresponding area by a preset ratio, and re-execute the filtering.

[0132] Extract the coordinate change trend of key feature points of the filtered vehicle body point cloud and compare it with the retained low-frequency approximate components. If the change trend is inconsistent, it is determined to be a filtering anomaly. Then, the interference type and interference period are re-identified to ensure the accuracy of the filtering period.

[0133] Specifically, the steps for weighted point cloud registration and pose calculation include:

[0134] Because sensors at different viewpoints are installed in different locations (e.g., the front sensor faces the direction of movement directly, while the rear sensor faces the direction of movement from the side), they are significantly affected by the intensity of motion and the energy of interference. Therefore, the instantaneous intensity of motion and the energy of interference at different viewpoints are normalized to obtain normalized intensity and energy of motion and interference, eliminating the influence of dimensional differences on the weights. Based on the weight ratios of intensity of motion and energy of interference, the basic weight for each viewpoint is calculated. The expression is as follows:

[0135]

[0136] In the formula, , These represent the weighting percentages of motion intensity and disturbance energy, respectively. To normalize the intensity of motion, To normalize the interference energy;

[0137] The basic weights are corrected by combining the regional compensation coefficients corresponding to the viewpoints. The confidence weights of the corresponding viewpoints are obtained through multiplication. The confidence weights of different viewpoints are normalized to avoid registration errors caused by the total weights exceeding the range, so as to generate a set of confidence weights for each viewpoint.

[0138] The view with the highest confidence weight is selected as the reference view, the denoised vehicle point cloud corresponding to the reference view is used as the baseline point cloud, and the denoised vehicle point clouds of other views are used as the point clouds to be registered.

[0139] Using the weighted iterative nearest point algorithm, weighted point pair matching is performed between each point cloud to be registered and the reference point cloud. For each point in the point cloud to be registered, the point with the closest Euclidean distance in the reference point cloud is searched to form an initial weighted point pair. A confidence weight based on the viewpoint of the point to be registered in each pair is assigned, and a confidence weight based on the reference viewpoint is assigned to the reference point. The total weight of the point pair is the average of the two confidence weights, and a weighted registration error function is constructed. The weighted registration error function is the sum of the total weight of all point pairs multiplied by the squares of the coordinate differences between the two points. This avoids the limitations of equal-weight calculation errors, allowing high-weight point pairs to dominate the error calculation and reducing the interference of low-weight point pairs. The expression is shown below:

[0140]

[0141] In the formula, For the first Registration error at each viewpoint , For the number of viewpoints, For the first The total weight of each pair of points For point-to-point quantity, For the first The reference point in a pair of points For the first The first point in the pair The points to be registered from each perspective Let be a rotation matrix. It is a translation vector;

[0142] The rotation matrix and translation vector are solved by singular value decomposition algorithm, the weighted registration error function is minimized, and iterative calculation is performed until the registration error is less than the preset convergence threshold or the number of iterations reaches the upper limit. The registration transformation matrix of the point cloud to be registered relative to the reference point cloud is obtained. The same process is used to complete the registration of all point clouds to be registered with the reference point cloud. All view point clouds are mapped to the reference view coordinate system, i.e. the vehicle dynamic coordinate system, and integrated into a unified multi-view fused point cloud. This ensures that the fused point cloud has no view misalignment and no noise interference, and provides a high-precision unified point cloud foundation for subsequent pose calculation.

[0143] Vehicle body feature points are extracted from multi-view fused point clouds using predefined geometric feature points, such as wheel centers and roof corners. The curvature of each vehicle body feature point is calculated, and high curvature points exceeding a preset curvature threshold are selected. Clustering is performed based on the neighborhood search radius, and each cluster center is labeled as a candidate feature point. Combining the relative positional relationship of the candidate feature points on the vehicle body, such as the symmetrical distribution of wheel centers, the actual feature points are matched and their coordinates in the vehicle body dynamic coordinate system are recorded.

[0144] The average coordinates of all wheel centers are used as the vehicle body position coordinates to reflect the overall spatial position of the vehicle body. The attitude angles, including roll angle, pitch angle, and yaw angle, are calculated by combining the coordinate differences between the wheel centers and the corner points of the roof with trigonometric functions. Among them, the roll angle is calculated based on the vertical coordinate difference between the left and right wheel centers, reflecting the left and right tilt of the vehicle body; the pitch angle is calculated based on the vertical coordinate difference between the front and rear wheel centers, reflecting the front and rear tilt of the vehicle body; and the yaw angle is calculated based on the horizontal lateral coordinate difference between the front and rear wheel centers, reflecting the left and right yaw of the vehicle body.

[0145] Extract the dynamic calibration target feature points from the fused point cloud, map them to the global coordinate system through the solved pose parameters, and calculate the pose accuracy error between the mapped coordinates and the dynamic calibration target.

[0146] If the pose accuracy error is less than the preset accuracy threshold, the pose is deemed qualified, and the initial pose data of the target vehicle, including position coordinates and attitude angles, is output; otherwise, the vehicle body feature points are re-extracted and the calculation is repeated to ensure that the pose accuracy meets the requirements.

[0147] Specifically, the steps for calculating the pose estimation residuals include:

[0148] Perform a double integral on the acceleration data in the inertial motion data to obtain the inertial predicted position increment, and perform a single integral on the angular velocity data to obtain the inertial predicted attitude angle increment.

[0149] Initial pose data is used as the initial state vector. An extended Kalman filter is used to perform prediction. The inertial increment is converted into pose change through a rigid body kinematics model, resulting in the state vector of the inertial predicted pose. Simultaneously, the state covariance matrix is ​​calculated to quantify the uncertainty of the predicted pose. Finally, the inertial predicted pose, including the predicted position and predicted attitude angle, is output as a dynamic reference for calculating the pose residual. The inertial increment includes the inertial predicted position increment and the inertial predicted attitude angle increment. The rigid body kinematics model includes an input layer, a state transition layer, a covariance prediction layer, and an output layer. The input layer receives the inertial predicted position increment, the inertial predicted attitude angle increment, and the initial state vector. The inertial predicted position increment is the displacement obtained by quadratic integration of the acceleration data. The change quantities are as follows: the inertial predicted attitude angle increment is the attitude angle change obtained by integrating the angular velocity data once, corresponding to roll, pitch, and yaw angles respectively; the initial state vector is the pose calculation result of the previous moment; the state transition layer is used for position and attitude updates; the position update superimposes the inertial predicted position increment onto the position of the previous moment to obtain the predicted position; the attitude update uses quaternions or rotation matrices to convert the inertial predicted attitude angle increment into an attitude transformation operator, which is then combined with the attitude of the previous moment to obtain the predicted attitude; the covariance prediction layer is used to update the state covariance matrix based on the system noise covariance matrix and the state transition matrix to quantify the uncertainty of the predicted pose; and the output layer is used to output the theoretically calculated pose and the state covariance matrix.

[0150] The initial pose data is compared with the inertial predicted pose, and the single-dimensional residuals are calculated, including the position residual and the attitude residual. The position residual is calculated by the Euclidean distance between the position coordinates of two points, reflecting the overall deviation of the position estimation. The attitude angle residual is calculated by the Euclidean distance between the attitude angles of two points, reflecting the overall deviation of the attitude angle estimation.

[0151] Configure position residual weights and attitude residual weights, calculate the comprehensive residual of pose estimation, and set the pose residual threshold; if the comprehensive residual is not greater than the pose residual threshold, it is determined that the current pose deviation is small, and there is no need to update the parameters significantly, and a pose sequence is generated.

[0152] If the overall residual exceeds the pose residual threshold, indicating an error accumulation trend, parameter updates are performed to locate the source of the deviation. Based on this source, corresponding parameter updates are conducted, and the pose is recalculated and the residuals verified using the updated parameters. The source of the deviation includes: if the position residual primarily originates from a certain direction, parameter drift in the corresponding direction in the dynamic coupling error model is identified, and the recursive least squares algorithm is used to update the dynamic coupling error model parameters. The residual in the corresponding direction is used as the observation input to calculate the correction amount of the model parameters, resulting in the updated model parameters. If the pose residual primarily originates from a certain angle, the corresponding rotation component offset in the sensor extrinsic parameter matrix is ​​identified, and a Kalman filter is used to update the sensor extrinsic parameter matrix. The residual in the corresponding angle is used as the observation input to calculate the correction amount of the extrinsic parameter matrix, and the correction intensity is adjusted based on the Kalman gain to avoid over-updating, resulting in the updated extrinsic parameter matrix.

[0153] Specifically, the steps for reconstructing the deviation analysis include:

[0154] For each time stamp of the denoised vehicle point cloud, a rigid body transformation mapping is performed in combination with the corresponding pose sequence to transform the denoised vehicle point cloud from the vehicle dynamic coordinate system to the global coordinate system of the detection area. The pose confidence of each time step is calculated by the ratio of the comprehensive residual at the corresponding time step to the maximum allowable residual. The point clouds at multiple time steps in the global coordinate system are fused by pose confidence and contour completion is performed based on Poisson reconstruction to fill the point cloud gaps in the occluded areas of the vehicle body and generate a complete 3D contour model of the vehicle body.

[0155] A standard model of the target vehicle is obtained. In the global coordinate system, global ICP registration is performed in conjunction with the 3D contour model of the vehicle body to minimize the positional deviation between the standard model and the 3D contour model of the vehicle body in the global coordinate system. By iteratively optimizing the rotation matrix and translation vector, the Euclidean distance between the key feature points in the 3D contour model of the vehicle body and the corresponding points in the standard model is less than a preset distance threshold, ensuring that the spatial references of the two are completely consistent. The standard model is a 3D CAD model of the vehicle under ideal conditions provided by the vehicle design department.

[0156] For the registered 3D contour model of the vehicle body and the standard model, the registration deviation vector is calculated point by point based on the correspondence of each model point. The magnitude of the deviation vector is calculated using the Euclidean distance formula, thereby obtaining the absolute deviation value of each point, quantifying the difference between the actual contour and the design standard, and color coding of the point cloud in the 3D contour model of the vehicle body according to the magnitude of the absolute deviation value to generate a color map of the vehicle body size deviation.

[0157] Points whose absolute deviation values ​​exceed a preset defect judgment threshold are filtered out. Based on Euclidean distance clustering, adjacent points are divided into independent deviation regions. Each deviation region is marked with a unique number and its minimum bounding box coordinates are recorded to locate the defect position. The average deviation of each deviation region is calculated. The average deviation is the ratio of the sum of the absolute deviation values ​​of all points in the corresponding region to the number of points in the region.

[0158] To eliminate random measurement fluctuations and distinguish whether deviation areas are genuine defects, curvature features, normal vector change rate, area, and average deviation are extracted for each deviation area to generate deviation geometric features. Combined with the constructed defect classification model, the defect type of the current target vehicle is output. Simultaneously, based on the average deviation and area, the defect severity is calculated with weights and divided into three levels: mild, moderate, and severe, according to the defect severity value. Targeted rectification suggestions are given based on the defect type. In particular, a defect classification model is constructed based on a point cloud deep learning network. It is trained with massive labeled defect samples to learn the mapping relationship between geometric features and defect types. Real-time geometric features are input to output the defect type of the current target vehicle, such as dents, bulges, distortions, and edge chipping.

[0159] The time series of roll angle, pitch angle and yaw angle are extracted from the pose sequence to form three sets of one-dimensional data based on time order. For each set of attitude angle time series, the fluctuation amplitude, fluctuation frequency, standard deviation and peak factor are calculated and normalized. The total stability score is calculated by weighting multiple indicators and the stability level is divided. Finally, the stability assessment result is generated.

[0160] Working principle and effects:

[0161] By calibrating sensor timestamps through a pulse synchronization mechanism integrated with a sensor network and combining this with a pre-calibrated extrinsic parameter matrix, point cloud and inertial motion data are unified into the vehicle's dynamic coordinate system. This solves the sampling misalignment problem, lays a data foundation free of basic bias, and ensures the accuracy of subsequent processing. The inertial motion data is then analyzed to calculate motion intensity. A pre-trained dynamic coupling error model is used to reverse-correct the point cloud, linking vehicle motion with sensor deviations to eliminate systematic errors caused by motion and compensate for the lack of correlation in existing technologies. A sliding window is used to monitor high-frequency components of inertial motion data to identify instantaneous interference. Adaptive filtering based on interference energy preserves contour details while denoising, mitigating the impact of sudden interference and improving point cloud quality. Confidentiality weights are assigned to multi-view data, and weighted registration is used to calculate the initial pose, allowing less affected data to dominate and improving pose accuracy. Kalman filtering is used to update the model and extrinsic parameter matrix with pose residuals, preventing error accumulation. Global contours are reconstructed based on accurate poses, and deviations are analyzed, defects are identified, and stability is evaluated by comparing with a standard model. This eliminates coupling and interference effects, significantly improving detection accuracy and meeting the requirements for high-precision moving vehicle detection.

[0162] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for three-dimensional contour detection and deviation analysis of a target vehicle, characterized in that, include: Initial point cloud data and motion data are collected synchronously, mapped to the constructed vehicle dynamic coordinate system, and data alignment is verified to generate vehicle point cloud data and inertial motion data. The instantaneous motion intensity index is calculated, and the theoretical deviation value of the vehicle body point cloud data at the current moment is generated using the dynamic coupling error model, and the reverse deviation correction is performed. Identify instantaneous interference events, record the interference period, calculate the interference energy, and perform adaptive statistical filtering on the vehicle body point cloud data during the interference period. Assign real-time confidence weights and perform weighted point cloud registration and pose calculation to generate initial pose data; The inertial predicted pose is obtained, the pose estimation residual is calculated, and the parameters and extrinsic parameter matrix of the dynamic coupling error model are updated and fed back using a Kalman filter to generate a pose sequence. Based on the pose sequence, a point cloud fusion algorithm is used to reconstruct the three-dimensional contour, generate a three-dimensional contour model of the vehicle body, and perform reconstruction deviation analysis to generate a comprehensive vehicle body inspection report.

2. The method for three-dimensional contour detection and deviation analysis of a target vehicle according to claim 1, characterized in that, The steps for data alignment verification include: The initial point cloud data is segmented using feature points to obtain the initial centroid; Obtain the theoretical centroid of the target vehicle, perform dynamic centroid correction on the initial centroid, generate the centroid correction amount, and calculate the actual centroid coordinates; A dynamic coordinate system for the vehicle body is established with the actual centroid coordinates as the origin; Acquire a dynamic calibration target and perform feature point matching with the calibration target point cloud data in the initial point cloud data to generate an initial calibration point pair; By using the sensor extrinsic parameter matrix, the calibration target point cloud data is mapped to the vehicle body dynamic coordinate system to obtain the mapped point cloud coordinates; Based on the initial calibration point pair, the corrected extrinsic parameter matrix is ​​generated using the extrinsic parameter optimization model; The initial point cloud data is mapped to the vehicle dynamic coordinate system using the corrected extrinsic parameter matrix to obtain vehicle point cloud data. The positional deviation between the motion data and the actual centroid coordinates is corrected to generate inertial motion data.

3. The method for three-dimensional contour detection and deviation analysis of a target vehicle according to claim 2, characterized in that, The data alignment verification steps also include: Extract vehicle body feature points, obtain the coordinates of vehicle body feature points in adjacent frames, and calculate the difference in coordinate changes; Based on the motion direction of the inertial motion data, the spatial correlation error is calculated; if the spatial correlation error is not greater than the spatial error threshold, the spatial verification is deemed to have passed; otherwise, a data resampling command is triggered. Obtain the timestamp interval. If the timestamp interval is not greater than the verification multiple of the sampling period, the time verification is deemed to have passed; otherwise, a data resampling instruction is triggered. The number of point clouds in a single frame of vehicle body point cloud data and the number of complete parameters in a single set of inertial motion data are counted. The number of missing parameters is calculated, the theoretical number of parameters is obtained, and the data missing rate is calculated. If the data missing rate is not greater than the data missing threshold, the data verification is deemed successful; otherwise, a data re-sampling instruction is triggered.

4. The method for three-dimensional contour detection and deviation analysis of a target vehicle according to claim 3, characterized in that, The steps for correcting reverse deviation include: Extract the acceleration and angular velocity components at the current sampling moment, remove the static components, compare them one by one with the classification criteria of motion state, and record the current motion state label. Obtain the acceleration and angular velocity components corresponding to the current motion state, and call the corresponding component weights to calculate the basic intensity of the current motion state; Calculate the intensity of dynamic change based on the instantaneous rate of change of the current motion state; The instantaneous motion intensity is obtained by weighted summation of the base intensity and the dynamically changing intensity. Based on the motion state labels, the corresponding sub-model is called from the pre-trained dynamic coupling error model to generate theoretical deviation values; In the vehicle body point cloud data, the correspondence in the initial calibration point pair is used to obtain the vehicle body calibration point cloud corresponding to the prime number dynamic calibration target, and the actual deviation is calculated. The model deviation is calculated by the difference between the theoretical deviation value and the actual deviation value, and the sub-model is adjusted to generate the coupling deviation value predicted by the fine-tuned sub-model.

5. The method for three-dimensional contour detection and deviation analysis of a target vehicle according to claim 4, characterized in that, The steps for correcting reverse bias also include: The point cloud region is divided and a region compensation coefficient is set to correct the coupling deviation value and obtain the region compensation value. For the point cloud points in each point cloud region, the region compensation value is used to perform reverse compensation to generate a compensated vehicle body point cloud. Feature points corresponding to the dynamic calibration target are extracted from the point cloud of the compensated vehicle body, and the compensation error is calculated. If the compensation error in all directions is not greater than the compensation error threshold, the compensation is deemed qualified. If the compensation error in any direction exceeds the compensation error threshold, a secondary correction process is initiated to calculate the secondary correction value and generate a new compensated vehicle point cloud, until the compensation error in all directions meets the threshold requirement.

6. The method for three-dimensional contour detection and deviation analysis of a target vehicle according to claim 5, characterized in that, The steps for identifying transient disturbance events include: By using wavelet transform for denoising, the inertial motion data is split into low-frequency approximate components and high-frequency detail components. Set a soft threshold to quantize high-frequency detail components and generate transient interference characteristic signals; Traverse the instantaneous interference feature signals, extract the acceleration interference amplitude and angular velocity interference amplitude at each moment, and use dual thresholds to initially screen suspected interference events; For suspected interference events, calculate the interference energy, analyze the signal waveform of the suspected interference event to determine the type of interference, and generate interference event markers; The corresponding filtering algorithm is invoked based on the type of interference, and the filtering parameters are calculated through parameter matching. By combining point cloud region labels, the compensation vehicle body point cloud during the interference period is divided into corresponding regions. At the same time, based on the point cloud curvature threshold, detail sub-regions and planar sub-regions are marked in each region. Differentiated filtering is performed for different types of interference. For point clouds that are not associated with interference events, only a first-level smoothing filter is performed to obtain a denoised vehicle body point cloud.

7. The method for three-dimensional contour detection and deviation analysis of a target vehicle according to claim 6, characterized in that, The steps of weighted point cloud registration and pose calculation include: The instantaneous motion intensity and interference energy from different perspectives are normalized to obtain normalized motion intensity and normalized interference energy, and the basic weight of each perspective is calculated. The basic weights are corrected by combining the regional compensation coefficients corresponding to the viewpoints to obtain confidence weights, and then normalized to generate a set of confidence weights for each viewpoint. Divide the reference point cloud and the point cloud to be registered, and use the weighted iterative nearest point algorithm to generate a fused point cloud; Extract vehicle body feature points and calculate curvature, filter candidate feature points, and combine the relative positional relationship of candidate feature points on the vehicle body to match and obtain actual feature points; Obtain the vehicle's position coordinates, and calculate the attitude angles by using the coordinate differences between the wheel center and the corner of the roof. Extract the dynamic calibration target feature points from the fused point cloud, map them to the global coordinate system through the solved pose parameters, and calculate the pose accuracy error; Once the pose accuracy error is not less than the preset accuracy threshold, the vehicle body feature points are re-extracted and the calculation is repeated.

8. The method for three-dimensional contour detection and deviation analysis of a target vehicle according to claim 7, characterized in that, The steps for calculating pose estimation residuals include: Perform a double integral on the acceleration data in the inertial motion data to obtain the inertial predicted position increment, and perform a single integral on the angular velocity data to obtain the inertial predicted attitude angle increment. Using the initial pose data as the initial state vector, the extended Kalman filter is enabled to perform prediction, generating an inertial predicted pose, including the predicted position and the predicted attitude angle. The initial pose data is compared with the inertial predicted pose, and the single-dimensional residuals are calculated, including position residuals and attitude residuals. Configure position residual weights and attitude residual weights, and calculate the comprehensive residual of pose estimation; Once the overall residual exceeds the pose residual threshold, it is determined that there is an error accumulation trend. Then, the parameters are updated, the source of deviation is located, and the corresponding parameters are updated based on the source of deviation. The pose is then recalculated based on the updated parameters, and the residual is verified.

9. The method for three-dimensional contour detection and deviation analysis of a target vehicle according to claim 8, characterized in that, The steps of reconstructing deviation analysis include: For each timestamp of the denoised vehicle point cloud, combined with the corresponding pose sequence, a rigid body transformation mapping is performed to map it to the global coordinate system of the detection area, and the pose confidence at each time point is calculated. The multi-moment point cloud in the global coordinate system is fused by pose confidence and contour completion is performed based on Poisson reconstruction to generate a three-dimensional contour model of the vehicle body. Obtain the standard model of the target vehicle, and perform global ICP registration in the global coordinate system in conjunction with the three-dimensional contour model of the vehicle body; For the registered 3D contour model of the vehicle body and the standard model, the registration deviation vector is calculated point by point, and the absolute deviation value of each point is obtained by using the Euclidean distance formula; Points with absolute deviation values ​​exceeding a preset defect judgment threshold are selected, and deviation regions are divided based on Euclidean distance clustering. The average deviation of each deviation region is then calculated.

10. The method for three-dimensional contour detection and deviation analysis of a target vehicle according to claim 9, characterized in that, The steps of reconstructing deviation analysis also include: For each deviation region, extract curvature features, rate of change of normal vector, region area, and average deviation to generate the deviation geometric features of the deviation region; Based on the constructed defect classification model, the defect type of the current target vehicle is output, and the defect severity is calculated by weighting based on the average deviation and the area of ​​the region. Extract the time series of roll angle, pitch angle and yaw angle from the pose sequence respectively. For each set of attitude angle time series, calculate the fluctuation amplitude, fluctuation frequency, standard deviation and peak factor. The stationarity score is calculated by weighting multiple indicators, and the stationarity level is classified.

Images