A real-time terrain anomaly detection method based on multi-source data analysis

Abstract

The present application relates to the technical field of vehicle environment perception analysis, and particularly relates to a real-time terrain anomaly detection method based on multi-source data analysis. The method comprises the following steps: collecting laser point cloud terrain monitoring data, terrain image feature data and vehicle motion pose characteristic data; performing front-vehicle perception terrain map analysis, i.e. optimization processing, based on the laser point cloud terrain monitoring data and the vehicle motion pose characteristic data, by using a multi-source error transfer mechanism, to generate optimized front-vehicle perception terrain map data; transmitting the terrain image feature data to the optimized front-vehicle perception terrain map data for visual feature matching and mapping processing, to generate optimized front-vehicle perception visual terrain map data; and performing real-time terrain anomaly detection processing of front-vehicle perception based on the optimized front-vehicle perception visual terrain map data, to generate front-vehicle perception real-time terrain anomaly detection data. The present application realizes real-time terrain anomaly detection of multi-source errors under rainy conditions.

Description

Technical Field

[0001] This invention relates to the field of vehicle environmental perception and analysis technology, and in particular to a real-time terrain anomaly detection method based on multi-source data analysis. Background Technology

[0002] In the field of emergency rescue, land rescue, as a core means of responding to sudden events such as natural disasters and accidents, is highly dependent on the mobility of emergency rescue vehicles. Emergency rescue vehicles often face severe challenges in unstructured rainy scenarios. In such scenarios, roads are muddy, flooded, and littered with gravel, and the terrain features are complex and dynamically changing, directly affecting vehicle ride comfort, maneuverability, and driving safety. Traditional terrain detection methods are insufficient for the needs of unstructured rainy scenarios. While direct measurement methods offer high accuracy, they are expensive and difficult to implement real-time pre-aiming. Inverse recognition methods based on vehicle dynamics response, while practical, cannot obtain accurate temporal information, making it difficult to support the vehicle's ability to predict the terrain ahead. With the development of sensor technology, multi-sensor fusion methods are gradually becoming the mainstream direction for terrain perception. However, existing real-time terrain anomaly detection methods still have significant bottlenecks in unstructured rainy scenarios: they cannot accurately analyze the errors caused by monitoring information in rainy scenarios, resulting in poor terrain detection accuracy. Furthermore, the kinematic errors of the vehicle itself in rainy scenarios and the lack of analysis of the differences in mechanical properties of different terrains make it difficult to identify hidden dangers where the elevation is normal but the vehicle's adhesion coefficient is abnormal. Summary of the Invention

[0003] Based on this, the present invention provides a real-time terrain anomaly detection method based on multi-source data analysis to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a real-time terrain anomaly detection method based on multi-source data analysis includes the following steps:

[0005] Step S1: Use the vehicle's built-in multi-source monitoring sensors to collect multi-source terrain monitoring data of the rain scene monitoring area, which includes laser point cloud terrain monitoring data, terrain image feature data and vehicle motion pose characteristic data.

[0006] Step S2: Based on laser point cloud terrain monitoring data and vehicle motion pose characteristic data, perform multi-source error propagation mechanism analysis and processing of the vehicle front perception terrain map to generate vehicle front perception terrain map data.

[0007] Step S3: Optimize the vehicle front perception terrain map data by performing vehicle positioning cumulative error processing on the vehicle front perception terrain map data to generate optimized vehicle front perception terrain map data;

[0008] Step S4: Transmit the terrain image feature data to the optimized front-view terrain map data for visual feature matching and mapping processing to generate optimized front-view visual terrain map data; perform real-time terrain anomaly detection processing based on the optimized front-view visual terrain map data to generate real-time terrain anomaly detection data; and feed the real-time terrain anomaly detection data back to the terminal to perform real-time terrain anomaly detection feedback operation under rainy conditions.

[0009] Furthermore, step S1 includes the following steps:

[0010] The vehicle's built-in multi-source monitoring sensors include lidar equipment, monitoring camera equipment, and inertial measurement unit equipment;

[0011] The vehicle's built-in lidar equipment is used to perform lidar point cloud terrain monitoring in the rainy weather monitoring area, generating lidar point cloud terrain monitoring data.

[0012] The vehicle's built-in monitoring camera is used to monitor and process the terrain image features of the monitoring area in rainy weather, generating terrain image feature data.

[0013] The vehicle's motion posture characteristics are collected using the inertial measurement unit (IMU) built into the vehicle, generating vehicle motion posture characteristic data.

[0014] Furthermore, step S2 includes the following steps:

[0015] Step S21: Based on the vehicle motion posture characteristic data, the vehicle motion acceleration, vehicle motion angular velocity and vehicle motion speed are used to analyze the influence characteristics of point cloud motion distortion under rainy conditions, and generate point cloud motion distortion influence characteristic data under rainy conditions.

[0016] Step S22: Perform covariance matrix analysis on the laser point cloud error characterization based on the laser point cloud terrain monitoring data to generate the covariance matrix of the laser point cloud error characterization under rainy weather conditions.

[0017] Step S23: Based on the feature data of motion distortion of point cloud under rainy conditions and the covariance matrix of laser point cloud error characterization under rainy conditions, establish a point cloud calibration model for rainy conditions. Then, use the point cloud calibration model for rainy conditions to filter and remove distortion from the original point cloud topography of laser point cloud topography monitoring data under rainy conditions to obtain laser point cloud calibration topography data.

[0018] Step S24: Based on the laser point cloud calibration terrain data, perform point cloud terrain coordinate transformation and solution processing to generate point cloud terrain coordinate data;

[0019] Step S25: Perform elevation and horizontal dimensionality reduction processing on the point cloud terrain coordinate data to generate point cloud terrain elevation dimensionality reduction data and point cloud terrain horizontal dimensionality reduction data.

[0020] Step S26: Perform vehicle motion pose error analysis based on vehicle motion pose characteristic data to generate vehicle motion pose error data; perform horizontal and vertical error analysis on point cloud terrain elevation dimensionality reduction data and point cloud terrain horizontal dimensionality reduction data based on vehicle motion pose error data to generate pose error point cloud terrain elevation dimensionality reduction data and pose error point cloud terrain horizontal dimensionality reduction data respectively.

[0021] Step S27: Perform multi-source error propagation mechanism-based front-view topographic map fusion processing on the pose error point cloud terrain elevation dimensionality reduction data and the pose error point cloud terrain horizontal dimensionality reduction data to generate front-view topographic map data.

[0022] Furthermore, step S22 includes the following steps:

[0023] Based on the analysis of laser point cloud terrain monitoring data, the laser beam uncertainty parameters and terrain matching uncertainty parameters under rainy conditions are analyzed.

[0024] Based on the preset lidar beam model and the likelihood function, the covariance matrix analysis of the laser point cloud error characterization is performed on the laser beam uncertainty parameters and terrain matching uncertainty parameters under rainy conditions, generating the laser point cloud error characterization covariance matrix under rainy conditions.

[0025] Furthermore, the lidar beam model is used to perform characteristic analysis of the lidar-emitted laser beam during propagation and reflection under rainy conditions, based on the lidar beam uncertainty parameters. The likelihood function is used to perform analysis of the degree of matching between lidar measurements and actual terrain under rainy conditions, based on the terrain matching uncertainty parameters.

[0026] Furthermore, step S24 includes the following steps:

[0027] The true terms and Gaussian error terms of the laser point cloud calibration terrain data are analyzed using a pre-set Gaussian model. Then, the linear combination analysis of the laser point cloud coordinates and laser radar measurement error terms is performed on the true terms and Gaussian error terms of the laser point cloud calibration terrain data using a small angle approximation algorithm to obtain the linear feature data of laser point cloud coordinates-error terms.

[0028] The linear feature data of the point cloud coordinates and error terms of the LiDAR are processed to transform the point cloud terrain coordinates, generating point cloud terrain coordinate transformation data. The point cloud terrain coordinates are then solved using the Jacobian matrix to obtain the point cloud terrain coordinate data.

[0029] Furthermore, step S3 includes the following steps:

[0030] Step S31: Construct the IMU pre-integration model;

[0031] Step S32: Analyze the error parameters of the vehicle motion range based on the vehicle motion posture characteristic data to obtain the vehicle motion range error parameters;

[0032] Step S33: Transmit the vehicle motion range error parameters to the IMU pre-integration model to optimize the vehicle motion error parameters and generate optimized vehicle motion error parameters;

[0033] Step S34: Perform reverse correction processing on the vehicle motion pose characteristic data based on the vehicle motion error optimization parameters to generate corrected vehicle motion pose characteristic data.

[0034] Step S35: Perform laser mileage error analysis based on laser point cloud terrain monitoring data and generate laser mileage error parameters;

[0035] Step S36: Based on the corrected vehicle motion pose characteristic data and laser mileage error parameters, perform parameter tight coupling and factor graph optimization analysis of the vehicle positioning cumulative error to generate optimized vehicle positioning cumulative error data;

[0036] Step S37: Optimize the vehicle front perception terrain map data by using the vehicle positioning cumulative error optimization data to generate optimized vehicle front perception terrain map data.

[0037] Furthermore, step S35 includes the following steps:

[0038] Keyframes are extracted from the laser point cloud terrain monitoring data to obtain keyframe laser point cloud terrain monitoring data;

[0039] The laser point cloud terrain monitoring data of key frames is analyzed to perform laser point cloud change analysis between adjacent frames, and laser point cloud change data between adjacent frames is generated.

[0040] Based on the keyframe laser point cloud terrain monitoring data and the vehicle-front perception terrain map data, the keyframe and corresponding local map positioning error analysis is performed to obtain the keyframe positioning error data.

[0041] Laser mileage error parameters are generated by analyzing the laser point cloud change data of adjacent frames and the positioning error data of key frames.

[0042] Furthermore, step S36 includes the following steps:

[0043] Based on the corrected vehicle motion pose characteristic data and laser mileage error parameters, a tight coupling analysis of the corrected vehicle motion pose and laser mileage error is performed to generate corrected pose-mileage error tight coupling data.

[0044] The corrected pose-odometer error tightly coupled data is fused and weighted using a pre-defined error state Kalman filter algorithm to obtain the fused weighted corrected pose-odometer error tightly coupled data.

[0045] Based on the fusion weighted pose-mileage error tightly coupled data, factor graph optimization analysis of vehicle positioning cumulative error is performed to generate optimized vehicle positioning cumulative error data.

[0046] Furthermore, step S4 includes the following steps:

[0047] Step S41: Transmit the terrain image feature data to the optimized front-view perception terrain map data for visual feature matching and mapping processing to generate optimized front-view perception visual terrain map data.

[0048] Step S42: Perform visual terrain type classification processing on the optimized vehicle-front perception visual terrain map data for rainy weather scenarios to generate visual terrain classification data;

[0049] Step S43: Use a preset semantic segmentation model to label the visual terrain classification data with terrain type attributes, and generate attribute visual terrain classification data;

[0050] Step S44: Based on the pre-established terrain level prior database, perform terrain visual attachment coefficient mapping processing on the attribute visual terrain classification data for different attributes to generate terrain visual attachment coefficients.

[0051] Step S45: Extract the vehicle's longitudinal kinematics data and lateral kinematics data from the vehicle's motion pose characteristic data. Based on the vehicle's longitudinal kinematics data and lateral kinematics data, perform terrain kinematics attachment coefficient mapping processing on the visual terrain classification data for various types to generate terrain kinematics attachment coefficients.

[0052] Step S46: Analyze the actual probability distribution of terrain attachment fusion based on the terrain visual attachment coefficient and terrain kinematic attachment coefficient to generate terrain attachment fusion data;

[0053] Step S47: Based on the preset self-supervised learning model, perform self-supervised optimization processing on the probability distribution uncertainty of actual terrain attachment for the terrain visual attachment coefficient and terrain kinematic attachment coefficient, and generate terrain attachment self-supervised optimization data.

[0054] Step S48: Perform real-time terrain anomaly detection processing for front-view perception based on terrain attachment self-supervised optimization data to generate real-time terrain anomaly detection data for front-view perception.

[0055] Step S49: Feed back the real-time terrain anomaly detection data from the front perception of the vehicle to the terminal to perform real-time terrain anomaly detection and feedback operations under rainy conditions.

[0056] The beneficial effects of this application lie in the fact that the present invention collaboratively collects multi-source terrain monitoring data of the monitoring area in rainy weather scenarios through multi-source monitoring sensors (LiDAR equipment, monitoring camera equipment, and inertial measurement unit equipment) built into the vehicle. The LiDAR equipment can penetrate rain and fog interference to accurately capture the three-dimensional point cloud information of the terrain, providing basic data for terrain geometric modeling; the monitoring camera equipment can record the visual texture features of the terrain, such as road surface roughness and the distribution of water accumulation areas, supplementing the lack of texture details in the point cloud data; the inertial measurement unit collects the vehicle's motion posture characteristics in real time, including parameters such as acceleration and angular velocity, providing a kinematic benchmark for the spatiotemporal alignment of multi-source data. The LiDAR point cloud data ensures the accuracy of terrain geometric features, the visual data enhances the distinguishability of terrain types, and the inertial data realizes the temporal and spatial correlation of data from different sensors. By constructing a vehicle-aware terrain map through a multi-source error propagation mechanism, the core problem of error accumulation and propagation in terrain modeling in unstructured rainy scenarios is systematically solved. This study analyzes the impact of vehicle motion pose data on the distortion of laser point clouds and establishes a calibration model based on the laser point cloud error covariance matrix. This model can specifically filter out noise and distortion caused by vehicle bumps and rain / fog scattering in rainy weather point clouds, making the laser point cloud-calibrated terrain data closer to the real terrain. The point cloud terrain coordinate data undergoes elevation and horizontal dimensionality reduction processing, enabling separate analysis of terrain features in the vertical and horizontal directions. This facilitates the quantification of error propagation patterns in different directions. Further correction of the dimensionality-reduced data using vehicle motion pose error accurately compensates for terrain modeling deviations caused by vehicle motion uncertainty, ensuring the independent accuracy and collaborative consistency of elevation and horizontal data. By fusing elevation and horizontal dimensionality-reduced data through a multi-source error propagation mechanism, the generated vehicle-front perception terrain map not only integrates the error-compensated geometric features but also reflects the spatial distribution characteristics of complex terrain in rainy weather. By constructing an IMU pre-integration model, optimizing vehicle motion pose error and laser odometry error, and combining factor map optimization, the vehicle-front perception terrain map is accurately optimized. This effectively solves the problem of interference from accumulated vehicle positioning errors on terrain modeling in unstructured rainy scenarios, significantly improving the reliability of terrain perception. The IMU pre-integration model can accurately estimate and reversely correct the cumulative drift error of the IMU by optimizing the error parameters of the vehicle's motion range, thus avoiding pose deviation caused by long-term operation. Keyframe extraction and laser mileage error analysis focus on point cloud data with significant terrain features. Through the correlation analysis between changes in adjacent frames and local map positioning errors, the mileage error of the lidar under point cloud distortion and noise interference in rainy weather is quantified.A tight coupling analysis of pose correction and laser odometry error, combined with the fusion weight allocation of the error state Kalman filter algorithm, achieves dynamic balance of multi-source errors. When point cloud quality is high, laser odometry data is prioritized; when the vehicle experiences severe vibrations, IMU pose correction is enhanced, ensuring the complementarity and consistency of error data. Factor graph optimization constrains global localization errors, effectively suppressing the divergence of accumulated errors. This ensures that the optimized front-view terrain map closely matches the real scene in terms of spatial coordinates and terrain details, laying a high-precision terrain model foundation for subsequent visual feature fusion and anomaly detection. Through the fusion of visual features and optimized terrain maps, multi-source fusion of terrain adhesion coefficients, and self-supervised learning optimization, accurate detection of terrain anomalies in unstructured rainy scenes is achieved, taking into account both the geometric and mechanical properties of terrain features, significantly improving the robustness and real-time performance of anomaly detection. The matching and mapping of terrain image features with optimized terrain maps integrates visual texture information into the geometric model, compensating for the lack of texture details in laser point clouds. This enables the optimized visual terrain map for vehicle-front perception to possess both three-dimensional geometric structure and visual features, providing rich evidence for terrain classification. The combination of semantic segmentation models and a terrain level prior database enables rapid attribute labeling and adhesion coefficient mapping for terrain types, ensuring efficient conversion of visual features into mechanical properties. The terrain kinematic adhesion coefficient calculated using vehicle longitudinal and lateral kinematic data forms a multi-source complement with the visual adhesion coefficient. Visual data provides prior attributes of the terrain, while kinematic data reflects the real-time dynamic interaction between the vehicle and the terrain. The fusion of their probability distributions effectively balances the weights of static and dynamic characteristics. The self-supervised learning model further optimizes the uncertainty of adhesion coefficients caused by changes in local features of the same terrain (such as local water accumulation on wet roads). By dynamically sampling historical data, the model's inclusiveness of terrain variations is enhanced, avoiding misjudgments caused by fluctuations in local features. Anomaly detection based on terrain-attachment self-supervised optimization data can identify both explicit hazards such as abrupt elevation changes and implicit risks such as abnormal adhesion coefficients (e.g., black ice, deep mud pits). The detection results are fed back to the terminal in real time, providing accurate decision-making basis for emergency rescue vehicle speed adjustment and suspension control, ensuring the efficient advancement of rescue missions in complex rainy scenarios.

[0057] Therefore, the real-time terrain anomaly detection method based on multi-source data analysis of this invention, by constructing a complete multi-source error analysis and compensation system, can accurately quantify and correct the errors in monitoring information under rainy conditions. By combining the lidar beam model and the likelihood model, a lidar point cloud error covariance matrix is ​​established. Combined with vehicle motion pose data to analyze the point cloud motion distortion characteristics, the constructed rainy-condition point cloud calibration model can effectively filter out point cloud noise and distortion caused by rain and fog scattering and vehicle bumps. At the same time, based on the IMU pre-integration model and the cumulative error compensation strategy optimized by the factor graph, the coupling error between vehicle motion pose and lidar mileage can be dynamically corrected, significantly reducing the terrain modeling error caused by rainy sensor measurement deviation and vehicle motion uncertainty, and solving the problem that existing technologies cannot accurately analyze rainy-day errors, resulting in poor terrain detection accuracy. By establishing a visual adhesion coefficient mapping between a semantic segmentation model and a terrain-level prior database, and generating a kinematic adhesion coefficient by combining vehicle longitudinal / lateral kinematic data, and then optimizing the probability distribution uncertainty of both through a self-supervised learning model, this approach can not only capture explicit hazards such as terrain elevation anomalies, but also identify implicit risks (such as slippery black ice, mud traps, etc.) where the elevation is normal but the vehicle adhesion coefficient is abnormal. This detection logic, which takes into account both geometric features and mechanical properties, compensates for the deficiency of existing technologies in analyzing the differences in mechanical properties of different terrains, significantly improving the comprehensiveness and reliability of terrain anomaly detection in unstructured rainy scenarios, and providing a more accurate decision-making basis for safe vehicle passage. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating the steps of a real-time terrain anomaly detection method based on multi-source data analysis according to the present invention.

[0059] Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S2.

[0060] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0061] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0062] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. Functional entities may be implemented in software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods. The term "and / or" as used herein includes any and all combinations of one or more of the associated items listed.

[0063] To achieve the above objectives, please refer to Figures 1 to 2 This invention provides a real-time terrain anomaly detection method based on multi-source data analysis. In the embodiments of this invention, please refer to... Figure 1 The diagram shown is a flowchart illustrating the steps of a real-time terrain anomaly detection method based on multi-source data analysis according to the present invention. The real-time terrain anomaly detection method based on multi-source data analysis includes the following steps:

[0064] Step S1: Use the vehicle's built-in multi-source monitoring sensors to collect multi-source terrain monitoring data of the rain scene monitoring area, which includes laser point cloud terrain monitoring data, terrain image feature data and vehicle motion pose characteristic data.

[0065] In this embodiment of the invention, the vehicle-mounted lidar scans the monitoring area of ​​unstructured roads in suburban areas during rainy weather at a set frequency. The laser beam penetrates the rain and fog to illuminate the ground, and after reflection, forms laser point cloud terrain monitoring data containing three-dimensional terrain coordinate information. A high-definition camera mounted on the front of the vehicle captures the terrain of the monitoring area, acquiring terrain image feature data containing details such as road surface texture, water accumulation areas, and gravel distribution, clearly showing the road surface's wetness and texture changes. The vehicle's built-in inertial measurement unit collects data at a set sampling rate, obtaining the vehicle's acceleration in the longitudinal, lateral, and vertical directions through a three-axis accelerometer, and the angular velocity of the vehicle around the three axes through a three-axis gyroscope, integrating these data to generate vehicle motion posture characteristic data containing the vehicle's real-time motion state.

[0066] Step S2: Based on laser point cloud terrain monitoring data and vehicle motion pose characteristic data, perform multi-source error propagation mechanism analysis and processing of the vehicle front perception terrain map to generate vehicle front perception terrain map data.

[0067] In this embodiment of the invention, vehicle motion acceleration, angular velocity, and velocity parameters are extracted from vehicle motion posture characteristic data. The correlation between these parameters and laser point cloud terrain monitoring data is analyzed to determine the point cloud distortion patterns caused by vehicle bumps and steering, generating characteristic data of point cloud motion distortion impact under rainy conditions. Based on the technical parameters of the lidar, combined with the energy attenuation law of the laser beam during propagation due to raindrop scattering, and the matching probability distribution of the point cloud and the real terrain, a laser point cloud error covariance matrix is ​​constructed. The diagonal elements of the matrix correspond to the error variance of the laser point cloud in the x, y, and z axes, respectively, while the off-diagonal elements are the covariance. Based on the above distortion characteristic data and error covariance matrix, a point cloud calibration model for rainy conditions is established. The laser point cloud terrain monitoring data is filtered to remove noise points caused by rain and fog interference, and to correct distortions such as point cloud stretching and offset caused by vehicle motion, resulting in laser point cloud calibrated terrain data. The calibrated point cloud data is transformed from the LiDAR coordinate system to the terrain coordinate system. During the transformation, the transformation error is calculated using a coordinate transformation matrix and combined with the Jacobian matrix to generate point cloud terrain coordinate data. This coordinate data is then processed to separate elevation and horizontal information, generating point cloud terrain elevation and horizontal dimensionality-reduced data respectively. Vehicle motion pose characteristic data is analyzed to obtain parameters such as position deviation and attitude error during vehicle motion. These parameters are applied to the elevation and horizontal dimensionality-reduced data to correct terrain data deviations caused by vehicle motion errors, resulting in pose error point cloud terrain elevation and horizontal dimensionality-reduced data. Finally, based on the continuity characteristics of the terrain, the confidence ellipse algorithm is used to determine the fusion weights of point cloud data from different regions. The elevation and horizontal dimensionality-reduced data are then fused to generate a vehicle-facing perception terrain map, which accurately reflects the terrain undulations and planar distribution of the monitored area.

[0068] Step S3: Optimize the vehicle front perception terrain map data by performing vehicle positioning cumulative error processing on the vehicle front perception terrain map data to generate optimized vehicle front perception terrain map data;

[0069] In this embodiment of the invention, based on the working principle of the inertial measurement unit (IMU), an IMU pre-integration model is constructed. This model includes error parameters for the gyroscope and accelerometer. By integrating measurement data over a continuous time period, the correlation between vehicle motion states at different times is established. Multiple continuous motion intervals are divided from the vehicle motion pose characteristic data, and the acceleration and angular velocity errors within each interval are calculated to obtain vehicle motion interval error parameters. These error parameters are input into the IMU pre-integration model, and the parameters are optimized using the error compensation algorithm within the model to generate optimized vehicle motion error parameters. Based on the optimized error parameters, the position, attitude, and other parameters in the vehicle motion pose characteristic data are adjusted in reverse to correct the pose deviation caused by IMU drift, generating corrected vehicle motion pose characteristic data. The laser point cloud terrain monitoring data is processed, extracting a keyframe point cloud every 5 meters. By comparing the position changes of adjacent keyframe point clouds, the relative motion error measured by the lidar is calculated, generating laser mileage error parameters. The vehicle motion pose characteristic data is fused with laser odometer error parameters. A tight coupling method is used to correlate the error information of the two. An error state Kalman filter algorithm is applied to analyze the reliability of different data, dynamically allocating fusion weights and assigning higher weights to data with high reliability. A factor graph is constructed based on the fused weight data. Nodes in the graph represent vehicle pose and error parameters, and edges represent constraints between parameters. Through iterative solving, the total error of the factor graph is minimized, generating optimized cumulative error data for vehicle positioning. This optimized data is then used to correct the vehicle-ahead perception terrain map data, adjusting terrain coordinates and feature distribution to generate optimized vehicle-ahead perception terrain map data.

[0070] Step S4: Transmit the terrain image feature data to the optimized front-view terrain map data for visual feature matching and mapping processing to generate optimized front-view visual terrain map data; perform real-time terrain anomaly detection processing based on the optimized front-view visual terrain map data to generate real-time terrain anomaly detection data; and feed the real-time terrain anomaly detection data back to the terminal to perform real-time terrain anomaly detection feedback operation under rainy conditions.

[0071] In this embodiment of the invention, terrain image feature data is spatially aligned with optimized vehicle-aware topographic map data. By calculating the mapping relationship between image pixels and terrain three-dimensional coordinates, visual features are accurately mapped to terrain spatial locations, generating optimized vehicle-aware visual topographic map data. This data is then analyzed, and based on features such as road surface color, texture, and water distribution, the terrain is classified into types such as slippery asphalt roads, muddy dirt roads, gravel roads, and waterlogged areas, generating visual terrain classification data. An improved semantic segmentation model is used to process the visual terrain classification data. The model extracts deep image features through a deep convolutional neural network, combines skip connections to preserve detailed information, and assigns attribute labels to different terrain types, such as "slippery asphalt road - high waterlogging" and "muddy dirt road - deep depressions," generating attribute visual terrain classification data. A pre-established terrain level prior database is invoked, which stores the correspondence between visual features and adhesion coefficient ranges for various terrain types. By querying the database, the corresponding terrain visual adhesion coefficient is matched to the attribute visual terrain classification data. Parameters such as longitudinal acceleration and lateral acceleration are extracted from vehicle motion posture data. Combined with the vehicle dynamics model, the adhesion force of the vehicle when driving on different terrains is calculated, generating the terrain kinematic adhesion coefficient. A Gaussian mixture model is used to analyze the probability distribution of the visual and kinematic terrain adhesion coefficients. The probability density functions of the two are weighted and fused to generate terrain adhesion fusion data. A self-supervised learning model is used to process the fusion data. The model learns the variation law of terrain adhesion coefficient from a large amount of historical data in an unsupervised manner, optimizing the uncertainty of probability distribution in the current data, such as correcting the abrupt change in adhesion coefficient caused by local water accumulation, generating self-supervised optimized terrain adhesion data. Based on this data, judgment criteria are set. When the terrain adhesion coefficient exceeds the normal range or the terrain elevation change rate exceeds the threshold, it is judged as terrain anomaly, generating real-time terrain anomaly detection data for front-view perception. This data includes the three-dimensional coordinates of the anomaly location, the anomaly type, and the severity. After being transmitted to the vehicle control system, the system adjusts parameters such as vehicle speed and suspension stiffness according to the data to achieve safe driving control in rainy conditions.

[0072] Furthermore, step S1 includes the following steps:

[0073] The vehicle's built-in multi-source monitoring sensors include lidar equipment, monitoring camera equipment, and inertial measurement unit equipment;

[0074] The vehicle's built-in lidar equipment is used to perform lidar point cloud terrain monitoring in the rainy weather monitoring area, generating lidar point cloud terrain monitoring data.

[0075] The vehicle's built-in monitoring camera is used to monitor and process the terrain image features of the monitoring area in rainy weather, generating terrain image feature data.

[0076] The vehicle's motion posture characteristics are collected using the inertial measurement unit (IMU) built into the vehicle, generating vehicle motion posture characteristic data.

[0077] In this embodiment of the invention, the multi-source monitoring sensors built into the vehicle include a lidar device, a monitoring camera device, and an inertial measurement unit device.

[0078] The vehicle's built-in LiDAR device is used to perform laser point cloud terrain monitoring in a rainy scene. After activation, the LiDAR emitter emits laser beams at preset angle intervals. These beams penetrate rain and fog, illuminating the ground, rocks, puddles, and other surfaces in the monitoring area. Part of the laser energy is reflected back to the LiDAR receiver. The receiver records the round-trip time interval of the laser beam and calculates the distance between the LiDAR and the reflecting point based on the speed of light. Combining this with the horizontal and vertical angles at the time of laser beam emission, trigonometric calculations are used to convert the distance information into three-dimensional coordinates of the reflecting point in the LiDAR coordinate system. Each frame of point cloud data contains the three-dimensional coordinates of all reflecting points within the monitoring area. These coordinates are arranged chronologically, forming a continuous laser point cloud stream that comprehensively records the geometric changes in the terrain ahead as the vehicle travels. The resulting laser point cloud terrain monitoring data can be directly used for subsequent terrain modeling and analysis. The vehicle's built-in monitoring camera device is also used to perform terrain image feature monitoring and processing in the rainy scene monitoring area. After the monitoring camera is activated, the lens automatically focuses on the monitoring area in front of the vehicle. The image sensor converts the optical image of the monitoring area into an electrical signal through photoelectric conversion. After amplification and noise reduction, the electrical signal is converted into a digital image. The three channel values ​​of each pixel in the digital image correspond to the intensity of red, green, and blue colors, respectively. These intensity values ​​can distinguish color differences on the terrain surface; for example, waterlogged areas are brighter due to stronger light reflection, while muddy areas are darker due to greater light absorption. Simultaneously, the pixel grayscale changes in the image reflect the texture characteristics of the terrain surface; for example, gravel roads exhibit a rough texture distribution, while asphalt roads exhibit a relatively smooth texture distribution. The monitoring camera continuously captures images at a fixed frame rate, generating a sequence of images that completely records the changes in the visual characteristics of the monitoring area over time. The resulting terrain image feature data can be used for subsequent terrain type identification and visual matching. The vehicle's motion and posture characteristics are collected using the vehicle's built-in inertial measurement unit (IMU). After the IMU is activated, the three-axis accelerometer senses the vehicle's acceleration in three directions in real time. When the vehicle accelerates, decelerates, or experiences bumps, the mass block inside the accelerometer displaces, and the acceleration value is obtained by measuring the corresponding electrical signal change. The three-axis gyroscope senses the vehicle's rotational motion in real time. When the vehicle turns, pitches, or rolls, the vibrating mass inside the gyroscope deflects due to the Coriolis force, and the angular velocity value is obtained by measuring the corresponding electrical signal change. Acceleration and angular velocity data are continuously output at the sampling frequency. Through integration, the vehicle's velocity and displacement changes can be further obtained. These data together constitute the vehicle's motion and posture characteristic data, comprehensively reflecting the vehicle's motion state during driving. This provides a foundation for subsequent analysis of the impact of vehicle motion on terrain measurement and for achieving spatiotemporal alignment of multi-source data.

[0079] Furthermore, as an embodiment of the present invention, reference is made to... Figure 2 As shown, Figure 1A detailed flowchart of step S2 is shown below. In this embodiment, step S2 includes the following steps:

[0080] Step S21: Based on the vehicle motion posture characteristic data, the vehicle motion acceleration, vehicle motion angular velocity and vehicle motion speed are used to analyze the influence characteristics of point cloud motion distortion under rainy conditions, and generate point cloud motion distortion influence characteristic data under rainy conditions.

[0081] In this embodiment of the invention, continuous sequences of vehicle motion acceleration, angular velocity, and velocity are extracted from vehicle motion posture characteristic data. These sequences are then strictly aligned with concurrently collected laser point cloud terrain monitoring data by timestamp. By comparing the point cloud morphology when the vehicle is stationary with that under different motion states, the impact of acceleration changes on the point cloud is analyzed. Longitudinal acceleration stretches the point cloud along the driving direction, with the degree of stretching increasing significantly with the absolute value of acceleration. Lateral acceleration causes the point cloud to shift towards the turning side of the vehicle, with the shift amount positively correlated with the magnitude of lateral acceleration. Combined with angular velocity data, point cloud distortion caused by vehicle pitch and roll movements is identified. Changes in pitch angle cause layered shifts in the vertical direction of the point cloud, with the upper and lower layers shifting in opposite directions. Changes in roll angle cause horizontal distortion of the point cloud, with the degree of distortion intensifying as the roll angle increases. These distortion features are correlated with corresponding motion parameters to quantify distortion patterns under different motion parameters, generating characteristic data on the impact of point cloud motion distortion under rainy conditions.

[0082] Step S22: Perform covariance matrix analysis on the laser point cloud error characterization based on the laser point cloud terrain monitoring data to generate the covariance matrix of the laser point cloud error characterization under rainy weather conditions.

[0083] In this embodiment of the invention, feature points with known true coordinates (such as fixed curbs and the bottom of markers) in the laser point cloud terrain monitoring data are selected, and the deviations between the measured coordinates and the true coordinates of these feature points are calculated. The distribution patterns of the deviations in the x, y, and z axes are analyzed to determine the distance measurement error caused by raindrop scattering during laser beam propagation, which shows a trend of increasing deviation with increasing measurement distance. Simultaneously, the deviation values ​​of a large number of feature points are statistically analyzed, and the variance of the deviation in each axis direction and the covariance of the inter-axis deviations are calculated to construct a laser point cloud error covariance matrix. The diagonal elements in the matrix correspond to the error variances in the x, y, and z axis directions, reflecting the dispersion of measurement errors along each axis; the off-diagonal elements represent the covariance, reflecting the correlation between errors along different axes. This matrix comprehensively characterizes the error distribution characteristics of laser point clouds in rainy weather scenarios, generating a rainy weather laser point cloud error characterization covariance matrix.

[0084] Step S23: Based on the feature data of motion distortion of point cloud under rainy conditions and the covariance matrix of laser point cloud error characterization under rainy conditions, establish a point cloud calibration model for rainy conditions. Then, use the point cloud calibration model for rainy conditions to filter and remove distortion from the original point cloud topography of laser point cloud topography monitoring data under rainy conditions to obtain laser point cloud calibration topography data.

[0085] In this embodiment of the invention, a distortion correction sub-model is constructed based on the characteristic data of motion distortion affecting point clouds under rainy conditions. This model queries the corresponding distortion coefficients according to vehicle motion parameters (acceleration, angular velocity, and speed) and corrects the stretching, offset, and distortion of the point cloud through coordinate transformation. A noise filtering sub-model is constructed by combining the covariance matrix representing the error of the laser point cloud under rainy conditions. This model uses a covariance-based Gaussian filtering algorithm to statistically analyze the neighborhood points of each point cloud point and remove noise points that deviate from the normal distribution range (such as isolated points caused by raindrop reflection). The two sub-models are integrated into a point cloud calibration model for rainy conditions. First, the distortion correction sub-model corrects the geometric shape of the point cloud, and then the noise filtering sub-model removes residual noise, ultimately obtaining laser point cloud calibration terrain data. This data more accurately reflects the true geometric shape of the terrain.

[0086] Step S24: Based on the laser point cloud calibration terrain data, perform point cloud terrain coordinate transformation and solution processing to generate point cloud terrain coordinate data;

[0087] In this embodiment of the invention, the transformation relationship between the lidar coordinate system and the terrain coordinate system is determined. The lidar coordinate system has its origin at the lidar center, with the x-axis along the vehicle's forward direction, the y-axis perpendicular to the x-axis pointing to the left of the vehicle, and the z-axis pointing vertically upwards. The terrain coordinate system has its origin at a fixed point within the monitoring area, with the x-axis extending along the road centerline, the y-axis perpendicular to the x-axis pointing to one side of the road, and the z-axis perpendicular to the ground pointing upwards. By measuring the relative positions of the origins of the two coordinate systems and the angles between their axes, a rotation matrix and a translation vector are established. The coordinates of each point in the lidar point cloud calibration terrain data are transformed using the rotation matrix, and then the position is adjusted by superimposing the translation vector, thus completing the transformation from the lidar coordinate system to the terrain coordinate system. During the transformation process, the Jacobian matrix is ​​calculated to analyze the amplification effect of the coordinate transformation on the error, ensuring the accuracy of the transformed coordinates and generating point cloud terrain coordinate data.

[0088] Step S25: Perform elevation and horizontal dimensionality reduction processing on the point cloud terrain coordinate data to generate point cloud terrain elevation dimensionality reduction data and point cloud terrain horizontal dimensionality reduction data.

[0089] In this embodiment of the invention, for each point in the point cloud terrain coordinate data, its z-axis coordinate is extracted as elevation information, and its x-axis and y-axis coordinates are extracted as horizontal position information, thus separating elevation and horizontal data. The elevation information is then gridded, dividing the terrain region into equally sized grids. The average elevation of all points within each grid is calculated, forming elevation grid data reflecting the terrain undulations, i.e., point cloud terrain elevation dimensionality-reduced data. The same gridding process is applied to the horizontal position information, calculating the x and y coordinate distribution density of points within each grid, forming horizontal grid data reflecting the terrain planar distribution, i.e., point cloud terrain horizontal dimensionality-reduced data. This dimensionality reduction process reduces the data volume while preserving key terrain features, facilitating subsequent analysis and correction of errors in the elevation and horizontal directions.

[0090] Step S26: Perform vehicle motion pose error analysis based on vehicle motion pose characteristic data to generate vehicle motion pose error data; perform horizontal and vertical error analysis on point cloud terrain elevation dimensionality reduction data and point cloud terrain horizontal dimensionality reduction data based on vehicle motion pose error data to generate pose error point cloud terrain elevation dimensionality reduction data and pose error point cloud terrain horizontal dimensionality reduction data respectively.

[0091] In this embodiment of the invention, continuous position, velocity, and attitude sequences are extracted from vehicle motion pose characteristic data and compared time-by-time with synchronously acquired high-precision reference data (such as the true position and attitude provided by differential GPS). The differences between the measured position and the true position at each time point are calculated in the x, y, and z axes to obtain the three-dimensional position error; the differences between the measured attitude angles (pitch, roll, and yaw) and the true attitude angles are calculated to obtain the attitude error. These position and attitude errors are integrated in chronological order to form vehicle motion pose error data. For point cloud terrain elevation dimensionality reduction data, the impact of z-axis position error and pitch angle error is analyzed: z-axis position error causes overall elevation data offset, which is corrected by subtracting the z-axis position error at the corresponding time point from the z-value of each elevation grid; pitch angle error causes elevation data to tilt along the driving direction, which is calculated by establishing a geometric relationship between pitch angle and elevation correction, and then superimposed to generate pose error point cloud terrain elevation dimensionality reduction data. For point cloud terrain horizontal dimensionality reduction data, the impact of x and y axis position errors and roll and heading angle errors is analyzed: x and y axis position errors directly cause horizontal grid offset, which is corrected by subtracting the position error value from the x and y coordinates of the corresponding grid; roll angle error causes the horizontal grid to tilt laterally, and heading angle error causes the horizontal grid to rotate. The tilt and rotation effects are eliminated by coordinate rotation transformation, respectively, to generate pose error point cloud terrain horizontal dimensionality reduction data.

[0092] Step S27: Perform multi-source error propagation mechanism-based front-view topographic map fusion processing on the pose error point cloud terrain elevation dimensionality reduction data and the pose error point cloud terrain horizontal dimensionality reduction data to generate front-view topographic map data.

[0093] In this embodiment of the invention, the elevation-level dimensionality-reduced data and the horizontal-level dimensionality-reduced data of the pose error point cloud are aligned by timestamp to ensure a one-to-one correspondence in spatial location. Based on a multi-source error propagation mechanism, the error correlation between the two types of data is analyzed. For example, the heading angle error in the horizontal direction indirectly affects the forward and backward distribution accuracy of the elevation data, and the pitch angle error in the vertical direction is related to the near and far position accuracy of the horizontal data. This correlation is quantitatively described by the error propagation coefficient. For continuous terrain areas (such as flat roads), a weighted fusion algorithm is used, assigning weights according to the error variance of the two types of data, with data having smaller errors receiving higher weights. The elevation grid and the horizontal grid are superimposed according to their weights to form a continuous transition terrain feature. For areas with abrupt terrain changes (such as road edges and steep slopes), a boundary enhancement mechanism is enabled to retain vertical abrupt changes in the elevation data and horizontal boundaries in the horizontal data, enhancing boundary clarity through morphological operations. After fusion, a consistency check is performed, calculating the matching degree between the elevation and horizontal coordinates of each grid. Grids with low matching degrees are re-applied to the error propagation model for correction. The final generated vehicle-ahead perception terrain map data maintains a high degree of spatial consistency, accurately reflecting the undulation and planar distribution of the terrain.

[0094] Furthermore, step S22 includes the following steps:

[0095] Based on the analysis of laser point cloud terrain monitoring data, the laser beam uncertainty parameters and terrain matching uncertainty parameters under rainy conditions are analyzed.

[0096] Based on the preset lidar beam model and the likelihood function, the covariance matrix analysis of the laser point cloud error characterization is performed on the laser beam uncertainty parameters and terrain matching uncertainty parameters under rainy conditions, generating the laser point cloud error characterization covariance matrix under rainy conditions.

[0097] In this embodiment of the invention, subsets of point clouds with different distance ranges are selected from the laser point cloud terrain monitoring data. Each subset contains distance information between the lidar and the reflecting point, as well as reflection intensity data. The characteristics of the laser beam during rain propagation are analyzed. By comparing the difference in reflection intensity between sunny and rainy days at the same distance, the laser energy attenuation law caused by raindrop scattering is determined, and the changes in the laser beam divergence angle and propagation path offset are quantified. These parameters together constitute the laser beam uncertainty parameters. Simultaneously, point cloud clusters corresponding to known real terrain features (such as fixed obstacle edges and road markings) from the laser point cloud terrain monitoring data are selected. The positional deviation of each point in the point cloud cluster from the real terrain feature is calculated, and the distribution range and density of the deviation are statistically analyzed. The matching deviation caused by the blurring of the reflected signal due to the wet terrain surface and water accumulation is determined, and this is used as the terrain matching uncertainty parameter. The lidar beam model includes the relationship between laser beam propagation distance and energy attenuation, and the correlation between the beam divergence angle and the measurement range. The laser beam uncertainty parameters are substituted into this model to calculate the measurement error components of the laser point cloud in the x, y, and z axes at different propagation distances. The likelihood function is constructed based on terrain matching uncertainty parameters. By calculating the matching probability distribution between the laser point cloud and the real terrain, the probability of error occurrence is determined; a lower probability indicates a larger error. The error components output by the beam model are combined with the probability distribution obtained from the likelihood function to calculate the variance of the error in each axis direction, reflecting the degree of dispersion of the error in that direction. The covariance of the errors between different axes is also calculated, reflecting the degree of mutual influence between the errors of each axis. Based on these variances and covariances, a laser point cloud error covariance matrix is ​​constructed. The diagonal elements of the matrix represent the variance of each axis, and the off-diagonal elements represent the inter-axis covariance, comprehensively characterizing the error distribution characteristics of the laser point cloud in rainy weather scenarios, generating a rainy weather laser point cloud error characterization covariance matrix.

[0098] Furthermore, the lidar beam model is used to perform characteristic analysis of the lidar-emitted laser beam during propagation and reflection under rainy conditions, based on the lidar beam uncertainty parameters. The likelihood function is used to perform analysis of the degree of matching between lidar measurements and actual terrain under rainy conditions, based on the terrain matching uncertainty parameters.

[0099] Furthermore, step S24 includes the following steps:

[0100] The true terms and Gaussian error terms of the laser point cloud calibration terrain data are analyzed using a pre-set Gaussian model. Then, the linear combination analysis of the laser point cloud coordinates and laser radar measurement error terms is performed on the true terms and Gaussian error terms of the laser point cloud calibration terrain data using a small angle approximation algorithm to obtain the linear feature data of laser point cloud coordinates-error terms.

[0101] The linear feature data of the point cloud coordinates and error terms of the LiDAR are processed to transform the point cloud terrain coordinates, generating point cloud terrain coordinate transformation data. The point cloud terrain coordinates are then solved using the Jacobian matrix to obtain the point cloud terrain coordinate data.

[0102] In this embodiment of the invention, the preset Gaussian model includes a mean function and a variance function. The mean function characterizes the central tendency of the real terrain features in the laser point cloud calibration terrain data, while the variance function describes the dispersion of the data around the mean. The laser point cloud calibration terrain data is input into the Gaussian model. The mean function extracts the true terms reflecting the actual terrain morphology. These true terms are composed of the statistical average of a large amount of point cloud data, representing the terrain coordinates after eliminating error interference. The variance function separates the error terms that conform to a Gaussian distribution. These error terms include random errors caused by equipment noise and environmental interference during the laser radar measurement process. A small-angle approximation algorithm is introduced to simplify the sine and cosine values ​​of the horizontal and vertical angles, decomposing the three-dimensional coordinates of the true terms into a trigonometric function combination of distance, horizontal angle, and vertical angle. Simultaneously, the Gaussian error terms are decomposed into linear expressions of distance error, horizontal angle error, and vertical angle error. Matrix operations are used to integrate the true terms and error terms, establishing a linear correspondence between the laser radar point cloud coordinates and each measurement error term, forming linear feature data of laser radar point cloud coordinates and error terms. The transformation relationship between the LiDAR coordinate system and the terrain coordinate system is determined, including the translation parameters between the origins of the two coordinate systems and the rotation parameters between the coordinate axes. The spatial relationship between the LiDAR coordinate system and the terrain coordinate system is clarified: the LiDAR coordinate system has its origin at the LiDAR installation center, with the x-axis along the vehicle's forward direction, the y-axis pointing to the left of the vehicle, and the z-axis vertically upward; the terrain coordinate system has its origin at a fixed reference point within the monitoring area, with the x-axis along the road's extension direction, the y-axis perpendicular to the x-axis pointing to one side of the road, and the z-axis perpendicular to the ground upward. The translation parameters between the origins of the two coordinate systems and the rotation angles between the coordinate axes are obtained through field measurements, establishing a rotation matrix and a translation vector. For each point coordinate in the linear feature data of the LiDAR point cloud coordinates - error term, the orientation is first transformed using the rotation matrix to align the coordinate axes of the LiDAR coordinate system with those of the terrain coordinate system. Then, a translation vector is superimposed to correct the positional offset, completing the transformation from the LiDAR coordinate system to the terrain coordinate system and generating point cloud terrain coordinate transformation data. A Jacobian matrix is ​​constructed, where each element corresponds to the partial derivative of the coordinate transformation function with respect to the translation and rotation parameters. This matrix quantifies the impact of minute changes in the transformation parameters on the coordinate transformation result. The point cloud terrain coordinate transformation data is substituted into the Jacobian matrix, and matrix operations are used to calculate the coordinate deviation caused by parameter errors during the transformation process. Based on the calculated deviation, the initially transformed coordinate data is corrected to eliminate the deviation introduced by parameter errors, ultimately obtaining the accurate point cloud coordinates in the terrain coordinate system, i.e., the point cloud terrain coordinate data.

[0103] In another embodiment of the present invention, the small-angle approximation algorithm for linear combination analysis of lidar point cloud coordinates and lidar measurement error terms based on the true term and Gaussian error term of the lidar point cloud calibration terrain data is as follows: ,in, , as well as These are respectively identified as noisy 3D coordinates in the lidar coordinate system, true 3D coordinates, and coordinate error terms caused by measurement noise; , as well as These represent the actual distance value, the actual horizontal angle, and the actual vertical angle in the lidar coordinate system, respectively. , as well as Noise in distance measurement, horizontal angle measurement, and vertical angle measurement in the lidar coordinate system.

[0104] Furthermore, step S3 includes the following steps:

[0105] Step S31: Construct the IMU pre-integration model;

[0106] In this embodiment of the invention, the IMU pre-integration model is: ; ; ,in, To monitor the start and end times of the time window, The sampling time interval of the inertial measurement unit. Represented as the discrete time-time index in a time series; , The inertial measurement unit monitors specific moments within the monitoring time window. Angular velocity and acceleration; , , as well as The inertial measurement unit monitors specific moments within the monitoring time window. The gyroscope drift bias, accelerometer offset bias, gyroscope measurement noise, and accelerometer measurement noise; as well as Let be the vehicle's attitude rotation matrix at the start time and the attitude rotation matrix at the end time; , The vehicle's speed at the start and speed at the end. It is the acceleration due to gravity; , The coordinates of the vehicle's position at the start time and at the end time; , , These represent the relative rotation matrix, relative velocity change, and relative position change within the time window, respectively. From the start time to a specific time within the time window The relative rotation matrix and From the start time to a specific time within the time window The relative speed after rotation adjustment.

[0107] Step S32: Analyze the error parameters of the vehicle motion range based on the vehicle motion posture characteristic data to obtain the vehicle motion range error parameters;

[0108] In this embodiment of the invention, continuous acceleration and angular velocity sequences are extracted from vehicle motion posture characteristic data. Multiple vehicle motion intervals are divided according to time intervals, each interval corresponding to a continuous motion process. Time-domain analysis is performed on the acceleration data within each motion interval to calculate the mean and variance of the acceleration, determining the acceleration deviation caused by zero drift of the inertial measurement unit. The same analysis is performed on the angular velocity data output by the gyroscope to obtain the angular velocity deviation. Combining the vehicle's motion state (e.g., constant speed, acceleration, turning), the deviation is correlated with the motion state to determine the variation law of error parameters under different motion states, such as the amplification characteristics of acceleration deviation during acceleration and the fluctuation range of angular velocity deviation during turning. These deviations and variation laws are quantified into specific parameters to form the vehicle motion interval error parameters.

[0109] Step S33: Transmit the vehicle motion range error parameters to the IMU pre-integration model to optimize the vehicle motion error parameters and generate optimized vehicle motion error parameters;

[0110] In this embodiment of the invention, the vehicle motion interval error parameters are input into the IMU pre-integration model. The model first analyzes the type of error parameters, distinguishing between different error terms such as zero bias in acceleration, zero bias in angular velocity, and scale factor error. For zero bias in acceleration, the model performs preliminary correction by directly subtracting the zero bias value from the original acceleration measurement. For zero bias in angular velocity, after subtracting the zero bias value from the original angular velocity measurement, the model further performs cross-compensation on the three-dimensional angular velocity components based on non-orthogonal error parameters. The corrected acceleration and angular velocity data are substituted into the integral equation of the pre-integration model to recalculate the vehicle's position, velocity, and attitude sequence. By comparing the deviations of the motion sequence before and after correction with the reference trajectory (synchronously acquired by high-precision equipment), the model iteratively adjusts the compensation coefficients of the error parameters to minimize the deviation between the corrected motion sequence and the reference trajectory. The finally determined error parameters are the optimized parameters for vehicle motion error.

[0111] Step S34: Perform reverse correction processing on the vehicle motion pose characteristic data based on the vehicle motion error optimization parameters to generate corrected vehicle motion pose characteristic data.

[0112] In this embodiment of the invention, a reverse correction function is constructed based on vehicle motion error optimization parameters. This function uses time as a variable and outputs the position, velocity, and attitude correction amounts at each moment. For position data, the function calculates the position correction value at each moment in reverse chronological order according to the cumulative position deviation law in the error optimization parameters, subtracts the correction value from the original position data, and eliminates the cumulative deviation. For velocity data, the function calculates the velocity correction value at each moment based on the correlation formula between velocity error and acceleration error in the error optimization parameters and adds it to the original velocity data. For attitude data, the function performs a reverse rotation correction on the original attitude angles through Euler angle rotation transformation based on the attitude angle deviation in the error optimization parameters, ensuring that the corrected attitude is consistent with the true attitude. All correction operations strictly follow the time correspondence principle, and each moment's motion pose data corresponds to a unique correction amount, ultimately generating corrected vehicle motion pose characteristic data.

[0113] Step S35: Perform laser mileage error analysis based on laser point cloud terrain monitoring data and generate laser mileage error parameters;

[0114] In this embodiment of the invention, keyframes are extracted from laser point cloud terrain monitoring data at fixed time intervals. Each point cloud frame contains three-dimensional coordinate information within the monitoring area. A point cloud registration algorithm is used to calculate the relative pose changes between adjacent keyframes, obtaining the relative motion parameters (translation and rotation) measured by the lidar. These relative motion parameters are accumulated to generate a laser mileage trajectory. Simultaneously, the actual trajectory within the same time period is acquired using a high-precision positioning device, and the positional deviation between the laser mileage trajectory and the actual trajectory at each keyframe is calculated. These deviations are statistically analyzed to determine the cumulative pattern of the deviation over time, such as a linear growth trend or periodic fluctuation characteristics. Combined with the registration residual (average distance error between point clouds) generated during the point cloud registration process, the error components caused by point cloud noise and terrain changes in laser mileage measurement are quantified, and finally integrated into laser mileage error parameters.

[0115] Step S36: Based on the corrected vehicle motion pose characteristic data and laser mileage error parameters, perform parameter tight coupling and factor graph optimization analysis of the vehicle positioning cumulative error to generate optimized vehicle positioning cumulative error data;

[0116] In this embodiment of the invention, the corrected vehicle motion pose characteristic data and laser odometry error parameters are tightly coupled to establish an error correlation equation between the two. In the equation, the corrected vehicle motion pose data provides the absolute pose constraint of the vehicle, while the laser odometry error parameters provide the error boundary for relative motion. The error information of the two types of data is fused through weighted summation, with the weights determined by their respective error variances; the smaller the variance, the greater the weight. The fused data is used to construct a factor graph. Nodes in the graph represent the vehicle's pose and error parameters at each time step, while edges represent the constraint relationships between poses at adjacent time steps (defined by the laser odometry error parameters) and the absolute pose constraints (defined by the corrected vehicle motion pose data). A Gaussian-Newton iterative algorithm is used to optimize the factor graph. By adjusting the node parameters, the total error of all constraints is minimized. The error parameters obtained after the iteration terminates are the optimized data for the vehicle positioning cumulative error.

[0117] Step S37: Optimize the vehicle front perception terrain map data by using the vehicle positioning cumulative error optimization data to generate optimized vehicle front perception terrain map data.

[0118] In this embodiment of the invention, the position and attitude errors in the vehicle positioning cumulative error optimization data are applied to the vehicle-ahead perception terrain map data. For each point cloud coordinate in the terrain map, a translation adjustment is performed based on the position error correction amount to ensure that the absolute position of the point cloud in the terrain coordinate system is consistent with its true position; a rotation transformation is performed based on the attitude error correction amount to eliminate the tilt or distortion of the terrain map caused by vehicle attitude deviation. During the adjustment process, the relative positional relationship between point clouds remains unchanged to ensure that terrain features (such as the relative size and position of protrusions and depressions) are not affected. After the correction is completed, the continuity of the terrain map data is checked, the connection error of adjacent area point clouds is calculated, and areas with errors exceeding the threshold are reapplied with error optimization data for fine-tuning. The final optimized vehicle-ahead perception terrain map data has significantly improved positional accuracy in the global coordinate system, and the spatial distribution of terrain features is consistent with the real terrain.

[0119] Furthermore, step S35 includes the following steps:

[0120] Keyframes are extracted from the laser point cloud terrain monitoring data to obtain keyframe laser point cloud terrain monitoring data;

[0121] The laser point cloud terrain monitoring data of key frames is analyzed to perform laser point cloud change analysis between adjacent frames, and laser point cloud change data between adjacent frames is generated.

[0122] Based on the keyframe laser point cloud terrain monitoring data and the vehicle-front perception terrain map data, the keyframe and corresponding local map positioning error analysis is performed to obtain the keyframe positioning error data.

[0123] Laser mileage error parameters are generated by analyzing the laser point cloud change data of adjacent frames and the positioning error data of key frames.

[0124] In this embodiment of the invention, laser point cloud terrain monitoring data is extracted at fixed time intervals, and the extraction time must strictly correspond to the sampling time of vehicle motion pose characteristic data. During the extraction process, the information content of each frame of point cloud is calculated. The information content is determined by the density, coverage area, and number of terrain features contained in the point cloud. Point cloud frames with information content reaching a set threshold are retained as key frames. The extracted key frames are deduplicated. If the overlapping area of ​​two consecutive frame point clouds exceeds a set proportion and the change in terrain features is less than a set threshold, the latter frame is discarded. This ensures that the key frames can fully reflect the terrain changes without causing excessive data volume due to redundancy, ultimately obtaining key frame laser point cloud terrain monitoring data. Two consecutive key frame laser point clouds are selected, and a point cloud registration algorithm is used to calculate the spatial transformation relationship between them. This algorithm finds corresponding feature point pairs in the two frame point clouds, solves for the translation vector and rotation matrix, and describes the position and attitude changes of the latter frame relative to the former frame. The distance deviation of corresponding points in the overlapping area of ​​the two frame point clouds after transformation is calculated, and the mean and variance of the deviation are statistically analyzed to reflect the accuracy of point cloud registration. Simultaneously, the point cloud features of non-overlapping areas are analyzed to identify newly detected terrain areas and terrain areas outside the detection range due to vehicle movement, quantifying the changes in area and point cloud quantity of these areas. Translation vectors, rotation matrices, deviation statistics, and regional change information are integrated to generate laser point cloud change data for adjacent frames. Local map regions corresponding to the laser point cloud of each keyframe are extracted from the vehicle-ahead perception terrain map data. The scope of the local map is based on the detection center of the keyframe and covers the same terrain area. A feature matching algorithm is used to find matching terrain features (such as raised edges and recessed boundaries) between the keyframe laser point cloud and the corresponding local map, calculating the coordinate deviation of feature points in the keyframe coordinate system and the local map coordinate system. The deviations of all matching feature points are statistically analyzed, calculating the average and maximum deviations in each coordinate axis direction to determine the positioning offset of the keyframe in the local map. Combining the known position of the local map in the global coordinate system, the positioning offset of the keyframe is converted into an error in the global coordinate system, generating keyframe positioning error data. The translation vectors and rotation matrices in the laser point cloud change data of adjacent frames are cumulatively calculated to obtain the odometer trajectory based on the laser point cloud. The trajectory is compared with the global positioning information in the keyframe positioning error data, and the trajectory deviation at each keyframe is calculated. The deviation is the difference between the laser mileage trajectory position and the global positioning position. The variation of the deviation over time is analyzed to determine the error accumulation rate, which is determined by the increment of the deviation per unit time. The distribution characteristics of the deviation under different terrain types (such as flat roads, slopes, and waterlogged areas) are statistically analyzed to determine the degree of influence of terrain on laser mileage error. The error accumulation rate, error distribution characteristics under different terrains, and deviation statistics are integrated to generate laser mileage error parameters.

[0125] Furthermore, step S36 includes the following steps:

[0126] Based on the corrected vehicle motion pose characteristic data and laser mileage error parameters, a tight coupling analysis of the corrected vehicle motion pose and laser mileage error is performed to generate corrected pose-mileage error tight coupling data.

[0127] The corrected pose-odometer error tightly coupled data is fused and weighted using a pre-defined error state Kalman filter algorithm to obtain the fused weighted corrected pose-odometer error tightly coupled data.

[0128] Based on the fusion weighted pose-mileage error tightly coupled data, factor graph optimization analysis of vehicle positioning cumulative error is performed to generate optimized vehicle positioning cumulative error data.

[0129] In this embodiment of the invention, the position, velocity, and attitude information in the corrected vehicle motion pose characteristic data are used as the basic framework, and the relative motion error and cumulative error rules in the laser mileage error parameters are integrated into this framework. A correlation equation between pose and mileage error is established, with the corrected vehicle pose parameters on the left and the laser mileage measurement value and corresponding error term on the right. Matrix operations are used to mathematically fuse the absolute reference information of the vehicle pose with the relative error constraint of the laser mileage, enabling the pose data to correct the cumulative deviation of the laser mileage, while the laser mileage error parameter constrains the drift of the pose data. The consistency of the time dimension is maintained during the fusion process; the pose data at each moment is coupled with the laser mileage error parameter at the corresponding moment, ultimately generating corrected pose-mileage error tightly coupled data containing the correlation information between pose and mileage error. The preset error state Kalman filter algorithm includes a state equation and an observation equation. The state equation describes the dynamic change law of the error parameters in the corrected pose-mileage error tightly coupled data, and the observation equation establishes the relationship between the error parameters and the measured values. The corrected pose-odometer error tightly coupled data is input into a filtering algorithm to calculate the error covariance matrix of the corrected vehicle motion pose data and the error covariance matrix of the laser odometry error parameters. The fusion weights are determined based on the diagonal elements of the covariance matrices (i.e., the variance of each error term), with smaller variances resulting in larger weights for the error terms. The pose data and the error information of the laser odometry error parameters are fused through weighted summation to generate intermediate data containing weight factors. This data is then fed back into the filtering algorithm's state update process to adjust the estimated values ​​of the error parameters, ultimately yielding the corrected pose-odometer error tightly coupled data with fused weights. A factor graph is constructed based on this fused weighted data. Nodes in the graph represent the vehicle's pose and error parameters at each time step, and edges between nodes represent factor constraints. Factors include prior factors (provided by the corrected vehicle motion pose data), odometry factors (provided by the laser odometry error parameters), and weight factors (determined by the fusion weights). The prior factors constrain the absolute reference of the vehicle pose, the odometry factors constrain the relative changes in pose between adjacent time steps, and the weight factors adjust the degree of influence of different constraints. The gradient descent algorithm is used to optimize the factor map, and the total error of all factors is minimized by iteratively adjusting the node parameters. During the iteration process, the error residual of the factor map is calculated after each adjustment. The iteration stops when the residual is less than a set threshold. The error parameters obtained at this time are the optimized data of vehicle positioning cumulative error, which can quantify the magnitude and distribution pattern of the positioning cumulative error.

[0130] Furthermore, step S4 includes the following steps:

[0131] Step S41: Transmit the terrain image feature data to the optimized front-view perception terrain map data for visual feature matching and mapping processing to generate optimized front-view perception visual terrain map data.

[0132] In this embodiment of the invention, visual features such as edges, textures, and colors are extracted from terrain image feature data. These features must have a spatial correspondence with the terrain geometric features in the optimized vehicle-front perception terrain map data. The visual features are mapped to the three-dimensional coordinate system of the optimized vehicle-front perception terrain map data through spatial coordinate transformation, establishing a one-to-one correspondence between each visual feature point and the point cloud in the terrain map. During the mapping process, a feature point matching algorithm is used to determine the matching relationship by calculating the similarity between the visual features and the point cloud features (such as edge direction consistency and texture distribution correlation), and mismatched feature point pairs are eliminated. Geometric correction is performed on the matched visual features to ensure that the spatial position of the visual features completely matches the terrain structure in the terrain map, ensuring spatial consistency between visual information and point cloud information, and finally generating optimized vehicle-front perception visual terrain map data that integrates visual features and terrain geometric features.

[0133] Step S42: Perform visual terrain type classification processing on the optimized vehicle-front perception visual terrain map data for rainy weather scenarios to generate visual terrain classification data;

[0134] In this embodiment of the invention, visual feature differences in rainy weather scenarios are extracted from optimized visual topographic map data for vehicle-front perception. These differences include the specular reflection intensity of waterlogged areas, the degree of texture blurring on wet roads, and the color saturation of dry areas. Classification thresholds are set based on these feature differences. Areas with reflection intensity exceeding the threshold are classified as waterlogged areas, areas with texture blurring reaching the threshold are classified as wet roads, and areas with color saturation within a specific range are classified as dry areas. Simultaneously, terrain geometric features are used to assist classification; for example, areas with slopes exceeding a set value are classified as either slope waterlogged areas or slope dry areas based on visual features. The spatial continuity of regions is maintained during classification; areas with similar adjacent pixel features are merged into the same category to avoid scattered classification results, generating visual terrain classification data that includes the spatial distribution of different terrain types.

[0135] Step S43: Use a preset semantic segmentation model to label the visual terrain classification data with terrain type attributes, and generate attribute visual terrain classification data;

[0136] In this embodiment of the invention, the preset semantic segmentation model includes an encoding module and a decoding module. The encoding module extracts deep features from the visual terrain classification data through convolution operations, and the decoding module maps the deep features back to the original data size through deconvolution operations, achieving pixel-level classification. Visual terrain classification data is input into the model, and the model determines the terrain type of each pixel based on terrain type features learned during training (such as the reflection pattern of puddles and the texture structure of the road surface). After determination, a unique attribute label is assigned to different terrain types, such as using a specific symbol to identify puddles and another symbol to identify dry roads. The label must include the terrain type name and key feature parameters (such as the reflection intensity range of puddles). During the labeling process, it is ensured that the label completely matches the spatial range of the corresponding area, ultimately generating attribute visual terrain classification data with attribute labels.

[0137] Step S44: Based on the pre-established terrain level prior database, perform terrain visual attachment coefficient mapping processing on the attribute visual terrain classification data for different attributes to generate terrain visual attachment coefficients.

[0138] In this embodiment of the invention, a priori database of terrain levels is pre-established using historical data from visual terrain classification data. This database contains the correspondence between different terrain attributes and visual adhesion coefficients. Each terrain attribute in the database (e.g., water accumulation, dry road surface, mud) is associated with a unique range of visual adhesion coefficients, determined based on extensive measured data on the friction characteristics of terrain surfaces and tires. The terrain type attribute labels in the attribute visual terrain classification data are precisely matched with the terrain attributes in the database. This matching process is achieved by comparing the terrain type name and feature parameters in the labels. Upon successful matching, the visual adhesion coefficient range for the corresponding terrain attribute in the database is retrieved. Based on the specific visual characteristics of the terrain area in the attribute visual terrain classification data (e.g., the reflection intensity of water accumulation, the texture distribution of mud), a unique value is determined within this range and used as the terrain visual adhesion coefficient for that area. This operation is repeated for all terrain areas to ensure that each area is mapped to a corresponding terrain visual adhesion coefficient.

[0139] Step S45: Extract the vehicle's longitudinal kinematics data and lateral kinematics data from the vehicle's motion pose characteristic data. Based on the vehicle's longitudinal kinematics data and lateral kinematics data, perform terrain kinematics attachment coefficient mapping processing on the visual terrain classification data for various types to generate terrain kinematics attachment coefficients.

[0140] In this embodiment of the invention, longitudinal kinematic data (such as longitudinal acceleration and rate of change of longitudinal velocity) and lateral kinematic data (such as lateral acceleration and rate of change of lateral displacement) are extracted from vehicle motion posture characteristic data. For each terrain type in the visual terrain classification data, a mapping rule between kinematic data and adhesion coefficient is established. In the rule, the change amplitude of longitudinal kinematic data is related to the resistance of the terrain to the longitudinal movement of the vehicle, and the change amplitude of lateral kinematic data is related to the constraint of the terrain on the lateral sliding of the vehicle. The extracted longitudinal and lateral kinematic data are substituted into the mapping rule of the corresponding terrain type to calculate the kinematic adhesion coefficient of that terrain type under the current motion state. The calculation process needs to take into account the area and distribution of the terrain region to ensure that the kinematic adhesion coefficient can reflect the overall influence of the terrain on the vehicle motion, and finally generate a corresponding terrain kinematic adhesion coefficient for each terrain type.

[0141] Step S46: Analyze the actual probability distribution of terrain attachment fusion based on the terrain visual attachment coefficient and terrain kinematic attachment coefficient to generate terrain attachment fusion data;

[0142] In this embodiment of the invention, the visual adhesion coefficient and the kinematic adhesion coefficient of the terrain are collected, and their numerical ranges and distribution characteristics are determined. Probability density functions are used to describe the actual probability distributions of the two adhesion coefficients. The probability distribution of the visual adhesion coefficient is determined based on the consistency of the terrain visual features; the more consistent the features, the more concentrated the probability peaks. The probability distribution of the kinematic adhesion coefficient is determined based on the stability of the vehicle motion data; the more stable the data, the more significant the probability peaks. The two probability distributions are fused through joint probability distribution calculation. During the fusion process, weights are assigned according to the reliability of the two adhesion coefficients. The reliability is determined by the acquisition accuracy of each data point; higher accuracy results in higher weights. The resulting terrain adhesion fusion data includes the comprehensive adhesion coefficient and corresponding probability distribution for each terrain region. This data can simultaneously reflect the influence of visual features and kinematic features on terrain adhesion performance.

[0143] Step S47: Based on the preset self-supervised learning model, perform self-supervised optimization processing on the probability distribution uncertainty of actual terrain attachment for the terrain visual attachment coefficient and terrain kinematic attachment coefficient, and generate terrain attachment self-supervised optimization data.

[0144] In this embodiment of the invention, the pre-defined self-supervised learning model includes a feature extraction layer, a contrastive learning layer, and a probability optimization layer. The feature extraction layer extracts high-dimensional features from the visual and kinematic attachment coefficients of the terrain through convolutional operations. These features cover the numerical distribution, spatial correlation, and dynamic trends of the two attachment coefficients. The contrastive learning layer constructs positive and negative sample pairs from the extracted features. Positive sample pairs represent features of the same terrain region at different times, while negative sample pairs represent features of different terrain regions. By calculating the similarity loss of the sample pairs, the model learns the consistency patterns of terrain attachment features. The probability optimization layer adjusts the probability distribution of terrain attachment based on the feature patterns obtained from self-supervised learning. It expands the probability range of regions with high uncertainty (such as transitional zones with ambiguous features) and narrows the probability range of regions with low uncertainty (such as waterlogged areas with clear features). The resulting self-supervised optimized terrain attachment data more accurately reflects the probability distribution characteristics of actual terrain attachment.

[0145] Step S48: Perform real-time terrain anomaly detection processing for front-view perception based on terrain attachment self-supervised optimization data to generate real-time terrain anomaly detection data for front-view perception.

[0146] In this embodiment of the invention, a terrain anomaly detection threshold is set based on terrain attachment self-supervised optimization data. This threshold is determined according to the distribution range of attachment coefficients of historical normal terrain. For each terrain region in the terrain attachment self-supervised optimization data, the deviation between its comprehensive attachment coefficient and the threshold is calculated. The deviation is quantified by the ratio of the absolute difference to the threshold. When the deviation exceeds the set standard, the region is marked as an anomaly candidate region. Combined with optimized vehicle-front perception visual terrain map data, spatial morphology analysis is performed on the anomaly candidate regions to check for continuous anomalous features (such as a sudden drop in attachment coefficient over a large area) and to exclude isolated anomalies caused by measurement noise. For confirmed anomaly regions, their location, range, and degree of anomaly are recorded, generating real-time vehicle-front perception terrain anomaly detection data containing all anomaly information.

[0147] Step S49: Feed back the real-time terrain anomaly detection data from the front perception of the vehicle to the terminal to perform real-time terrain anomaly detection and feedback operations under rainy conditions.

[0148] In this embodiment of the invention, the locations of abnormal areas in the real-time terrain anomaly detection data perceived by the vehicle are converted into spatial coordinates recognizable by the terminal. This coordinate conversion is based on a preset correspondence between the terminal's coordinate system and the terrain coordinate system. After receiving the data, the terminal analyzes the anomaly severity information and sorts them by severity level. The severity level is determined by the area of ​​the abnormal area and the degree of deviation in the adhesion coefficient. Based on the sorting results, the terminal initiates corresponding feedback operations: for slightly abnormal areas, it outputs text prompts indicating their location; for moderately abnormal areas, it simultaneously activates an audible and visual warning, prompting the driver to slow down; for severely abnormal areas, in addition to the warning, it outputs a suggested driving path based on the vehicle's current motion state. The path must avoid the abnormal area and comply with vehicle dynamics constraints. During the feedback operation, the terminal continuously updates the anomaly data to ensure that the feedback information remains consistent with the real-time terrain status.

[0149] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0150] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A real-time terrain anomaly detection method based on multi-source data analysis, characterized in that, The method comprises the following steps: Step S1: multi-source terrain monitoring and collection of the rain scene monitoring area is performed by using the multi-source monitoring sensor built in the vehicle to obtain multi-source terrain monitoring data, wherein the multi-source terrain monitoring data comprises laser point cloud terrain monitoring data, terrain image feature data, and vehicle motion pose characteristic data; Step S2: vehicle front perception terrain map analysis processing based on the laser point cloud terrain monitoring data and the vehicle motion pose characteristic data is performed to generate vehicle front perception terrain map data; Step S3: vehicle front perception terrain map optimization processing of the vehicle positioning cumulative error is performed on the vehicle front perception terrain map data to generate optimized vehicle front perception terrain map data; Step S3 comprises the following steps: Step S31: an IMU pre-integration model is constructed; Step S32: error parameter analysis of the vehicle motion interval is performed according to the vehicle motion pose characteristic data to obtain vehicle motion interval error parameters; Step S33: the vehicle motion interval error parameters are transmitted to the IMU pre-integration model for vehicle motion error parameter optimization to generate vehicle motion error optimization parameters; Step S34: reverse correction processing of the vehicle motion pose characteristic data is performed according to the vehicle motion error optimization parameters to generate corrected vehicle motion pose characteristic data; Step S35: laser mileage error analysis is performed based on the laser point cloud terrain monitoring data to generate laser mileage error parameters; Step S36: tight coupling analysis of the corrected vehicle motion pose and the laser mileage error is performed according to the corrected vehicle motion pose characteristic data and the laser mileage error parameters to generate corrected pose-mileage error tight coupling data; fusion weight analysis of the corrected pose-mileage error tight coupling data is performed by using a preset error state Kalman filtering algorithm to obtain fusion weight corrected pose-mileage error tight coupling data; factor graph optimization analysis of the vehicle positioning cumulative error is performed according to the fusion weight corrected pose-mileage error tight coupling data to generate vehicle positioning cumulative error optimization data; Step S37: vehicle front perception terrain map optimization processing of the vehicle positioning cumulative error is performed on the vehicle front perception terrain map data by using the vehicle positioning cumulative error optimization data to generate optimized vehicle front perception terrain map data; Step S4: visual feature matching and mapping processing is performed on the terrain image feature data by using the optimized vehicle front perception terrain map data to generate optimized vehicle front perception visual terrain map data; real-time terrain anomaly detection processing of the vehicle front perception is performed based on the optimized vehicle front perception visual terrain map data to generate vehicle front perception real-time terrain anomaly detection data; the vehicle front perception real-time terrain anomaly detection data is fed back to the terminal to perform rain working condition real-time terrain anomaly detection feedback operation. 2.The real-time terrain anomaly detection method based on multi-source data analysis of claim 1, wherein, Step S1 comprises the following steps: The multi-source monitoring sensor built in the vehicle comprises a laser radar device, a monitoring camera device, and an inertial measurement unit device; The laser point cloud terrain monitoring data is generated by using the laser radar device built in the vehicle to perform laser point cloud terrain monitoring on the rain scene monitoring area; The terrain image feature data is generated by using the monitoring camera device built in the vehicle to perform terrain image feature monitoring processing on the rain scene monitoring area; Collect vehicle motion pose characteristic data by using an inertial measurement unit device built in a vehicle. 3.The real-time terrain anomaly detection method based on multi-source data analysis of claim 1, wherein, Step S2 includes the following steps: Step S21: Perform point cloud motion distortion influence feature analysis under rain conditions according to vehicle motion acceleration, vehicle motion angular velocity, and vehicle motion speed of the vehicle motion pose characteristic data, and generate rain condition point cloud motion distortion influence feature data; Step S22: Perform covariance matrix analysis of laser point cloud error representation according to laser point cloud terrain monitoring data, and generate rain condition laser point cloud error representation covariance matrix; Step S23: Establish a rain condition point cloud calibration model according to the rain condition point cloud motion distortion influence feature data and the rain condition laser point cloud error representation covariance matrix, and perform terrain original point cloud filtering and distortion removal processing on the laser point cloud terrain monitoring data under rain conditions through the rain condition point cloud calibration model to obtain laser point cloud calibration terrain data; Step S24: Perform point cloud terrain coordinate conversion and solving processing based on the laser point cloud calibration terrain data, and generate point cloud terrain coordinate data; Step S25: Perform elevation and horizontal dimension reduction processing of the point cloud terrain on the point cloud terrain coordinate data respectively, and generate point cloud terrain elevation dimension reduction data and point cloud terrain horizontal dimension reduction data; Step S26: Perform vehicle motion pose error analysis according to the vehicle motion pose characteristic data, and generate vehicle motion pose error data; perform horizontal error and elevation error analysis of the motion pose on the point cloud terrain elevation dimension reduction data and the point cloud terrain horizontal dimension reduction data respectively according to the vehicle motion pose error data, and generate pose error point cloud terrain elevation dimension reduction data and pose error point cloud terrain horizontal dimension reduction data respectively; Step S27: Perform front-perception terrain map fusion processing of the pose error point cloud terrain elevation dimension reduction data and the pose error point cloud terrain horizontal dimension reduction data according to the multi-source error transmission mechanism, and generate front-perception terrain map data. 4.The real-time terrain anomaly detection method based on multi-source data analysis of claim 3, wherein, Step S22 includes the following steps: Analyze the laser beam uncertainty parameter under rain conditions and the terrain matching uncertainty parameter under rain conditions according to the laser point cloud terrain monitoring data; Perform covariance matrix analysis of laser point cloud error representation on the laser beam uncertainty parameter under rain conditions and the terrain matching uncertainty parameter under rain conditions according to the preset laser radar beam model combined with the likelihood function, and generate rain condition laser point cloud error representation covariance matrix. 5.The real-time terrain anomaly detection method based on multi-source data analysis of claim 4, wherein, The laser radar beam model is used to perform characteristic analysis of the laser radar emitted laser beam under rain conditions in the propagation and reflection process, and the likelihood function is used to perform matching degree analysis between the laser radar measurement and the real terrain under the terrain matching uncertainty parameter under rain conditions. 6.The real-time terrain anomaly detection method based on multi-source data analysis of claim 3, wherein, Step S24 includes the following steps: The real term and Gaussian error term of the laser point cloud calibration terrain data are analyzed by using a preset Gaussian model, and the linear combination of the laser radar point cloud coordinates and the laser radar measurement error term is analyzed by using an algorithm under a small angle approximation for the real term and Gaussian error term of the laser point cloud calibration terrain data, to obtain laser radar point cloud coordinate-error term linear feature data; Point cloud terrain coordinate conversion data is generated by performing point cloud terrain coordinate conversion processing on the laser radar point cloud coordinate-error term linear feature data, and point cloud terrain coordinate data is obtained by solving the point cloud terrain coordinate conversion data through a Jacobian matrix.

7. The method for real-time terrain anomaly detection based on multi-source data analysis according to claim 1, characterized in that, Step S35 includes the following steps: Key frame extraction is performed on the laser point cloud terrain monitoring data to obtain key frame laser point cloud terrain monitoring data; Adjacent frame laser point cloud change data is generated by performing adjacent frame laser point cloud change analysis on the key frame laser point cloud terrain monitoring data; Key frame positioning error data is obtained by performing key frame and corresponding local map positioning error analysis on the key frame laser point cloud terrain monitoring data and the vehicle front perception terrain map data; Laser mileage error analysis is performed through the adjacent frame laser point cloud change data and the key frame positioning error data to generate laser mileage error parameters. 8.The real-time terrain anomaly detection method based on multi-source data analysis of claim 1, wherein, Step S4 includes the following steps: Step S41: The terrain image feature data is transmitted to the optimized vehicle front perception terrain map data for visual feature matching and mapping processing to generate optimized vehicle front perception visual terrain map data; Step S42: Visual terrain type classification processing is performed on the optimized vehicle front perception visual terrain map data in a rainy weather scenario to generate visual terrain classification data; Step S43: A preset semantic segmentation model is used to perform terrain type attribute label identification on the visual terrain classification data to generate attribute visual terrain classification data; Step S44: The attribute visual terrain classification data is subjected to different attribute terrain visual adhesion coefficient mapping processing according to a pre-established terrain level priori database to generate terrain visual adhesion coefficients; Step S45: Vehicle longitudinal kinematics data and vehicle lateral kinematics data of vehicle motion pose characteristic data are extracted, and each type of terrain kinematics adhesion coefficient mapping processing is performed on the visual terrain classification data according to the vehicle longitudinal kinematics data and the vehicle lateral kinematics data to generate terrain kinematics adhesion coefficients; Step S46: Actual probability distribution analysis processing of terrain adhesion fusion is performed according to the terrain visual adhesion coefficients and the terrain kinematics adhesion coefficients to generate terrain adhesion fusion data; Step S47: Actual terrain adhesion probability distribution uncertainty self-supervised optimization processing is performed on the terrain visual adhesion coefficients and the terrain kinematics adhesion coefficients according to a preset self-supervised learning model to generate terrain adhesion self-supervised optimization data; Step S48: Real-time terrain anomaly detection processing of vehicle front perception is performed according to the terrain adhesion self-supervised optimization data to generate vehicle front perception real-time terrain anomaly detection data; Step S49: The vehicle front perception real-time terrain anomaly detection data is fed back to the terminal to perform real-time terrain anomaly detection feedback work in a rainy weather condition.

Images