Pareto optimal map matching method for autonomous vehicle positioning

Abstract

The invention discloses a Pareto optimal map matching method for autonomous vehicle positioning, and the method comprises the steps: obtaining sensor data of an autonomous vehicle, and building a brightness frame of the autonomous vehicle according to the sensor data and road boundary information obtained by a preset map; converting a laser radar point corresponding to the brightness frame into a UTM coordinate according to the current position and direction of the autonomous vehicle reported by the GNSS, calculating the conversion from the WGS coordinate to the UTM coordinate, obtaining a laser radar frame from the conversion from the UTM coordinate to the laser radar coordinate, obtaining a boundary point in the laser radar frame by adopting a spatial index according to the laser radar point, and obtaining the boundary point according to the boundary point. According to the method and the system, the boundary points are obtained, the 3D laser radar data are subjected to point cloud data conversion into the navigation coordinates according to the boundary points, so that the bright point cloud picture is generated, the point cloud map density is reduced, the calculation amount of a map matching algorithm is reduced, dead reckoning is combined with NDT map matching under the Pareto optimization algorithm, and the vehicle positioning precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of map matching, and in particular relates to a Pareto optimal map matching method for autonomous vehicle positioning. Background Art

[0002] In autonomous driving, to support positioning, a range of sensor modalities are used, including Global Navigation Satellite Systems (GNSS), Inertial Measurement Units (IMUs), cameras, LiDAR, and radar. Each sensor forms part of a rich stream, but often their performance is affected by weather conditions such as heavy rain, fog, or snow, which can reduce the quality and reliability of sensor data. For example, GNSS suffers from signal attenuation and multipath errors when operating in challenging environments such as urban canyons or forests; on the other hand, LiDAR systems suffer from beam scattering in rainy conditions, further leading to lower object detection accuracy. These environmental parameters introduce variations in sensor data, thereby increasing the computational burden of real-time processing while reducing the overall positioning accuracy.

[0003] To overcome these limitations of single-sensor systems, multi-sensor fusion has recently received great attention. In fact, it can meet the specific errors of each sensor, such as the NLOS problem of GNSS, the bias in IMU measurements, and the accumulated errors in visual or LiDAR odometry systems, which are usually integrated into an INS and fused with GNSS data through Kalman filtering to estimate the vehicle's position, velocity and heading. Under normal GNSS operating conditions, the integration works well; in cases where GNSS is denied or the signal is weak, positioning drifts due to the accumulation of inertial navigation system errors. Other sensors, such as odometers, or auxiliary GNSS signals (such as Doppler frequency) can be used to limit these errors. In cases where GNSS signals are unavailable or unreliable, high-definition (HD) maps have become an important asset for enhanced positioning. These maps are generated from dense point cloud data maps (DPCD-maps) generated by LiDAR and can provide very accurate global position information. On the other hand, the computational cost of real-time map matching is very high, thus posing a very serious challenge. Traditional positioning methods based on extended Kalman filtering often require complex linearization and simplification, which reduces flexibility and robustness in dynamic environments. Therefore, there is an urgent need to provide a Pareto optimal map matching method for autonomous vehicle positioning to solve the above-mentioned technical problems. Summary of the invention

[0004] In view of this, the present invention provides a Pareto optimal map matching method for autonomous vehicle positioning, proposes a new light map generation method, greatly reduces the point cloud map density, and reduces the computational complexity of the map matching algorithm, which is specifically implemented using the following technical solutions.

[0005] The present invention provides a Pareto optimal map matching method for autonomous vehicle positioning, comprising the following steps: Acquire sensor data of an autonomous vehicle, and create a brightness frame of the autonomous vehicle based on the sensor data and road boundary information obtained from a preset map, wherein the sensor data includes 3D lidar data and GNSS data, and the road boundary information is represented by a list of four discrete points of respective boundaries; Convert the laser radar point corresponding to the brightness frame into UTM coordinates according to the current position and direction of the autonomous vehicle reported by GNSS, wherein the nearest lateral position point is found from the four boundary lists and marked as the laser radar point, and the laser radar point is deleted from the four discrete point lists in the list; Calculate the conversion from WGS coordinates to UTM coordinates, and obtain the lidar frame from the conversion from UTM coordinates to lidar coordinates, where WGS coordinates are World Geodetic System coordinates and UTM coordinates are Universal Transverse Mercator coordinates; According to the laser radar points, the boundary points are obtained in the laser radar frame by using the spatial index, and the 3D laser radar data is converted into the navigation coordinates according to the boundary points to generate the bright spot cloud map.

[0006] The present invention provides a Pareto optimal map matching method for autonomous vehicle positioning, which obtains sensor data of the autonomous vehicle, creates a brightness frame of the autonomous vehicle according to the sensor data and road boundary information obtained from a preset map, converts a laser radar point corresponding to the brightness frame into a UTM coordinate according to the current position and direction of the autonomous vehicle reported by a GNSS, calculates the conversion from the WGS coordinate to the UTM coordinate, and obtains a laser radar frame from the conversion from the UTM coordinate to the laser radar coordinate, obtains boundary points in the laser radar frame using a spatial index according to the laser radar point, and converts point cloud data of 3D laser radar data into navigation coordinates according to the boundary points to generate a bright spot cloud map, thereby reducing the density of the point cloud map and the calculation amount of the map matching algorithm, combining dead reckoning with NDT map matching under a Pareto optimization algorithm, realizing the fusion of distance, speed and direction, without any linearization processing, and improving the vehicle positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0008] Figure 1 A flow chart of the Pareto optimal map matching method for autonomous vehicle positioning provided by the present invention. DETAILED DESCRIPTION

[0009] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0010] See also Figure 1 , a Pareto optimal map matching method for autonomous vehicle positioning, characterized in that it includes the following steps: S1: Acquire sensor data of an autonomous vehicle, and create a brightness frame of the autonomous vehicle according to the sensor data and road boundary information obtained from a preset map, wherein the sensor data includes 3D lidar data and GNSS data, and the road boundary information is represented by a list of four discrete points of each boundary; S2: converting the laser radar point corresponding to the brightness frame into UTM coordinates according to the current position and direction of the autonomous vehicle reported by GNSS, wherein the nearest lateral position point is found from the four boundary lists and marked as a laser radar point, and the laser radar point is deleted from the four discrete point lists in the list; S3: Calculate the conversion from WGS coordinates to UTM coordinates, and obtain the lidar frame from the conversion from UTM coordinates to lidar coordinates, wherein the WGS coordinates are World Geodetic System coordinates and the UTM coordinates are Universal Transverse Mercator coordinates; S4: Acquire boundary points in the laser radar frame using a spatial index according to the laser radar points, and convert point cloud data of the 3D laser radar data into navigation coordinates according to the boundary points to generate a bright spot cloud map.

[0011] In this embodiment, the light point cloud map generation technology using GNSS, INS and 3D LiDAR data is used to retain only basic features such as buildings and roads, while excluding seasonal variable elements. The light map is then combined with a dead reckoning model and a map matching algorithm based on normal distribution transformation (NDT). The Pareto optimization method is used to solve the problem of minimizing the mean and variance of the positioning error, replacing the traditional Kalman filter method, providing a robust solution to the multi-objective challenge of positioning accuracy and reliability. Based on the given rough data, the vegetation and road boundary information on the left and right sides of the ego vehicle can be segmented. Each class is constrained by boundaries, which are considered to be close to the road and depict the edge of the vegetation or other unwanted features. The leftmost and rightmost boundaries of the boundary area contain these features. The boundary information is given in the form of a list of four discrete points that define the respective boundaries. The LiDAR points are then converted to UTM coordinates (east and north) based on the current position and direction of the ego vehicle reported by the GNSS. For each translated LiDAR point, the nearest lateral position point is found from the four boundary lists. This area is the area surrounded by the two points on the left and right boundary tables. These LiDAR points are then marked for removal from the original list, resulting in a frame of “brighter” but relevant LiDAR points.

[0012] It is noteworthy that the prior information from the predefined map is used to segment the non-road area and vegetation, while the low-cost GNSS is used to reduce the point cloud data of the LiDAR frame. The brightness frame concept proposed in this invention is very effective in reducing the computational cost while maintaining the accuracy of the NDT-based map matching algorithm. It also implements the automatic recognition algorithm of roads and vegetation in order to segment the area of ​​interest of the NDT map matching algorithm. The transformation from World Geodetic System (WGS) coordinates to Universal Transverse Mercator (UTM) coordinates and then from UTM to LiDAR coordinates is pre-calculated, which will prevent the conversion calculation from being repeated in each iteration and greatly reduce the computational overhead. The algorithm saves these transformations in advance and avoids recalculating them for each LiDAR frame. In the context of autonomous vehicle navigation, localization needs to be accurate and reliable. Most traditional localization methods combine various sensor input data to continuously estimate the position and orientation of the vehicle. The most widely used is the dead reckoning model, which uses onboard sensors to measure wheel speed and yaw rate. This model calculates the position and heading in navigation coordinates and provides a basic reference for the data of other sensors such as LiDAR.

[0013] It should be understood that by acquiring sensor data of an autonomous vehicle, and creating a brightness frame of the autonomous vehicle based on the sensor data and road boundary information obtained from a preset map, converting the lidar points corresponding to the brightness frame into UTM coordinates based on the current position and direction of the autonomous vehicle reported by GNSS, calculating the conversion from WGS coordinates to UTM coordinates, and obtaining a lidar frame from the conversion from UTM coordinates to lidar coordinates, acquiring boundary points in the lidar frame using spatial indexes based on the lidar points, and converting point cloud data of 3D lidar data into navigation coordinates based on the boundary points to generate a bright spot cloud map, thereby reducing the density of the point cloud map and the amount of calculation of the map matching algorithm, combining dead reckoning with NDT map matching under the Pareto optimization algorithm, realizing the fusion of distance, speed and direction without any linearization processing, and improving the vehicle positioning accuracy.

[0014] Optionally, converting the 3D laser radar data into point cloud data into navigation coordinates according to the boundary points includes: Convert the point cloud data of the brightness frame into navigation coordinates and superimpose the converted point clouds together. The corresponding expression is: (1) in, are the coordinates of each point in the brightness frame, is the translation vector from the lidar frame to the navigation frame, is the rotation matrix between the lidar coordinates and the navigation coordinates, the rotation matrix The definition expression is: (2) in, , and is the heading, pitch and roll information, and ; After there is a point cloud in the navigation frame, the brightness frames are accumulated together to form a bright spot cloud map.

[0015] In this embodiment, based on each translated LiDAR point, the nearest lateral position point is found from the four boundary lists. This area is the area surrounded by two points on the left and right boundary tables, and then these LiDAR points are marked as deleted from the original list, thereby generating a LiDAR point frame. It is worth noting that prior information from a predefined map (Google Maps) is used to segment non-road areas and vegetation, while low-cost GNSS is used to reduce the point cloud data of the LiDAR frame. The light frame concept proposed by the present invention is very effective in reducing the computational cost while maintaining the accuracy of the NDT-based map matching algorithm. It also implements road and vegetation automatic recognition algorithms, which can be integrated into the frame to segment the area of ​​interest of the NDT map matching algorithm. Using the Applanix-LVX system with RTK correction installed on a high-end vehicle, the three-dimensional LiDAR coordinates are converted to navigation coordinates. This accurate conversion is used to convert the point cloud data of the brightness frame into navigation coordinates. These converted point clouds are stacked to create a bright spot cloud map.

[0016] Optionally, converting the laser radar point corresponding to the brightness frame into UTM coordinates according to the current position and direction of the autonomous vehicle reported by GNSS includes: The dead reckoning model is built by estimating the vehicle's position based on its current position, speed, and direction. The wheel speed and yaw rate of the onboard sensors are used to calculate the position in the navigation coordinates, as well as the vehicle's heading between the lidar coordinates and the navigation coordinates. hour along The expression for the dead reckoning estimate of the axis position is: (3) in, for The estimated position at the moment, is the time interval between two measurements, for Observe the speed at all times. yes Always observe the azimuth, The real position at the moment It is expressed as: (4) in, for The true speed of time, for The true azimuth at the moment, the wheel speed measured by the wheel speed sensor and direction , is the heading angle of the vehicle, is the heading angle, is the vehicle's sideslip angle, The expression is: (5) in, is the yaw rate, measured by and Given, the corresponding noise term and Destruction, the measured value follows the variance and The zero-mean Gaussian distribution of Given the dead reckoning data and the observed velocity and direction Calculate the eastward speed and northward speed, the corresponding expressions are: (6) (7) in, is the measured speed, Indicates the direction being measured.

[0017] In this embodiment, the dead reckoning model estimates the position of the vehicle based on the previous position, speed and direction of the vehicle to establish a dead reckoning model, and uses the wheel speed and yaw rate provided by the on-board sensors such as the inertial navigation system to calculate the position in the navigation coordinates and the heading of the vehicle between the LiDAR coordinates and the navigation coordinates. The yaw rate is due to the small lateral distance of the vehicle, which means that the sideslip angle is small and can be ignored, and can be approximated by the heading angle. The measured values ​​are given by and are corrupted by the corresponding noise terms and, which follow a zero-mean Gaussian distribution with variances and, respectively. In order to calculate the speed to the east and north, these components are first derived from the given dead reckoning data, and the speed to the east and north can be calculated using the observed speed and the observed vehicle direction. The tire slip ratio is not considered. If the slip ratio is not considered, there will be a speed error because the rotation speed of the tire is not exactly the same as the speed of the wheel, resulting in a cumulative position error. At the same time, the yaw angle and sideslip angle of the yaw rate gyro are not considered when solving the heading angle, which will lead to a heading error. Therefore, the growth of the position error is purely a matter of time. This again justifies the deployment of the map matching algorithm, as it reduces the accumulated error and develops a Pareto-based approach to estimate the error inherent in the dead reckoning model itself. In addition, the operating frequency of the dead reckoning model depends on only two sensors, the wheel speed and yaw rate sensors. It is set to 50 Hz, which is higher than the frequency of the map matching algorithm.

[0018] Optionally, the bias in the dead reckoning estimate is quantified by the difference between the estimated position and the true position, and the bias is calculated by evaluating the expected difference between the dead reckoning estimate and the true position. , the corresponding mathematical expression is: (8) The replacement expression is and for (9) Further simplified to: (10) Among them, the preset is zero at time The estimate is unbiased, with a preset direction angle Additive Gaussian white noise If the damage is done, the observed direction is expressed as: (11) in, Indicates that the variance The expected value of the cosine of the noise angle can be approximated by using the properties of the Gaussian distribution to approximate the expected value of the low noise variance as: (12) Substituting formula (12) into formula (10) to calculate the deviation, the result is: (13) Introducing a correction factor based on system characteristics or noise factors, the deviation is rewritten as: (14) in, The bias in the dead reckoning model is based on the expected value of the deviation between the estimated position and the true position, considered as an adjustment related to system characteristics or empirical calibration.

[0019] Optionally, the variance of the estimation error in the dead reckoning model can be determined and evaluated by analyzing the propagation of velocity and orientation measurement noise within the system, where the variation in position estimate is affected by the measured velocity. and direction The variance of the dead reckoning position estimate can be expressed as: (15) Using the characteristics of Gaussian noise, the variance of the product of velocity and direction angle cosine is: (16) in, represents the expected value, Approximately , the subsequent approximation is applicable and uses: (17) Here, variance is incorporated into the fusion process and the dead reckoning estimates are weighted relative to the range measurements.

[0020] In this embodiment, the bias in the dead reckoning estimate is quantified by the difference between the estimated position and the true position. To calculate the bias, the expected difference between the dead reckoning estimate and the true position is evaluated. The bias in the dead reckoning model is based on the expected value of the difference between the estimated position and the true position, which is a function of the nonlinear effect of noise in the heading angle on the cosine function. This bias term depends on the speed of the vehicle, the time interval, the heading angle itself, and the variation of noise in the heading measurement. The error state constrained Pareto method aims to determine the ideal fusion parameters and strike a balance between minimizing the position and heading errors in the estimate and managing the noise variance. The purpose of this method is to reduce the mean square error of the estimate while keeping the estimation error within the allowable range and taking into account the uncertainty caused by sensor noise.

[0021] It should be noted that the goal of the error state constrained Pareto method is to estimate the position and heading errors in the dead reckoning model and obtain the error measurements of the northeast position, velocity, and heading. This method optimizes the fusion parameters to minimize the estimation error and noise variance while ensuring that the error state constraints are met, and to maintain a balance between the two competing goals. The optimization framework is used to select fusion parameters to reduce the positioning and heading errors while controlling the overall noise variance. The error constraint makes the estimation error not exceed a certain maximum value, and the sum of the weights of the sensor fusion does not exceed 1. A balance is achieved between reducing the mean square error (MSE) of the estimate and minimizing the uncertainty caused by sensor noise, thereby ensuring the robust positioning of the autonomous vehicle.

[0022] Optionally, calculate the conversion from WGS coordinates to UTM coordinates, and obtain the lidar frame from the conversion from UTM coordinates to lidar coordinates, including: The normal distribution transform (NDT) algorithm is used to convert the lidar point cloud into navigation coordinates to calculate the reference map, align the lidar with the navigation frame, and provide appropriate map matching for positioning. Formula (18) represents the lidar frame To Navigation Frame Transformation: (18) in, Indicates a point in the navigation frame The coordinates of represents the corresponding point in the laser radar coordinate system, represents the rotation matrix of the lidar relative to the navigation frame, Vector representing the translation between the LiDAR and navigation coordinate systems, rotation matrix Defined as: (19) in, , and Indicates the roll angle, pitch angle and heading angle, , , , and , .

[0023] In this embodiment, an Applanix-LVX system with RTK correction installed on a high-end vehicle is used to achieve the conversion of 3D LiDAR coordinates to navigation coordinates. This precise transformation is used to convert the point cloud data of the brightness frame into navigation coordinates. These converted point clouds are superimposed together to form a bright spot cloud map. Since the vegetation that changes with the seasons has been deleted, on the one hand, the density of the map can be reduced, and on the other hand, since the vegetation is changing, the map may not be updated. For the former, the advantage is that based on the light frame with reduced density, the computational load of the map matching algorithm can be significantly reduced. For the latter, since there is no vegetation that changes in both the map and the frame, the positioning accuracy can be improved.

[0024] Optionally, convert each lidar scan point Convert to navigation frame and aggregate to form point cloud : (20) in, It is an aggregation to generate a point cloud map The total number of points, for the point cloud Optimizations include: Vegetation removal: filter out points corresponding to vegetation; Map downsampling: Downsample the map and keep the points that are most important for accurate positioning.

[0025] In this embodiment, the normal distribution transform (NDT) algorithm utilizes a reference map to calculate the reference map by converting the lidar point cloud into navigation coordinates. This transformation aligns the lidar with the navigation frame, thereby providing appropriate map matching for positioning.

[0026] Optionally, a normal distribution transform (NDT) algorithm is used to convert the lidar point cloud into navigation coordinates to calculate a reference map, including: Use the given dead reckoning model to calculate the vehicle’s The estimated position at time is: (twenty one) in, express The estimated easting and northing position at the moment, express The estimated speed at each moment, express Estimated heading at the moment, express Moment and The time interval between moments is transformed by finding the best alignment between the current lidar scan and the reference image , the transformation expression is: (twenty two) in, is the rotation angle, , are the eastward and northward translation components respectively. The NDT map matching process is to determine the best transformation , which will reduce the current scanning point and reference image The problem of minimizing the difference between is described as: (twenty three) in, Indicates the number of points in the current scan, determining the best transformed Used to correct the dead reckoning position estimate, after implementing NDT-based map matching, the corrected position estimate can be expressed as: (twenty four) A weighted fusion algorithm is used to fuse the dead reckoning and NDT estimates. Fusion position estimation at time It is expressed as: (25) in, and Represents the fusion weight, satisfying the condition , the difference of the merged position estimate is expressed as: (26) in, Indicates that the vehicle is The actual location at the moment.

[0027] In this embodiment, in order to optimize the map and enable practical positioning, the following operations are performed: (1) Vegetation removal: Filter out points corresponding to vegetation. This will be an important stage in reducing map density and maintaining uniformity in different seasons. Eliminating seasonal vegetation enhances the reliability of map positioning under a wide range of environmental conditions; (2) Map downsampling: Downsampling the map to retain only the most important points for accurate positioning is another way to further reduce computational complexity. Downsampling (downsampling) technology ensures that the map retains key features and provides effective navigation while still maintaining computational efficiency. A lightweight point cloud map is generated. By minimizing the presence of variable features such as seasonal vegetation, the stability of the reference map over time can be maintained.

[0028] Optionally, the uncertainty estimate in the NDT correction estimate is combined with the uncertainty estimate in the dead reckoning estimate, given the estimated error exist The covariance at time can be propagated at time k+1 using formula (27): (27) in, is the covariance estimated based on NDT map matching, is the covariance of the dead reckoning estimate, is the cross covariance between the NDT-based estimate and the dead reckoning estimate, and is calculated as: (28) in, represents the position estimate after dead reckoning, is the expected value of the dead reckoning estimate, represents the expected value of the NDT corrected estimate, is the expected value of the NDT corrected estimate, represents the expectation operator that computes the mean of a random variable, is the cross covariance of the two estimated errors to reflect the linear relationship between them.

[0029] In this embodiment, the transformed point cloud is accumulated to generate a bright spot cloud map as a reference for the NDT algorithm, which is optimized by removing vegetation to reduce density, thereby improving computational efficiency. Since these steps are to create a map, where the LiDAR scan is matched, high-precision position correction and positioning are allowed through the NDT-based map matching process. The Normal Distribution Transformation (NDT) algorithm aligns the current LiDAR scan with the reference map to find the optimal transformation that minimizes the difference between the transformed scan points and the reference map points.

[0030] Optionally, obtain the error of the fused position estimate, Error in time The norm is less than or equal to The error at the moment, the system state transfer matrix The spectral radius is less than 1: and (29) in, is the spectral radius of the system state transfer matrix, and Represents the error between the fusion estimate and the true position, ensuring This means that the estimation error will not diverge over time.

[0031] In this example, how the uncertainty in one estimate interacts with the error in another estimate, cross-covariance is crucial to understand and mitigate these compound uncertainties that combine multiple estimation techniques. Stability is essential for the continued performance of autonomous vehicles, especially in dynamic and uncertain environments. Notably, combining the map generation process with NDt-based map matching can improve positioning accuracy by retaining the reference map in an accurate and computational manner. By reducing the density of the map and minimizing unnecessary elements, the computational load is reduced, allowing real-time processing, which is the fusion between NDT-corrected estimates and dead reckoning, will produce more robust and stable position estimates, which are indispensable in autonomous navigation and vehicle localization. In fact, the extended accuracy of this approach can not only directly contribute to improving the accuracy of the position estimate, but also provide additional benefits to the reliability of the entire autonomous setting by making it more robust to environmental changes or sensor noise. Therefore, this joint approach provides a wide range of solutions to positioning challenges in autonomous vehicle applications, greatly improving other traditional methods of single sensors or single methods.

[0032] Specifically, multiple data acquisition systems were used to capture point cloud data for high-precision mapping and localization. The HDL-32E 3D LiDAR sensor mounted on the vehicle provided the required high-frequency point cloud data at 20 Hz, while the ROS-developed algorithms processed this data in real time on an industrial-grade PC. The vehicle pose updates were localized in response at 5Hz, while the ground truth was determined by the integrated GNSS and IMU system, Planar LVX. For large-scale data collection, the RIEGL VMX-2HA mobile laser scanning system was used, which provided very accurate 3D data with an accuracy of 8 mm and a precision of 5 mm, and could obtain up to 1.1 million measurements per second due to its dual scanning mechanism. Fixed GPS stations also allowed the establishment of control points to enhance georeferencing, allowing the data to be precisely aligned with real-world coordinates. The MLS system reached speeds of up to 60 km / h during operation, achieving a point density of approximately 7500-8500 points / m2. Due to the very high point density, even the finest details can be represented in the data, while also obtaining the granularity required for mapping and subsequent analysis. Four different independent MLS datasets were used and provided in LAS file format. Dataset 1 is a 30-meter-long urban road corridor segment. The number of points in the study area is 1,827,963, with a density of 8,500 points / m²; Dataset 2 is a length of 176.5 m for other road segments, with 9,850,578 points respectively. These datasets span different sections of the highway, each with different positioning and mapping challenges. In these datasets, the high level of point density allows thorough and detailed scanning of road markings, curbs, and even building facades, which can easily capture details and provide robust analysis for applications in autonomous navigation.

[0033] Specifically, real-time performance is the most critical issue in autonomous vehicle positioning. The rotation rate of the Velodyne LiDAR and the update frequency of the algorithm ensure that the vehicle's posture is generated at a frequency of 5Hz, thereby achieving continuous and responsive positioning. The use of industrial PCs to handle the computational workload brought by real-time processing of large data sets ensures the accuracy and speed of the algorithm. A powerful data acquisition and positioning system is proposed, which integrates high-performance LiDAR, precision GNSS / IMU and a powerful data processing platform. These data sets are very valuable for further verification and continuous improvement of positioning algorithms in challenging real-world scenarios.

[0034] The tests were conducted on a loop of approximately 2.7 km, with this part of the 0.5 km section of the track being specifically chosen for analysis as it faces a complex environment with a highly complex built environment and diverse vegetation. The performance of the brightness map was considered from the identification of 5 invariant features from 1D, 2D and 3D structures. The impact of the environment on the accuracy of the algorithm was analyzed in detail, considering Areas 1 and 2, where there are buildings and vegetation. The tests were conducted under heavy rain and fog conditions, which are common in the region. The harsh conditions introduced a lot of variability in the LiDAR point cloud, mainly due to the scattering effect caused by raindrops and the reduced visibility in foggy conditions. The density of raindrops as well as the diffusion effect of light in fog require frequent updates and adaptations of the map to achieve accurate real-time positioning. In fact, the ability to consider and adapt to such environmental changes is an important consideration when evaluating algorithms.

[0035] To simulate real driving conditions, a shuttle bus was driven along a ring road in a normal pattern of acceleration, deceleration and steering. During the test, the longitudinal speed was kept between 0 and 25 km / h by following the speed limit of the ring road. Therefore, the focus of the test was on the robustness of the map matching algorithm in adverse weather conditions, especially rain, which affects the data acquisition of the lidar sensor and the processing required to correctly localize the vehicle. The algorithm was initially executed in clear weather to provide a baseline for the actual performance of localization, and this result has been used as a reference for further evaluation. Subsequent experiments have been completed to test the combination of rainfall with lidar scanning and feature extraction. The test also explored how weather affects the ability to distinguish between static features (such as buildings) and dynamic environmental features (such as vegetation), given the nature of their appearance when it rains. This part of the evaluation is crucial to determine whether economically efficient navigation can be relied upon in changing weather conditions, especially in areas where heavy rainfall is expected.

[0036] The proposed brightness map generation was tested under adverse weather conditions, when the weather was clear and when it rained heavily. Two types of point cloud maps were created: a dense map that includes data from all available LiDAR frames and a lightmap that retains only invariant features such as buildings and road surfaces, while transient objects like vegetation are filtered out. The dense map contains all the information from the LiDAR points and provides a complete representation of the environment. It is worth noting that the lightmap retains less point density because sensitive points due to weather conditions are removed. Vegetation, characterized by areas 1 and 2, shows a greater reduction in point density in the brightness map compared to the dense map, which is particularly welcome during heavy rain, as this condition tends to make the LiDAR data worse due to scattering and reflection effects. This is because the brightness map, by removing these weather-sensitive points, presents fewer artefacts and provides a clearer view of the campus infrastructure.

[0037] As a result, the size of the point cloud map is reduced from 118 MB for the dense map to 48.3 MB for the light map, which indicates a 59% reduction in density. In addition, although vegetation changes require updating the dense map at a seasonal level, the coarse positioning light map generated based on a low-cost GNSS receiver is robust to seasonal and weather changes. This allows filtering out different vegetation points, making it denser in the map and reducing the update frequency, making it more efficient, reliable and robust to use this map matching algorithm in very adverse weather conditions (such as heavy rain).

[0038] Specifically, dense LiDAR frames are processed to effectively filter out vegetation points, thereby reducing the point cloud density. This filtering step is critical because during heavy rain, the ability of LiDAR to capture accurate data is severely affected. LiDAR scans acquired under heavy rain conditions appear relatively sparse, darker, and have an increased presence of noise points compared to LiDAR scans collected during clear weather. It is worth noting that these noise points are mainly generated by raindrops falling off the laser beam and only account for 1.22% of the total points received per scan, which is a negligible amount. Under heavy rain conditions, the number of measured points decreases, the intensity values ​​decrease, and some noise points are randomly present in the scan due to laser reflections from raindrops. Table 1 summarizes the statistics for three additional rainfall conditions in the environment, indicating that on average, 23% of the points are lost in a scan cycle. However, even after losing these points, the impact of noise points due to reflected raindrops is very small, which again strengthens the strength of the filtering method under weather conditions.

[0039] Table 1 Evaluation of rainfall impact

[0040] The reduction in the number of reception points affects not only the number of features to be extracted that are relevant for a good localization, but also their quality. The accumulated ground reflectance features for about 200 scans during the localization test. In contrast, it can be observed that some areas do not have any ground reflectance features during rainfall, while others show no change. This is due to the different permeability of surface materials, which is why some areas, such as grass on the roadside, which easily allows groundwater to enter the soil, reflect less changes during rainfall. Table 2 gives the number of features corresponding to each scene. As the rainfall intensity increases, the number of observable ground objects decreases significantly. However, the features that are less affected by rainfall change very little. Although falling raindrops can absorb some lidar points, it is unlikely that they will absorb all the points that hit the surface.

[0041] When responding to current fair weather conditions, the LL method has the shortest processing time, averaging 107.5 ms, while the LD and DD methods have higher processing times, 136.1 ms and 243.8 ms, respectively. The processing time of all methods increases from light to heavy rain, with the DD method reaching a critical peak of 284.7 ms during heavy rain. This behavior changes beneficially when the integrated Pareto method is used: efficiency is greatly improved across methods and conditions. This is particularly significant when considering the DD method for heavy rain, where the processing time is reduced from 284.7 ms to 104.4 ms. The optimization highlights the efficiency of the Pareto method in finding the right balance between computational efficiency and accuracy; therefore, it may be applicable to real-time map matching under adverse weather conditions.

[0042] Table 2 Characteristic observations under different weather conditions

[0043] After preprocessing, an integrated localization algorithm was implemented using the light map and dense map generated under heavy rain conditions, denoting the brightness frame and bright point cloud map as LL, the brightness frame and dense point cloud map as LD, and the dense frame and dense point cloud map as DD. The brightness frame image can be accurately registered on the brightness frame image and on the dense map, similar to the registration of dense frame images on dense maps. In response to the current fair weather conditions, the LL method has the shortest processing time, averaging 107.5 ms, while the LD and DD methods have higher processing times, 136.1 ms and 243.8 ms, respectively. The processing time of all methods increases from light rain to heavy rain, with the DD method reaching a critical peak of 284.7 ms during heavy rain. This behavior changes beneficially when the integrated Pareto method is used: the efficiency is greatly improved across methods and conditions. When the DD method is considered for heavy rain, the processing time for heavy rain is reduced from 284.7 ms to 104.4 ms. The optimization highlights the efficiency of the Pareto method in finding the right balance between computational efficiency and accuracy; thus, it may be applicable for real-time map matching in adverse weather conditions.

[0044] Table 3 summarizes the comparison of processing time under different environmental conditions and map matching methods, highlighting the efficiency of individual and combined Pareto methods. The data shows that the brightness frame and brightness map (LL) method has the shortest processing time in good weather conditions, with an average of 107.5 ms and an RMS of 32.1 ms. In contrast, the brightness frame and dense map (LD) method has an average processing time of 136.1 ms and an RMS of 47.4 ms, while the dense frame and dense map (DD) method has the longest processing time, with an average processing time of 243.8 ms and an RMS of 65.3 ms. When the weather gets worse, the processing time of each method increases. During heavy rain, the LL method has an average of 125.2 ms with a standard deviation of 37.39 ms. On the other hand, the average time for the LD and DD methods under the same weather conditions is 155.5 ms and 284.7 ms, and the RMS value is 54.3 ms and 76.2 ms, respectively. This increase is due to the increased complexity and noise brought by bad weather, which will eventually affect the frame quality and map matching accuracy.

[0045] The processing time of the integrated Pareto method was reduced under all conditions and methods, with the average time for the LL method under heavy rain conditions reduced from 125.2 ms to 70.8 ms, the LD method reduced from 155.5 ms to 81.5 ms, and the DD method reduced from 284.7 ms to 104.4 ms. The same pattern was maintained regardless of the weather changes. For example, in good weather conditions, it can be reduced to 60.8 ms using the LL method, 71.5 ms using the LD method, and 89.3 ms using the DD method. This is not only accurate, but also proves that the method used in this paper has high processing efficiency. Overall, the above data shows that the brightness frame reduces the calculation time by more than 43.4%. It also shows that the RMS value of both the LL and LD methods has a slight reduction. Although the utilization of the light map (brightness map) does not significantly reduce the processing time, it does effectively reduce the map size. At present, the method known as the integrated Pareto is effective in optimizing real-time map matching systems, especially in difficult conditions, because it balances computational performance and accuracy.

[0046] Table 3 Comparison of processing time

[0047] Table 4 shows the comparison of localization estimation errors for different settings in the map matching algorithm under different weather conditions, using three different methods: (1) light frame and light map (LL), (2) light frame and dense map (LD), and (3) dense frame and dense map (DD). For this comparison, the error metrics will include AME and RMSE. Since the localization performance is evaluated as a balance under different environmental conditions, the integrated Pareto method is applied to optimize the trade-off between these error metrics.

[0048] Table 4 Comparison of positioning estimation errors

[0049] The dead reckoning model is combined with an NDT-based map matching algorithm to estimate the position, so that for a given time step , the error can be Defined as the estimated position used by dead reckoning and NDT and estimated position used by NDT The actual position of the vehicle The difference between: (30) in, is the time step The estimated position of the vehicle in the east and north directions at time is the true position of the vehicle in the east and north directions. In this method, the dead reckoning estimate and the NDT-based map matching position are combined to estimate and NDT-based map locations and fusion weight and , so that: (31) in, is the dead reckoning position estimate, given by

[0050] Calculated; yes The speed of is the time interval between time steps, yes The estimated heading angle at It is an NDT-based position estimate obtained by matching the LiDAR scan to a reference (ground) map, using the ground truth east and north position of the Applanix LVX system and ,The results show that the LiDAR frame achieves a localization accuracy comparable to that of the original dense point cloud.

[0051] As mentioned above, a comprehensive analysis of the error performance in various sensor fusion methods has been conducted, and clear patterns of error distribution are observed in the odometry, dead reckoning, and proposed Pareto sensor fusion methods. In the LL case, the proposed Pareto sensor fusion has a lower cumulative error distribution compared to odometry and dead reckoning, indicating that this method has better performance in minimizing the absolute position error. For the LD scenario, although the error level increases compared to LL, the error level in the empirical cumulative distribution function (CDF) of the simulated error data is also more favorable for the proposed Pareto sensor fusion. With further reduction in the error rate, the DD method extends the advantage of the proposed Pareto sensor fusion, giving it the lowest cumulative error among all methods. Further confirmation of these observations is obtained from the values ​​of the root mean square error, as the proposed Pareto sensor fusion produces the smallest RMSE value in each case, which confirms its ability to produce accurate and reliable position estimates. These results once again strengthen the strength of the proposed Pareto sensor fusion and can effectively improve the error performance of different magnitudes of errors and complexities in sensor data.

[0052] In good weather conditions, the localization error is lowest for brightness frames vs. brightness maps, so this setting provides a fair baseline for position accuracy: it balances computational efficiency with reduced frame and map density to avoid overfitting. However, the dense setting, which provides detailed environmental features, does not provide a significant improvement in accuracy under good conditions. As weather conditions worsen, localization performance deteriorates for all configurations. Moderate and heavy rain lead to a large increase in position error, with more significant error increases shown in the denser configurations (brightness frames vs. dense maps and dense frames vs. dense maps). This is likely because the lidar sensor interacts more with rain, i.e., transient environmental features such as raindrops and wet surfaces introduce noise.

[0053] In light rain conditions, both AME and RMSE increase by about 50% due to the adverse effects of bad weather on positioning accuracy. Despite these increases, the Pareto method remains resilient, limiting the errors to approximately 0.0161 meters for LL and LD and 0.0168 meters for DD. This verifies the reliability of the method in a dynamic environment. In moderate rain conditions, the performance gap widens, with AME values ​​ranging from 0.0498 to 0.0515 meters and RMSE values ​​ranging from 0.0474 to 0.0485 meters. The Pareto methods continue to show superior performance, especially the DD method, which benefits the most from the optimization and has the lowest error of 0.0278 meters.

[0054] It can be seen that the most difficult conditions in which heavy rain occurs are characterized by a sharp increase in AME and RMSE, with AME varying between 0.1002 and 0.1022 m and RMSE being slightly lower, between 0.0934 and 0.0965 m. Under these poor conditions, the comprehensive Pareto method shows a significant reduction in the range of error increase within fairly low values, between 0.0564 and 0.0566 meters. This shows that the Pareto method provides stable and more accurate positioning in harsh environmental conditions compared to the other methods. Therefore, in harsh weather conditions, the density of frames and maps does not affect the positioning accuracy as much as the optimization method used, so robust optimization techniques in harsh environments are very important.

[0055] As expected from the relationship between the two metrics, RMSE is slightly smaller in all conditions. Due to the large errors, RMSE describes the consistency of the error values, while AME provides a broader view of the error size over time. However, the most important conclusion is that the integrated Pareto method consistently outperforms both AME and RMSE, thus emphasizing the efficiency of this multi-objective optimization strategy in minimizing localization errors.

[0056] Severe and adverse weather conditions, especially rain, have a more profound impact on the respective methods and metrics as it leads to higher positioning errors. However, the integrated Pareto method leads to a slightly slower increase in the error rate, which may indicate its robustness to changes in environmental conditions. This may be explained by the ability to simultaneously optimize multiple objectives, such as accuracy and variance. In addition, different mapping methods, including brightness frame and brightness map (LL), brightness frame and dense map (LD), dense frame and dense map (DD), show little difference in clear weather conditions, but show higher degradation in rainy conditions. Compared with LL and DD, the LD method has slightly lower AME and RMSE values, indicating that the combination of brightness frame and dense map achieves an effective balance between computational efficiency and information richness, especially in light to moderate rain scenarios.

[0057] Thus, the comparative analysis confirms the effectiveness of the integrated Pareto approach in reducing localization errors across different methods and under different environmental conditions. While AME and RMSE remain useful metrics, the Pareto optimization approach performs better across a wide range of conditions, with lower errors and significantly better robustness to adverse weather conditions. Considering that reliability and adaptability are two of the most important issues, this highlights the potential of the integrated Pareto approach as a real-world autonomous vehicle localization solution. Of course, further refinement can be achieved through more central methods and applications in more complex environments, such as canyons or dynamic obstacles, which could partially extend its practical applicability.

[0058] The brightness frames with brightness maps are fairly stable in light rain conditions, with low localization errors; even in moderate and heavy rain conditions, this configuration suffers some accuracy loss. Still, it performs better than denser ones that are resistant to ambient noise. While dense maps may provide benefits in ideal conditions, more lightweight approaches (especially brightness frames with brightness maps) may provide more robust localization performance in adverse weather conditions (such as heavy rain), as noise can compromise the integrity of dense point clouds.

[0059] The east and north position errors of the three selected frame and map density configurations: Light frame with Light map (LL), Lightframe with Dense map (LD), and Dense frame with Dense map (DD). The results are given for all four different weather conditions: sunny, light rain, moderate rain, and heavy rain. The change in the configuration's east position error over time reflects the fluctuation of positioning accuracy when the environment changes. (1) LL (Light Frame with Light Map) configuration, where the LL (Light Frame with Light Map) configuration performs stably under good weather conditions, with an east position error generally less than 0.5 meters. It performs roughly the same in light rain weather. However, in light rain and rainy conditions, its east position error begins to deteriorate, sometimes exceeding 0.75 meters. Although the LL configuration is robust under benign conditions, its sparse map cannot capture and compensate for dynamic environmental changes (such as raindrops and wet surfaces), resulting in large position accuracy deviations. (2) LD (LightFrame with Dense Map) configuration, where the LD configuration has a slightly higher error in adverse weather conditions. The performance is relatively stable on sunny days, with a position error of around -0.25~0.25 m. Light and heavy rain cause large fluctuations in power and position errors. In most heavy rain conditions, the positioning error exceeds 0.5m. The increased map accuracy captures transient features such as raindrops and reflections that the positioning algorithm cannot evenly filter out, thereby reducing accuracy. (3) DD (Dense Frame with Dense Map) configuration, where the maximum position error values ​​under different weather conditions are the DD configuration. In good weather conditions, the errors are very close to the values ​​of the other two configurations, mostly less than 0.5 meters. However, as the rain starts to get heavier, these position errors become unstable, with most exceeding 0.75 meters and approaching 1 meter in heavy rain. The combination of dense frame and dense map greatly amplifies environmental noise and captures too many transient or irrelevant features, such as artifacts caused by rain. This will lead to overfitting and severe loss of positioning accuracy.

[0060] The comparative analysis shows that the LL configuration has the most stable performance, as the position error is constant and lower compared to other configurations, whether it is sunny or light rain. This means that the frame and map density in LL optimally balance the representation of the environment to resist noise, and is therefore the most reliable for general use, especially in adverse conditions. In contrast, the LD and DD configurations can provide finer details of environmental identification, but are more susceptible to environmental noise, especially when it is raining heavily outside. In the dense configuration, redundant data points are introduced, adding a lot of noise and reducing the positioning accuracy, which means that too much detail in the environment model is negative in dynamic and noisy conditions, because the positioning system cannot filter out useful information from the noise. It can be seen that LL has the best performance in any weather conditions, especially in highly disturbing conditions such as heavy rain. Although the denser LD and DD can capture more details, they are also more susceptible to noise, which can lead to significant errors in weather degradation. The frame and map accuracy are appropriately selected to achieve the best positioning accuracy, at least in the presence of environmental noise, especially in highly dynamic and challenging outdoor conditions.

[0061] By comparing the different analyses, it can be noted for all configurations that rain significantly increases the localization errors, where the severity of these errors increases with the intensity of rain. However, the most consistent performance is provided by the LL configuration, as the error deviations are smaller in both sunny and rainy conditions. The lightweight frame and map settings are more effective in handling noisy and dynamic environments, without capturing irrelevant details that interfere with localization accuracy. However, in heavy rain conditions, the LD setting is more sensitive to rain, while the dense map provides more details of the environment. It may be useful in ideal conditions, but it introduces more noise in bad weather, leading to larger localization errors. This leads to the highest position error in all weather types in the DD configuration, which uses both dense frames and dense maps. Such a configuration, especially in rainy conditions, captures too much noise, which can prevent the system from maintaining a good localization, making it the least suitable configuration for use in challenging environments. Therefore, the LL configuration is the most robust and reliable in localization performance under different weather conditions, especially in rainy conditions. The denser configurations, whether involving dense maps or frames, or both, are more prone to noise and position errors, especially in rainy environmental conditions. These results bring back to the question of optimizing the localization system by balancing environmental detail and noise immunity, so the brightness frame and map settings may perform better when dealing with dynamic and challenging outdoor conditions, which are often accompanied by rain or other types of environmental interference.

[0062] It should be noted that these results provide important insights into the performance and optimization possibilities of sensor fusion methods in terms of processing time and position estimation errors. A very effective strategy to remove vegetation from LiDAR point clouds to reduce point cloud density and improve computational efficiency is this reduction, especially for light frame configurations, significantly reduces the processing load that the map matching algorithm needs to perform. Although vegetation removes some low-quality features, it does not significantly affect the localization accuracy because structural type features (such as buildings) dominate the map matching process. These features are more stable and invariant, therefore, they provide a more reliable reference for localization.

[0063] Computational results show that using lighter maps alone provides limited benefit in terms of increased processing time, while merging it with light frame generation can reduce computational requirements and appear to shrink further in severe weather conditions. This is evident from the examples that have been demonstrated using moderate and heavy rain, where processing times are significantly lower than those using dense frames and maps. As autonomous vehicle systems further evolve, computational complexity increases. When sensor fusion systems will start to integrate many features and functions, the optimization of each subsystem will become more important. By simplifying the map to reduce the computational load, in addition to enhancing real-time processing capabilities, the proposed method also opens new avenues for deploying complex localization and map matching algorithms in challenging dynamic environments.

[0064] Localization in rain-affected environments is one of the most challenging tasks for any autonomous driving localization system, given the rain-induced variability in sensor data. LiDAR point cloud data exhibits high variability in rain conditions, as each beam scatters and attenuates the signal. These effects degrade the accuracy of map matching algorithms, especially those that rely on dense point clouds representing transient environmental features such as vegetation. These features like vegetation are very sensitive to rain, thus increasing the variability of point-to-point localization performance during moderate and heavy rain. The proposed Pareto sensor fusion approach achieves optimal map generation by judiciously excluding these rain-sensitive features. As a result, the reference map will be very stable and reliable, as it filters out points associated with vegetation, which tend to fluctuate continuously with weather conditions and seasons. This reduction in point cloud density will increase the efficiency of the process and reduce the adverse effects of rain on LiDAR-based map matching.

[0065] By integrating brightness frames and lightmaps to make the system more robust in rain-affected environments, this setup is well balanced in light or moderate rainfall as it minimizes the computational load without compromising localization accuracy. When the rain is heavy, the additional contribution of temporal removal of the point cloud becomes more important. While dense point clouds can have higher spatial resolution, the increased distortion of the LiDAR signals makes them more susceptible to noise and errors in heavy rainfall conditions. The presented lightweight map will be more suitable for these conditions and therefore focuses on keeping features that are more stable and not affected by rain, such as buildings and road boundaries, which can serve as reliable markers for localization.

[0066] Therefore, rain-affected environments need to be further processed in a more automated and adaptive manner. In practice, this will involve manually removing rain-sensitive features such as vegetation through tools combined with coarse GNSS data. Automating this using machine learning and computer vision will greatly improve accuracy and efficiency. For example, deep learning models can be used to identify and exclude real-time elements in the lidar point cloud that are affected by rain. This will provide more consistent lidar performance under different weather conditions without the need for human intervention. Such adaptive filtering can also further enhance the system's ability to maintain high positioning accuracy under heavy rainfall conditions, thereby improving the reliability of autonomous vehicle navigation in rain-affected environments. In addition, integrating data from weather forecasts or rainfall sensors into the positioning framework can dynamically adjust the filtering process of the point cloud according to real-time weather conditions. Although the proposed Pareto sensor fusion method has greatly improved the positioning performance under rain-affected conditions, there is still a lot of room for optimization. Automating the process of removing rain-sensitive features and adaptive filtering techniques using real-time environmental data are very promising for further enhancing the robustness and accuracy of future autonomous positioning systems.

[0067] The result is stronger if GNSS data is also considered in map matching in good signal conditions to provide large-scale resilience to dead reckoning errors that grow larger over time. This hybrid approach, which fuses GNSS with lidar and inertial sensors in a consensus framework, has the potential to further extend the system’s operating range by improving its adaptability in real-world conditions. The Pareto-optimized localization framework optimally balances accuracy and computational efficiency in different conditions. This multi-objective approach excels at providing high adaptability and accuracy, especially in complex urban environments. Combining Pareto optimization with lightweight maps and sensor fusion enhances the system’s resilience and accuracy, making it suitable for real-time, weather-variable localization for autonomous vehicle applications. The Pareto-optimized localization approach shows large improvements in all metrics, with a 40-45% reduction in RMSE (from 0.0965m to 0.0566m in heavy rain), significantly improving localization flexibility and accuracy over traditional methods. By using a brightness map that retains only essential invariant features, the approach reduces point cloud density by 59%, reducing computational load and improving performance in adverse weather conditions. At the same time, the Pareto optimized brightness frame and brightness map (LL) method reduced the processing time from 107.5 ms to 60.8 ms under sunny conditions and from 284.7 ms to 104.4 ms under heavy rain conditions, proving the efficiency and accuracy of the method. The real-time application of the Pareto method in dynamic and climate change environments is effective, while pursuing higher positioning accuracy and higher processing efficiency.

[0068] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limiting, and thus other examples of the exemplary embodiments may have different values.

[0069] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0070] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A Pareto optimal map matching method for autonomous vehicle positioning, characterized in that: The following steps are involved: Acquire sensor data of an autonomous vehicle, and create a brightness frame of the autonomous vehicle based on the sensor data and road boundary information obtained from a preset map, wherein the sensor data includes 3D lidar data and GNSS data, and the road boundary information is represented by a list of four discrete points of respective boundaries; Convert the laser radar point corresponding to the brightness frame into UTM coordinates according to the current position and direction of the autonomous vehicle reported by GNSS, wherein the nearest lateral position point is found from the four boundary lists and marked as the laser radar point, and the laser radar point is deleted from the four discrete point lists in the list; Calculate the conversion from WGS coordinates to UTM coordinates, and obtain the lidar frame from the conversion from UTM coordinates to lidar coordinates, where WGS coordinates are World Geodetic System coordinates and UTM coordinates are Universal Transverse Mercator coordinates; According to the laser radar points, the boundary points are obtained in the laser radar frame by using the spatial index, and the 3D laser radar data is converted into the navigation coordinates according to the boundary points to generate the bright spot cloud map.

2. The Pareto optimal map matching method for autonomous vehicle positioning according to claim 1, characterized in that: The 3D laser radar data is converted into point cloud data into navigation coordinates according to the boundary points, including: Convert the point cloud data of the brightness frame into navigation coordinates and superimpose the converted point clouds together. The corresponding expression is: （1） in, are the coordinates of each point in the brightness frame, is the translation vector from the lidar frame to the navigation frame, is the rotation matrix between the lidar coordinates and the navigation coordinates. The definition expression is: （2） in, , and is the heading, pitch and roll information, and ; After there is a point cloud in the navigation frame, the brightness frames are accumulated together to form a bright spot cloud map.

3. The Pareto optimal map matching method for autonomous vehicle positioning according to claim 1, characterized in that: The laser radar points corresponding to the brightness frame are converted to UTM coordinates according to the current position and direction of the autonomous vehicle reported by GNSS, including: The dead reckoning model is built by estimating the vehicle's position based on its current position, speed, and direction. The wheel speed and yaw rate of the onboard sensors are used to calculate the position in the navigation coordinates, as well as the vehicle's heading between the lidar coordinates and the navigation coordinates. hour along The expression for the dead reckoning estimate of the axis position is: （3） in, for The estimated position at the moment, is the time interval between two measurements, for Observe the speed at all times. yes Always observe the azimuth, The real position at the moment It is expressed as: （4） in, for The true speed of time, for The true azimuth at the moment, the wheel speed measured by the wheel speed sensor and direction , is the heading angle of the vehicle, is the heading angle, is the vehicle's sideslip angle, The expression is: （5） in, is the yaw rate, measured by and Given, the corresponding noise term and Destruction, the measured value follows the variance and The zero-mean Gaussian distribution of Given the dead reckoning data and the observed velocity and direction Calculate the eastward and northward velocities, and the corresponding expressions are: （6） （7） in, is the measured speed, Indicates the direction being measured.

4. The Pareto optimal map matching method for autonomous vehicle positioning according to claim 3, characterized in that: Also includes: The bias in the dead reckoning estimate is quantified by the difference between the estimated position and the true position. The bias is calculated by evaluating the expected difference between the dead reckoning estimate and the true position. , the corresponding mathematical expression is: （8） The replacement expression is and for: （9） Further simplified to: （10） Among them, the preset is zero at time The estimate is unbiased, with a preset direction angle Additive Gaussian white noise If the damage is done, the observed direction is expressed as: （11） in, Indicates that the variance The expected value of the cosine of the noise angle can be approximated by using the properties of the Gaussian distribution to approximate the expected value of the low noise variance as: （12） Substituting formula (12) into formula (10) to calculate the deviation, the result is: （13） Introducing a correction factor based on system characteristics or noise factors, the deviation is rewritten as: （14） in, The bias in the dead reckoning model is based on the expected value of the deviation between the estimated position and the true position, considered as an adjustment related to system characteristics or empirical calibration.

5. The Pareto optimal map matching method for autonomous vehicle positioning according to claim 3, characterized in that: The variance of the estimation error in the dead reckoning model can be determined and evaluated by analyzing the propagation of velocity and direction measurement noise within the system. The variation of the position estimate is affected by the measured velocity. and direction The variance of the dead reckoning position estimate can be expressed as: （15） Using the characteristics of Gaussian noise, the variance of the product of velocity and direction angle cosine is: （16） in, represents the expected value, Approximately , the subsequent approximation is applicable and uses: （17） Here, variance is incorporated into the fusion process and the dead reckoning estimates are weighted relative to the range measurements.

6. The Pareto optimal map matching method for autonomous vehicle positioning according to claim 1, characterized in that: Calculate the conversion from WGS coordinates to UTM coordinates, and convert from UTM coordinates to lidar coordinates to get the lidar frame, including: The normal distribution transform (NDT) algorithm is used to convert the lidar point cloud into navigation coordinates to calculate the reference map, align the lidar with the navigation frame, and provide appropriate map matching for positioning. Formula (18) represents the lidar frame To Navigation Frame Transformation: （18） in, Indicates a point in the navigation frame The coordinates of represents the corresponding point in the laser radar coordinate system, represents the rotation matrix of the lidar relative to the navigation frame, Vector representing the translation between the LiDAR and navigation coordinate systems, rotation matrix Defined as: （19） in, , and Indicates the roll angle, pitch angle and heading angle, , , , and , .

7. The Pareto optimal map matching method for autonomous vehicle positioning according to claim 6 ， It is characterized in that Each laser radar scan point Convert to navigation frame and aggregate to form point cloud : （20） in, Aggregation generates point cloud map The total number of points, for the point cloud Optimizations include: Vegetation removal: filter out points corresponding to vegetation; Map downsampling: Downsample the map and keep the points that are most important for accurate positioning.

8. The Pareto optimal map matching method for autonomous vehicle positioning according to claim 6, characterized in that: The normal distribution transform (NDT) algorithm is used to convert the lidar point cloud into navigation coordinates to calculate the reference map, including: Use the given dead reckoning model to calculate the vehicle’s The estimated position at time is: （21） in, express The estimated easting and northing position at the moment, express The estimated speed at each moment, express Estimated heading at the moment, express Moment and The time interval between moments is transformed by finding the best alignment between the current lidar scan and the reference image , the transformation expression is: （22） in, is the rotation angle, , are the eastward and northward translation components respectively. The NDT map matching process is to determine the best transformation , which will reduce the current scan point and reference image The problem of minimizing the difference between is described as: （23） in, Indicates the number of points in the current scan, determining the best transformed Used to correct the dead reckoning position estimate, after implementing NDT-based map matching, the corrected position estimate can be expressed as: （24） A weighted fusion algorithm is used to fuse the dead reckoning and NDT estimates. Fusion position estimation at time It is expressed as: （25） in, and Represents the fusion weight, satisfying the condition , the difference of the merged position estimate is expressed as: （26） in, Indicates that the vehicle is The actual location at the moment.

9. The Pareto optimal map matching method for autonomous vehicle positioning according to claim 8, characterized in that: The uncertainty estimates in the NDT correction estimate are combined with the uncertainty estimates in the dead reckoning estimate, given the estimated error exist The covariance at time can be propagated at time k+1 using formula (27): （27） in, is the covariance estimated based on NDT map matching, is the covariance of the dead reckoning estimate, is the cross covariance between the NDT-based estimate and the dead reckoning estimate, and is calculated as: （28） in, represents the position estimate after dead reckoning, is the expected value of the dead reckoning estimate, represents the expected value of the NDT corrected estimate, is the expected value of the NDT corrected estimate, represents the expectation operator that computes the mean of a random variable, is the cross covariance of the two estimated errors to reflect the linear relationship between them.

10. The Pareto optimal map matching method for autonomous vehicle positioning according to claim 8, characterized in that: Get the error of the fused position estimate, Error in time The norm is less than or equal to The error at the moment, the system state transfer matrix The spectral radius is less than 1: and (29) in, is the spectral radius of the system state transfer matrix, and Represents the error between the fusion estimate and the true position, ensuring This means that the estimation error will not diverge over time.

Images