High-precision perception-driven autonomous driving positioning method and system

Abstract

The application provides a high-precision perception-driven automatic driving positioning method and system, and relates to the technical field of automatic driving. The method comprises the following steps: collecting vehicle environment data in real time through a multi-modal sensor array to generate a multi-source perception dataset; performing spatiotemporal synchronization processing; performing motion distortion correction on the fused perception data stream to generate dynamic compensation perception data; inputting the multi-source fusion positioning model, combining high-precision map data to perform three-dimensional space matching, and obtaining real-time positioning coordinates; and iteratively optimizing the real-time positioning coordinates through an adaptive error correction module to generate a vehicle positioning result. The application solves the technical problem that sensor data is affected by vehicle motion, resulting in data distortion and affecting positioning accuracy. Through motion distortion correction and compensation, the distortion caused by vehicle motion is eliminated, and the positioning accuracy and stability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to an automatic driving positioning method and system driven by high-precision perception. BACKGROUND

[0002] Automatic driving technology is one of the core technologies to realize autonomous driving of vehicles, and high-precision positioning is a key link to ensure the safe and efficient operation of automatic driving systems. The automatic driving positioning technology needs to accurately determine the position of the vehicle according to the perception information of the environment around the vehicle, combined with the map and the motion model. The positioning accuracy directly affects the decision and control of the automatic driving system, and thus determines the safety and reliability of automatic driving. During the driving process of the vehicle, especially in high-speed driving, sudden braking, sharp turning and other situations, the data collected by the sensor may be distorted due to the motion state of the vehicle, causing data distortion. Due to the problem of data distortion, the automatic driving system may deviate during positioning, especially during a long driving process, the error will continuously accumulate, eventually affecting the accuracy and stability of the overall positioning, thereby affecting the driving safety.

[0003] In summary, the prior art has the technical problem that the sensor data will be affected by the motion of the vehicle, resulting in data distortion and affecting the positioning accuracy. SUMMARY

[0004] The purpose of the present application is to provide an automatic driving positioning method and system driven by high-precision perception, to solve the technical problem in the prior art that the sensor data will be affected by the motion of the vehicle, resulting in data distortion and affecting the positioning accuracy.

[0005] In view of the above problems, the present application provides an automatic driving positioning method and system driven by high-precision perception.

[0006] In a first aspect, the present application provides an automatic driving positioning method driven by high-precision perception, which is realized by an automatic driving positioning system driven by high-precision perception, wherein the automatic driving positioning method driven by high-precision perception comprises: collecting vehicle environment data in real time through a multi-modal sensor array to generate a multi-source perception data set; performing spatio-temporal synchronization processing on the multi-source perception data set to establish a unified spatio-temporal reference fusion perception data stream; performing motion distortion correction on the fusion perception data stream to generate dynamic compensation perception data; inputting the dynamic compensation perception data into a multi-source fusion positioning model, combining high-precision map data for three-dimensional space matching to obtain real-time positioning coordinates; and performing iterative optimization on the real-time positioning coordinates through an adaptive error correction module to generate a vehicle positioning result.

[0007] Optionally, the laser radar is configured to scan the target area according to a preset scanning frequency to obtain three-dimensional point cloud data; the three-dimensional point cloud data triggers the visual sensor to perform image acquisition to obtain a plurality of regional image data; the millimeter wave radar is controlled to perform sensing on the target area according to a multi-target tracking mode to obtain speed vector information of a moving object; the three-dimensional point cloud data, the plurality of regional image data, and the speed vector information are time-stamped and aligned to generate the multi-source perception data set.

[0008] Optionally, a conversion matrix between coordinate systems of the multi-modal sensor array is established, the laser radar coordinate system is mapped to the vehicle body coordinate system according to the conversion matrix to obtain a first perception data set; a plurality of regional image data of the visual sensor is dynamically re-projected through a sliding window mechanism to obtain a second perception data set; asynchronous sampling data extracted based on the millimeter wave radar is interpolated and compensated to obtain a third perception data set; a spatiotemporal consistency constraint condition is constructed to verify geometric consistency of the first perception data set, the second perception data set, and the third perception data set in a three-dimensional space, and if the geometric data in the three-dimensional space is consistent, the multi-source perception data set is generated.

[0009] Optionally, angular velocity data and linear acceleration data of the target vehicle are obtained by inertia measurement of the target vehicle; motion compensation is performed based on the angular velocity data and the linear acceleration data to obtain a vehicle pose change amount; reverse motion compensation is performed on the three-dimensional point cloud data combined with the vehicle pose change amount to denoise the point cloud track of the dynamic obstacle to generate the dynamic compensation perception data.

[0010] Optionally, laser radar geometric features are extracted based on the three-dimensional point cloud data, and visual semantic features are extracted based on the plurality of regional image data; a deep neural network model is constructed, the visual semantic features and the laser radar geometric features are cross-modal feature fused through the deep neural network model to obtain visual-laser radar data; motion state estimation is performed on the speed vector information of the moving object according to an extended Kalman filtering algorithm to obtain a motion state estimation result; a motion state estimation result is obtained by introducing an attention mechanism to dynamically adjust the visual-laser radar data and the motion state estimation result for weight distribution, and the multi-source fusion positioning model is constructed.

[0011] Optionally, the road topology in the high-precision map is matched with the real-time positioning coordinates for matching degree evaluation, and a matching degree evaluation result is obtained; particle filtering is performed according to the matching degree evaluation result, a particle filtering result is obtained, the real-time positioning coordinates are dynamically adjusted according to the particle filtering result, and a particle distribution density is obtained; a historical positioning trajectory of a target vehicle is obtained, batch nonlinear optimization is performed on the historical positioning trajectory through a sliding window, and positioning optimized trajectory data is obtained; error calculation is performed on the positioning optimized trajectory data combined with the particle distribution density, and positioning error data is obtained; when the positioning error exceeds a preset threshold, a positioning mode is triggered for error reset, and the vehicle positioning result is determined.

[0012] Optionally, the matching degree of the real-time positioning coordinates output by the multi-source fusion positioning model and the high-precision map is monitored in real time, the positioning error data is determined according to the matching degree, and it is judged whether the positioning error data exceeds a preset threshold; when the positioning error data exceeds the preset threshold, an error trigger signal is generated, multi-source data consistency verification is started based on the error trigger signal, a positioning mode is selected according to a verification result, the positioning mode includes a tight combination positioning mode or a redundant positioning mode; error correction is performed based on the tight combination positioning mode or the redundant positioning mode, and the vehicle positioning result is generated according to the correction result combined with the spatial feature constraint relationship of the real-time positioning coordinates and the high-precision map.

[0013] In a second aspect, the present application also provides a high-precision perception driven automatic driving positioning system for executing the high-precision perception driven automatic driving positioning method as described in the first aspect, wherein the high-precision perception driven automatic driving positioning system comprises: a data acquisition module for acquiring vehicle environment data in real time through a multi-modal sensor array to generate a multi-source perception data set; a space-time synchronization processing module for performing space-time synchronization processing on the multi-source perception data set to establish a unified space-time reference fusion perception data stream; a motion distortion correction module for performing motion distortion correction on the fusion perception data stream to generate dynamic compensation perception data; a spatial matching module for inputting the dynamic compensation perception data into a multi-source fusion positioning model, combining high-precision map data for three-dimensional spatial matching, and obtaining real-time positioning coordinates; and a vehicle positioning module for performing iterative optimization on the real-time positioning coordinates through an adaptive error correction module to generate a vehicle positioning result.

[0014] One or more technical solutions provided in the present application have at least the following beneficial effects:

[0015] The vehicle environment data is collected in real time through a multi-modal sensor array to generate a multi-source perception data set; the multi-source perception data set is processed in time and space synchronization to establish a unified time and space reference fusion perception data stream; the fusion perception data stream is corrected for motion distortion to generate dynamic compensation perception data; the dynamic compensation perception data is input into a multi-source fusion positioning model, combined with high-precision map data for three-dimensional space matching to obtain real-time positioning coordinates; the real-time positioning coordinates are iteratively optimized through an adaptive error correction module to generate vehicle positioning results. That is, by processing all sensor data in time and space synchronization, then correcting for motion distortion to eliminate distortion caused by vehicle motion, combined with high-precision map to determine positioning coordinates, and using an adaptive error correction module to continuously optimize positioning results, the accuracy and stability of positioning are improved, thereby improving the safety and reliability of autonomous driving.

[0016] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the following specific embodiments of the present application can be implemented according to the content of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described. It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only exemplary, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0018] Figure 1 Flowchart of the high-precision perception-driven autonomous driving positioning method of the present application;

[0019] Figure 2 Structure diagram of the high-precision perception-driven autonomous driving positioning system of the present application.

[0020] Explanation of reference signs: data acquisition module 11, time and space synchronization processing module 12, motion distortion correction module 13, space matching module 14, vehicle positioning module 15. DETAILED DESCRIPTION

[0021] The application provides a high-precision perception-driven automatic driving positioning method and system, which solves the technical problem that the sensor data is affected by vehicle movement, resulting in data distortion and affecting positioning accuracy in the prior art. By performing spatio-temporal synchronization processing on all sensor data, and then performing motion distortion correction, the distortion caused by vehicle movement is eliminated, the positioning coordinates are determined in combination with a high-precision map, and the positioning result is continuously optimized using an adaptive error correction module, thereby improving the positioning accuracy and stability, and improving the safety and reliability of automatic driving.

[0022] The technical solutions in the application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the application, rather than all the embodiments of the application. It should be understood that the application is not limited to the example embodiments described herein. Based on the embodiments of the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of the application. In addition, it should be noted that, for the convenience of description, only parts related to the application are shown in the drawings, rather than all parts.

[0023] Embodiment one, please refer to the accompanying Figure 1 The application provides a high-precision perception-driven automatic driving positioning method, wherein the high-precision perception-driven automatic driving positioning method is executed by a high-precision perception-driven automatic driving positioning system, and the high-precision perception-driven automatic driving positioning method specifically comprises the following steps:

[0024] S100: Real-time acquisition of vehicle environment data by a multi-modal sensor array to generate a multi-source perception data set.

[0025] Specifically, the vehicle is equipped with a multi-modal sensor array composed of a laser radar, a vision sensor and a millimeter wave radar. These sensors each have their own focus. The laser radar is used to acquire accurate three-dimensional point cloud data, the vision sensor captures high-resolution image information, and the millimeter wave radar performs stably in adverse weather such as rain, snow and fog. During vehicle movement, the multi-modal sensor array continuously acquires vehicle environment data, and at the same time of acquiring data, each sensor will attach a time stamp to the data acquired by it. The vehicle environment data includes information such as roads, obstacles, pedestrians and traffic signs around the vehicle.

[0026] Due to the different working frequencies and response times of different sensors, it is necessary to align the data collected by each sensor through timestamp synchronization. Through a unified clock signal, it is ensured that all sensor data have a common time reference. Interpolation and alignment of data with different clock sampling rates are performed using software algorithms to ensure that laser radar, camera, and millimeter wave radar data are available at each moment. Through multi-modal sensor array and precise timestamp synchronization alignment, a high-quality multi-source perception dataset is effectively generated, improving the accuracy and robustness of environmental perception.

[0027] Further, the S100 of the present application comprises:

[0028] The laser radar is configured to scan the target area according to a preset scanning frequency, and to obtain three-dimensional point cloud data. The three-dimensional point cloud data triggers the visual sensor to perform image acquisition, and obtains multiple regional image data. The millimeter wave radar is controlled to perceive the target area according to a multi-target tracking mode, and to obtain velocity vector information of moving objects. The three-dimensional point cloud data, the multiple regional image data, and the velocity vector information are time-stamped and aligned, and the multi-source perception dataset is generated.

[0029] A conversion matrix between multi-modal sensor array coordinate systems is established, the laser radar coordinate system is mapped to the vehicle body coordinate system according to the conversion matrix, and a first perception dataset is obtained. A sliding window mechanism is used to dynamically reproject the multiple regional image data of the visual sensor, and a second perception dataset is obtained. Interpolation compensation is performed on asynchronous sampling data extracted based on the millimeter wave radar, and a third perception dataset is obtained. A spatio-temporal consistency constraint condition is constructed, and the geometric consistency of the first perception dataset, the second perception dataset, and the third perception dataset in three-dimensional space is verified. If the geometric data in three-dimensional space is consistent, the multi-source perception dataset is generated.

[0030] Specifically, first, the laser radar is configured, and the scanning frequency (such as 10 scans per second) is set. The laser radar emits laser pulses according to the set frequency, and records the time difference of reflected light, thereby obtaining continuous three-dimensional point cloud data. The three-dimensional point cloud data is a set of discrete data points reflecting the spatial structure of the target area obtained by the laser radar, each point having specific three-dimensional coordinate information, which is used to construct a three-dimensional model of the environment. For example, in a 30-second test, the laser radar will collect 300 frames of point cloud data, each frame containing about 100,000 points on average, detailing the geometric structure of the environment around the vehicle.

[0031] According to the three-dimensional point cloud data of the laser radar, a visual sensor (such as a camera) is triggered to perform image acquisition. For example, when the laser radar detects that there is an object in a certain area, the camera will be triggered to capture the image of the area and obtain a plurality of area image data corresponding to the three-dimensional point cloud data. The camera captures the image according to the set resolution and frame rate (such as 1920x1080 resolution, 30 frames / second). The visual sensor generally refers to the camera, which can capture high-definition images, and is triggered according to the point cloud data generated by the laser radar to collect image data of a specific area, and supplement the texture and color information of the point cloud data.

[0032] At the same time, the millimeter wave radar is configured in a multi-target tracking mode to perceive the motion of a plurality of moving objects in a target area in real time and obtain the speed vector information of each target. The millimeter wave radar is a sensor that uses millimeter wave frequency band electromagnetic waves for target detection, and is mainly used for multi-target tracking and obtaining the speed vector information of moving objects, especially in bad weather such as rain and snow. Assuming that the millimeter wave radar works at a rate of 50Hz, about 1500 pieces of speed vector data can be obtained during the same test period, which describes the motion state of the target object in detail.

[0033] In order to ensure the consistency of various sensor data in space and time, a high-precision timestamp synchronization technology is adopted. According to the relative position and direction of the multi-modal sensor array, a conversion matrix between the multi-modal sensor array coordinate systems is established, which is used to convert the data of one sensor (such as a laser radar) from its own coordinate system to the vehicle body coordinate system. The conversion matrix is usually obtained by external parameter calibration to ensure that the data of various sensors are aligned in the same reference frame. Taking the laser radar as an example, by using a laser calibration board or a calibration target, the data of the laser radar and other sensors (such as a camera) on the vehicle are collected to realize the geometric relationship calibration between the sensors. Using calibration software (such as the camera calibration toolbox of MATLAB or special laser radar calibration software), the conversion matrix of the laser radar coordinate system to the vehicle body coordinate system is calculated. The matrix is usually a 4x4 homogeneous transformation matrix, which includes a rotation matrix and a translation vector. The three-dimensional point cloud data collected by the laser radar is subjected to coordinate transformation according to the homogeneous transformation matrix. Through matrix operation, each point is mapped to the vehicle body coordinate system to form a first perception data set.

[0034] For visual sensor data, a sliding window mechanism is used to dynamically re-project the continuously captured image data. With a 5-frame window, the real-time motion data of the vehicle (such as vehicle speed and angular velocity obtained through the IMU) is used to geometrically correct each frame of image according to its time position in the window, eliminating image distortion caused by vehicle acceleration, turning and other dynamic factors. For example, the original image has an average blur error of 0.1 meters caused by changes in vehicle speed, and after dynamic re-projection processing, the error is reduced by 80% to 0.02 meters, ensuring that the image accurately reflects the environment. The second perception dataset is a set of visual image data that accurately reflects the environmental features without motion distortion after the sliding window mechanism and dynamic re-projection processing.

[0035] For millimeter wave radar, due to the asynchronous sampling problem, an interpolation compensation technique is used to compensate the continuity of the collected data in time. The time interval data generated by the asynchronous sampling of the millimeter wave radar is mathematically compensated (such as linear interpolation or spline interpolation), so that the data is more continuous and consistent in time dimension. Assuming that the original sampling frequency of the millimeter wave radar is 45Hz, but the data has interval inconsistency, through the linear interpolation method, the time error of the interpolated data is reduced to within 2 milliseconds, thereby generating a third perception dataset, which contains accurate motion target speed vector information. The third perception dataset is a dataset obtained from the millimeter wave radar after interpolation compensation, which contains motion object speed vector information.

[0036] To ensure the consistency of the first, second and third perception datasets in space and time, a spatio-temporal consistency constraint condition is constructed to ensure the geometric consistency of data from different sensors in three-dimensional space by establishing matching conditions in space and time. The geometric positions of each dataset in three-dimensional space are compared and verified using known static features in the environment (such as building boundaries, road profiles) and vehicle dynamic parameters. The geometric consistency of the first, second and third perception datasets in three-dimensional space is verified, and if the geometric data of the first, second and third perception datasets in three-dimensional space are consistent, a multi-source perception dataset is obtained. Through the above steps, the problems of image distortion caused by coordinate system differences and vehicle motion between different sensors, as well as time deviation caused by asynchronous sampling, are eliminated, and the positioning accuracy and environmental perception stability of the automatic driving system are greatly improved.

[0037] S200: Spatio-temporal synchronization processing is performed on the multi-source perception dataset to establish a unified spatio-temporal reference fusion perception data stream.

[0038] Specifically, the aforementioned multi-source perception data set is processed in time and space synchronization, the multi-source data is calibrated and aligned in time and space, and it is ensured that the data of each sensor reflects the environment information under the same time and the same space reference. Each type of data in the laser radar, visual sensor and millimeter wave radar carries its own time stamp and space information. Due to the different sampling frequencies and data formats of these sensors, direct use will have inconsistency in time and space, and time and space synchronization processing is needed.

[0039] A high-precision clock synchronization technology such as PTP protocol is adopted to uniformly calibrate the time of all sensors, so that the sampling time of each sensor can be mapped to the same global clock. At the same time, through external parameter calibration and coordinate conversion, the data of each sensor is converted to the vehicle body coordinate system to form a unified space reference. For example, in an experiment, the time stamp error of the laser radar is controlled within 5 milliseconds, and after coordinate conversion, the geometric error between the laser radar point cloud and the vehicle body coordinate system is less than 0.02 meters. Through interpolation, nearest neighbor matching and other algorithms, the data collected at different times are corrected, so that the laser radar point cloud, image data and millimeter wave radar velocity vector information at each time can be accurately corresponded.

[0040] The specific process of time synchronization processing includes: selecting a sensor as the master clock, usually the sensor with the highest time resolution and stability, aligning the time stamps of all sensors to the time reference of the master clock. For sensors that cannot be completely synchronized, record and compensate the time delay between them and the master clock. The space synchronization processing includes: selecting the vehicle body coordinate system as the unified space reference system, converting the data of the sensors to the vehicle body coordinate system.

[0041] The data after time synchronization and space coordinate conversion are fused to form a continuous and unified data stream. This data stream contains information of all sensors and is arranged according to the unified time and space reference. The fused perception data stream refers to the continuous data stream formed by fusing data from different sensors under the unified time and space reference, which reflects the comprehensive perception information of the environment around the vehicle. Through time and space alignment processing, the data of each sensor is strictly aligned to the unified time and space reference, solving the error problem of multi-sensor data due to inconsistent sampling time and space reference.

[0042] S300: Motion distortion correction is performed on the fused perception data stream to generate dynamic compensation perception data.

[0043] Further, the S300 of the present application comprises:

[0044] The angular velocity data and linear acceleration data of the target vehicle are obtained by inertia measurement on the target vehicle; motion compensation is performed based on the angular velocity data and the linear acceleration data to obtain a vehicle pose change quantity; the three-dimensional point cloud data is combined with the vehicle pose change quantity to perform reverse motion compensation, and the point cloud track of the dynamic obstacle is denoised to generate the dynamic compensation perception data.

[0045] Specifically, motion distortion correction refers to correcting the data distortion caused by the displacement and rotation of the vehicle during the scanning process, so that the data collected by each sensor can reflect the real and static environment. The vehicle collects angular velocity and linear acceleration data in real time through the built-in inertial measurement unit (IMU), which reflects the motion state of the vehicle in three-dimensional space. For example, the IMU records the angular velocity of the vehicle as (0.1 rad / s, 0.2 rad / s, 0.05 rad / s) and the linear acceleration as (0.5 m / s 2 ,-0.3 m / s 2 ,0 m / s 2 ) at time t. The angular velocity data refers to the data collected by the inertial measurement unit (IMU) describing the rotation rate of the vehicle around each axis, usually in radians per second, and the linear acceleration data describes the change rate of acceleration or deceleration of the vehicle in the straight line direction, usually in meters per second squared.

[0046] Using angular velocity and linear acceleration data, the attitude (i.e., pose) change quantity of the vehicle during data collection is calculated by mathematical methods such as integration, so as to correct the distortion of sensor data caused by vehicle motion. For example, in a 0.1-second collection period, the average angular velocity recorded by the IMU is 0.03 rad / s, and the average linear acceleration is 0.5 m / s 2 , and the initial speed of the vehicle in this period is 0.5 m / s. In order to obtain the pose change quantity of the vehicle in this collection period, the angular velocity and acceleration data need to be integrated respectively. For angular velocity data, △θ = ω × △t = 0.003 rad, where △θ is the angular velocity, ω is the average angular velocity, and △t is the time length 0.1 s. For linear acceleration data, if the initial speed v0 is considered, the displacement is calculated by the formula △s = v0 × △t + 1 / 2a × (△t) 2 = 0.0525 m. Through integration operation, the pose change quantity of the vehicle in this period is obtained, which is approximately 0.05 meters of translation and 0.003 rad of rotation. According to this, the vehicle pose change quantity is obtained, which is the comprehensive quantity of the translation and rotation change of the vehicle in a certain period of time, usually represented in the form of displacement and rotation angle.

[0047] A motion transformation matrix is constructed using the vehicle pose change amount, and then its inverse matrix is taken to implement reverse motion compensation. Each point cloud data point of the three-dimensional point cloud data is multiplied by the reverse motion compensation matrix to obtain the corrected points, the motion of the vehicle during scanning is frozen, the data is converted to a static reference state, thereby eliminating the geometric distortion of the point cloud data caused by the vehicle displacement. For example, in an experiment, the trajectory of the dynamic obstacle point cloud before compensation showed an average position error of about 0.15 meters, and after reverse motion compensation, the error was reduced to about 0.05 meters.

[0048] After completing the reverse motion compensation, the point cloud data of the dynamic obstacle may still be affected by environmental noise or sensor errors, and the trajectory thereof may fluctuate. The trajectory data is subjected to denoising processing, and a median filter or Kalman filter algorithm is usually used to smooth the point cloud trajectory and further remove outliers. After reverse motion compensation and denoising processing, dynamic compensation perception data is generated, and after removing the scanning errors and noise interference caused by vehicle motion, the data set can more truly and stably reflect the environmental features and the motion state of the dynamic obstacle.

[0049] Motion distortion correction refers to calculating and compensating for sensor data distortion caused by vehicle motion. During vehicle driving, due to the influence of motion, the data collected by the sensor may deviate in position and direction, resulting in distortion of point cloud data, image data, etc. Motion distortion correction is to compensate for these distortions through algorithms to restore the original environmental data. After motion distortion correction, the error of the perception data caused by vehicle motion is corrected to obtain dynamic compensation perception data. By performing motion distortion correction on the fused perception data stream, the vehicle can eliminate the data distortion caused by motion, and denoising processing further improves the stability and accuracy of the point cloud data, ensuring high precision of environmental modeling and dynamic target detection.

[0050] S400: input the dynamic compensation perception data into a multi-source fusion positioning model, combine high-precision map data for three-dimensional space matching, and obtain real-time positioning coordinates.

[0051] Specifically, the dynamic compensation perception data after motion distortion correction and denoising processing is input into the multi-source fusion positioning model. The input perception data is matched with high-precision map data using the algorithm in the multi-source fusion positioning model. The static environmental features such as road network, building boundary, and road signs contained in the high-precision map data provide positioning reference for vehicle positioning. The input dynamic compensation perception data (such as laser radar three-dimensional point cloud) is compared with the static environmental features in the high-precision map data. For example, laser radar point cloud can help identify the boundary of the road and the location of the building, and then match on the map. Through the three-dimensional space matching of the multi-source fusion positioning model and the high-precision map data, the accurate position of the current vehicle is calculated. This positioning coordinate is real-time and continuously updated with the movement of the vehicle. For example, assuming that the vehicle is driving on a known road and perceives the obstacles (such as streetlights) on both sides of the road through the laser radar, and the visual sensor recognizes the stop sign. The multi-source fusion positioning model matches these perception data with the road information in the high-precision map to accurately calculate the position coordinate of the current vehicle. By combining dynamic compensation perception data with high-precision map data for three-dimensional space matching, high-precision real-time positioning is achieved, and the reference point on the map is accurately positioned, ensuring the accuracy of navigation and path planning.

[0052] Further, the S400 of the present application comprises:

[0053] Based on the three-dimensional point cloud data, laser radar geometric features are extracted, and based on the multiple region image data, visual semantic features are extracted; a deep neural network model is constructed, the visual semantic features and the laser radar geometric features are fused through the deep neural network model to obtain visual-laser radar data; the speed vector information of the moving object is estimated in motion state according to the extended Kalman filter algorithm to obtain a motion state estimation result; a attention mechanism is introduced to dynamically adjust the visual-laser radar data and the motion state estimation result for weight distribution, and the multi-source fusion positioning model is constructed.

[0054] Specifically, geometric features are extracted by processing the three-dimensional point cloud data collected by the laser radar. Laser radar geometric features refer to environmental geometric information extracted by processing three-dimensional point cloud data, including the shape, size, and edge features of obstacles. For example, a normal estimation-based algorithm is used to extract surface shape, edge, and other information in the point cloud. Taking the road and obstacles as examples, the point cloud data of the laser radar can accurately reflect the geometric shape of the obstacles, such as lane lines, three-dimensional buildings, and other structures.

[0055] Image data collected by visual sensors is used to extract semantic features from images using existing deep learning methods such as convolutional neural networks (CNNs). Visual sensors can identify and label objects in images, such as pedestrians, vehicles, and traffic signs. Visual semantic features include object categories (such as pedestrians, vehicles, and road signs), as well as object color and texture.

[0056] A deep neural network model is constructed to fuse visual semantic features with lidar geometric features. This is achieved through a multi-channel neural network. The model takes in lidar geometric information and image semantic information separately as input, and automatically learns the associations between them. For example, the model can learn how to combine visual pedestrian categories with lidar obstacle geometry to more accurately determine obstacle locations and types. Deep neural network models typically contain multiple input channels, one for visual semantic features and another for lidar geometric features. By fusing features from these two channels, the network learns the associations between them. Deep neural network design requires the inclusion of fusion layers in the middle or final layers of the network. These layers are specifically designed to fuse features from different modalities, including weighted sum fusion, concatenation fusion, and attention fusion. For example, the fusion layer concatenates visual and lidar features and maps them into a unified feature space through a fully connected layer.

[0057] After deep neural network fusion, a unified dataset is obtained, which contains the combined information of vision and lidar. Vision-lidar data is a dataset fused through a deep neural network model, which contains semantic information from the vision sensor (image) and geometric information from the lidar (point cloud).

[0058] For moving objects (such as other vehicles or pedestrians), the extended Kalman filter algorithm is used to estimate their motion state. Based on the current velocity vector information and a prediction model, the object's future position and velocity are estimated. For example, when the velocity vector of another vehicle is detected, the EKF can estimate the vehicle's future motion trajectory and provide an accurate state estimate. The extended Kalman filter (EKF) algorithm is a nonlinear filtering algorithm commonly used for state estimation of dynamic systems, especially in the presence of noise. The EKF can estimate the motion state (position, velocity, etc.) of an object. The motion state estimate obtained through the extended Kalman filter algorithm is the object's current state, including its position, velocity, acceleration, etc.

[0059] The weight of different data sources is dynamically adjusted through an attention mechanism combined with visual-lidar data and motion state estimation results. When a moving object with high speed is detected, more attention may be paid to the motion state estimation of the moving object, while in a static environment, more reliance is placed on the fusion features of vision and lidar. The attention mechanism is used to dynamically adjust the importance of different input data, automatically adjust the attention to different features according to the context information or the requirements of the current task, and thus improve the performance of the model.

[0060] The feature fusion and motion state estimation results in the above steps are integrated through a multi-source fusion positioning model, which can provide high-precision positioning information according to the input data of the sensor. For example, the system may combine the data of vision and lidar to provide the relative position of the current vehicle and other objects in the environment, such as obstacles, pedestrians, etc. The multi-source fusion positioning model is a model based on the fusion of data from multiple sensors such as vision, lidar, IMU, etc. for precise positioning and state estimation. Through the complementary characteristics of multiple sensors, the positioning accuracy and robustness are improved.

[0061] By constructing a multi-source fusion positioning model, combining the geometric features of lidar and the semantic features of vision, and dynamically adjusting the data weight through an attention mechanism, not only the accuracy of environment perception is improved, but also the tracking ability of moving objects is enhanced, making it possible to more accurately predict the state of moving objects, thereby improving the positioning and navigation performance of autonomous vehicles in complex environments.

[0062] S500: iteratively optimize the real-time positioning coordinates through an adaptive error correction module to generate a vehicle positioning result.

[0063] Further, the S500 of the present application comprises:

[0064] extract the road topology structure in the high-precision map and perform matching degree evaluation on the real-time positioning coordinates, and obtain a matching degree evaluation result; perform particle filtering according to the matching degree evaluation result to obtain a particle filtering result, dynamically adjust the real-time positioning coordinates according to the particle filtering result, and obtain a particle distribution density; obtain a historical positioning trajectory of a target vehicle, perform batch nonlinear optimization on the historical positioning trajectory through a sliding window to obtain positioning optimization trajectory data; combine the positioning optimization trajectory data with the particle distribution density to perform error calculation and obtain positioning error data; when the positioning error exceeds a preset threshold, trigger error reset in a positioning mode to determine the vehicle positioning result.

[0065] The matching degree of the real-time positioning coordinates output by the multi-source fusion positioning model with the high-precision map is monitored in real time, the positioning error data is determined according to the matching degree, and it is judged whether the positioning error data exceeds a preset threshold; when the positioning error data exceeds the preset threshold, an error trigger signal is generated, multi-source data consistency verification is started based on the error trigger signal, a positioning mode is selected according to the verification result, the positioning mode includes a tight combination positioning mode or a redundant positioning mode; error correction is performed based on the tight combination positioning mode or the redundant positioning mode, and the vehicle positioning result is generated according to the correction result in combination with the spatial feature constraint relationship between the real-time positioning coordinates and the high-precision map.

[0066] Specifically, the road topology structure in the high-precision map is the spatial relationship and connection mode between road elements in the map, such as road intersections, turns, road segment connections, etc. The road topology structure is extracted from the high-precision map, including road geometry information, intersections, turns, traffic signs, etc. The real-time positioning coordinates are matched with the road topology structure in the map, and the matching degree of real-time positioning and map data is evaluated. High matching degree indicates that the current position of the vehicle is highly consistent with the relationship of the road on the map, and low matching degree may indicate that there is an error in positioning. For example, if the current positioning coordinates of the vehicle are (100.5, 200.3), and the corresponding position in the high-precision map is (100.55, 200.35), the matching degree is evaluated by calculating the distance between the two points.

[0067] According to the evaluation result of the matching degree, a particle filter method is used for dynamic adjustment. Particle filtering generates multiple particles to represent the possible poses of the vehicle, and updates the weight and position of each particle through filtering algorithm. That is, according to the result of particle filtering, the positioning coordinates of the vehicle are corrected. In the particle filtering process, the distribution of particles represents the possibility of different positions in the state space. The particle distribution density refers to the concentration of particle distribution in space, and the area with higher density indicates that the area is more likely to be the true position of the vehicle. For example, assuming that the matching degree of the high-precision map and the real-time positioning coordinates is 0.95, then the particle filter will adjust the position and distribution of the particles according to this evaluation, so that the distribution of the particles is closer to the true position. Particle filtering can continuously adjust the position estimation of the vehicle in a dynamic environment (such as obstacles or other changes).

[0068] The historical positioning trajectory of the target vehicle is obtained, which is the record of the vehicle's trajectory in the past period of time, containing the historical position data of the vehicle. The historical positioning trajectory is divided into multiple data blocks using the sliding window technique. Each sliding window contains a fixed number of consecutive position data. Over time, the window slides in the historical data, adding new positioning data each time it is updated and removing the oldest data. In each window, the trajectory data is optimized using a nonlinear optimization algorithm such as Levenberg-Marquardt, least squares, etc. The goal of optimization is to minimize the error between the trajectory points. After the optimization process is complete, the optimized trajectory data in each sliding window is obtained, which is combined into the final positioning optimization trajectory.

[0069] Batch nonlinear optimization refers to the overall optimization of a series of data points in the historical positioning trajectory. Unlike traditional linear optimization, nonlinear optimization deals with problems that have nonlinear relationships and is very effective for complex motion trajectories in reality, such as turns and accelerations that may occur during vehicle travel. Batch optimization can obtain more comprehensive and accurate trajectory correction by processing multiple data points simultaneously. The positioning optimization trajectory data is the optimized trajectory data, representing the accurate path the vehicle has taken in the historical time period, closer to the actual path, and removing noise and errors in the original trajectory.

[0070] In the particle filter algorithm, particles represent multiple possible positions of the vehicle, and particle distribution density represents the concentration of particles in different positions in the state space. Areas with higher particle density represent higher probability of that position being the true position. Particle filtering dynamically estimates the state of the vehicle by updating the weights and positions of the particles. The optimized trajectory data is combined with the particle filtering results, and the difference between the particle positions generated by particle filtering and the positions of the optimized trajectory is evaluated. The difference between the positioning optimization trajectory and the actual vehicle path is quantified, and the positioning error data is obtained by calculating the difference between the optimized trajectory and the particle distribution density, which is used to evaluate the accuracy of the positioning system. The smaller the error, the more accurate the positioning system. By calculating the error for each trajectory point, all error data is aggregated into a total error data set, representing the deviation between the optimized positioning data and the actual path.

[0071] An error limit is set in advance to determine whether the positioning result is accurate. When the positioning error exceeds the preset threshold, the positioning mode is triggered to reset the error. The output from the multi-source fusion positioning model, i.e. the real-time positioning coordinates of the vehicle, is monitored in real time, and the matching degree between the real-time positioning coordinates and the high-precision map is calculated using the pre-set position and spatial information in the high-precision map. According to the matching degree, the positioning error data, i.e. the error between the real-time positioning coordinates and the map position, is determined, and it is judged whether the positioning error data exceeds the preset threshold, i.e. the maximum acceptable error range.

[0072] When the positioning error exceeds the threshold, the positioning accuracy is considered insufficient, error correction needs to be started, and an error trigger signal is generated to start multi-source data consistency verification. Check if the data from different sensors (such as GNSS, IMU, vision, and lidar) are consistent, and make adjustments according to the verification results. According to the results of data consistency verification, select the trigger positioning mode, including using the tight combination positioning mode or the redundant positioning mode. The tight combination positioning mode is to tightly integrate data from multiple sensors through a factor graph optimization framework to improve positioning accuracy. For example, combine GNSS pseudorange and IMU acceleration data to reduce the error impact of a single sensor. The redundant positioning mode refers to enabling other sensors when some sensor data is inaccurate, such as vision-lidar odometry, which uses a closed-loop detection mechanism for positioning to avoid relying on a single sensor.

[0073] If the tight combination positioning mode is selected, the factor graph optimization framework will be used to integrate multi-band GNSS pseudorange / carrier phase and IMU data to correct positioning errors. If the redundant positioning mode is selected, the closed-loop detection mechanism of the vision-lidar odometry will be activated to correct the positioning error by detecting repeated features in the environment. After selecting the positioning mode, the real-time positioning coordinates are corrected according to the selected mode, and the corrected positioning data is matched with the high-precision map to further optimize the positioning result and ensure that the positioning accuracy meets the preset requirements.

[0074] By dynamically monitoring the matching degree of real-time positioning coordinates and high-precision maps, errors can be detected and corrected in a timely manner, which can effectively improve positioning accuracy. According to the size of the error, different positioning modes (tight combination or redundant mode) are selected for correction to ensure that the positioning result maintains high accuracy in any environment. The error trigger mechanism ensures that when the positioning accuracy deviates from the predetermined range, it can respond in a timely manner and start the necessary correction steps to maintain high-precision real-time positioning in complex environments, ensuring the safety and reliability of the vehicle during autonomous driving.

[0075] In summary, the high-precision perception-driven autonomous driving positioning method provided by the present application has the following beneficial effects:

[0076] The vehicle environment data is collected in real time by a multi-modal sensor array to generate a multi-source perception data set; the multi-source perception data set is processed in time and space synchronization to establish a unified time and space reference fusion perception data stream; the fusion perception data stream is corrected for motion distortion to generate dynamic compensation perception data; the dynamic compensation perception data is input into a multi-source fusion positioning model, combined with high-precision map data for three-dimensional space matching to obtain real-time positioning coordinates; the real-time positioning coordinates are iteratively optimized by an adaptive error correction module to generate vehicle positioning results. That is, by processing all sensor data in time and space synchronization, then correcting for motion distortion to eliminate distortion caused by vehicle motion, combined with high-precision map to determine positioning coordinates, and using an adaptive error correction module to continuously optimize positioning results, the accuracy and stability of positioning are improved, thereby improving the safety and reliability of autonomous driving.

[0077] Embodiment Two, based on the same inventive concept as the high-precision perception-driven autonomous driving positioning method in the foregoing Embodiment One, the present application also provides a high-precision perception-driven autonomous driving positioning system, please refer to the attached Figure 2 , the high-precision perception-driven autonomous driving positioning system comprises:

[0078] The data acquisition module 11 is configured to collect vehicle environment data in real time by a multi-modal sensor array to generate a multi-source perception data set; the time and space synchronization processing module 12 is configured to process the multi-source perception data set in time and space synchronization to establish a unified time and space reference fusion perception data stream; the motion distortion correction module 13 is configured to correct the fusion perception data stream for motion distortion to generate dynamic compensation perception data; the space matching module 14 is configured to input the dynamic compensation perception data into a multi-source fusion positioning model, combined with high-precision map data for three-dimensional space matching to obtain real-time positioning coordinates; the vehicle positioning module 15 is configured to iteratively optimize the real-time positioning coordinates by an adaptive error correction module to generate vehicle positioning results.

[0079] Further, the data acquisition module 11 in the high-precision perception-driven autonomous driving positioning system is further configured to:

[0080] The laser radar is configured to scan the target area according to a preset scanning frequency to obtain three-dimensional point cloud data; the visual sensor is triggered to perform image acquisition according to the three-dimensional point cloud data to obtain a plurality of regional image data; the millimeter wave radar is controlled to perform perception on the target area according to a multi-target tracking mode to obtain velocity vector information of moving objects; the three-dimensional point cloud data, the plurality of regional image data, and the velocity vector information are time-stamp synchronized and aligned to generate the multi-source perception data set.

[0081] Further, the data acquisition module 11 in the high-precision perception-driven autonomous driving positioning system is further used for:

[0082] A conversion matrix between the multi-modal sensor array coordinate systems is established, the laser radar coordinate system is mapped to the vehicle body coordinate system according to the conversion matrix, a first perception data set is obtained, a plurality of regional image data of the vision sensor is dynamically re-projected through a sliding window mechanism, a second perception data set is obtained, asynchronous sampling data is extracted based on the millimeter wave radar, interpolation compensation is performed, a third perception data set is obtained, a spatio-temporal consistency constraint condition is constructed, and geometric consistency of the first perception data set, the second perception data set and the third perception data set in a three-dimensional space is verified, and if the geometric data in the three-dimensional space is consistent, the multi-source perception data set is generated.

[0083] Further, the motion distortion correction module 13 in the high-precision perception-driven autonomous driving positioning system is further used for:

[0084] By inertia measurement on the target vehicle, angular velocity data and linear acceleration data of the target vehicle are obtained, motion compensation is performed based on the angular velocity data and the linear acceleration data, a vehicle pose change amount is obtained, reverse motion compensation is performed on the three-dimensional point cloud data combined with the vehicle pose change amount, and the point cloud trajectory of the dynamic obstacle is denoised to generate the dynamic compensation perception data.

[0085] Further, the space matching module 14 in the high-precision perception-driven autonomous driving positioning system is further used for:

[0086] Laser radar geometric features are extracted based on the three-dimensional point cloud data, and vision semantic features are extracted based on the plurality of regional image data, a deep neural network model is constructed, the vision semantic features and the laser radar geometric features are cross-modal feature fused through the deep neural network model to obtain vision-laser radar data, the velocity vector information of the moving object is motion state estimated according to an extended Kalman filtering algorithm to obtain a motion state estimation result, and a attention mechanism is introduced to dynamically adjust the vision-laser radar data and the motion state estimation result for weight distribution, and the multi-source fusion positioning model is constructed.

[0087] Further, the vehicle positioning module 15 in the high-precision perception-driven autonomous driving positioning system is further used for:

[0088] The road topology in the high-precision map is matched with the real-time positioning coordinates for matching degree evaluation, and a matching degree evaluation result is obtained. Particle filtering is performed according to the matching degree evaluation result, a particle filtering result is obtained, the real-time positioning coordinates are dynamically adjusted according to the particle filtering result, and a particle distribution density is obtained. A historical positioning trajectory of a target vehicle is obtained, and the historical positioning trajectory is batched and nonlinearly optimized through a sliding window to obtain positioning optimized trajectory data. The positioning optimized trajectory data is combined with the particle distribution density for error calculation to obtain positioning error data. When the positioning error exceeds a preset threshold, a positioning mode is triggered for error reset to determine the vehicle positioning result.

[0089] Further, the vehicle positioning module 15 in the high-precision perception driven automatic driving positioning system is also used for:

[0090] The matching degree between the real-time positioning coordinates output by the multi-source fusion positioning model and the high-precision map is monitored in real time, the positioning error data is determined according to the matching degree, and it is determined whether the positioning error data exceeds a preset threshold. When the positioning error data exceeds the preset threshold, an error trigger signal is generated, multi-source data consistency verification is started based on the error trigger signal, a positioning mode is selected according to the verification result, the positioning mode includes a tight combination positioning mode or a redundant positioning mode, error correction is performed based on the tight combination positioning mode or the redundant positioning mode, and the vehicle positioning result is generated according to the correction result in combination with the spatial feature constraint relationship between the real-time positioning coordinates and the high-precision map.

[0091] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The foregoing Figure 1 The high-precision perception driven automatic driving positioning method and specific examples in embodiment one are also applicable to the high-precision perception driven automatic driving positioning system of the present embodiment. As described in detail above for the high-precision perception driven automatic driving positioning method, those skilled in the art can clearly understand the high-precision perception driven automatic driving positioning system in the present embodiment. Therefore, for the sake of brevity of the specification, it will not be described in detail here. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part is described in the method part.

[0092] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0093] Obviously, many modifications and changes can be made to the application without departing from the spirit and scope of the application. It is understood that the application is not to be limited to the particular embodiments disclosed, but it is intended to cover all modifications which are within the scope of the application as defined by the language of the claims.

Claims

1. A high-precision perception-driven autonomous driving positioning method, characterized by: include: Collect vehicle environment data in real time through a multimodal sensor array to generate a multi-source perception dataset; Performing spatiotemporal synchronization processing on the multi-source perception data set to establish a fused perception data stream with a unified spatiotemporal reference; performing motion distortion correction on the fused perception data stream to generate dynamic compensation perception data; The dynamic compensation perception data is input into a multi-source fusion positioning model, and is combined with high-precision map data for three-dimensional space matching to obtain real-time positioning coordinates; Iteratively optimizing the real-time positioning coordinates through an adaptive error correction module to generate a vehicle positioning result; The vehicle environment data is collected in real time through a multimodal sensor array to generate a multi-source perception dataset, including: Configure the laser radar to scan the target area at a preset scanning frequency to obtain three-dimensional point cloud data; triggering a visual sensor to perform image acquisition according to the three-dimensional point cloud data to obtain image data of multiple regions; Control the millimeter-wave radar to sense the target area in multi-target tracking mode and obtain the velocity vector information of the moving object; Performing time stamp synchronization alignment on the three-dimensional point cloud data, the plurality of regional image data, and the velocity vector information to generate the multi-source perception data set; Performing time stamp synchronization alignment on the three-dimensional point cloud data, the plurality of regional image data, and the velocity vector information to generate a multi-source perception data set includes: Establishing a transformation matrix between the multimodal sensor array coordinate systems, mapping the lidar coordinate system to the vehicle body coordinate system according to the transformation matrix, and obtaining a first perception data set; Dynamically reprojecting image data of multiple regions of the visual sensor through a sliding window mechanism to obtain a second perception data set; Extract asynchronous sampling data based on millimeter-wave radar and perform interpolation compensation to obtain a third perception data set; Constructing a spatiotemporal consistency constraint condition to verify the geometric consistency of the first perception dataset, the second perception dataset, and the third perception dataset in the three-dimensional space. If the geometric data of the three-dimensional space are consistent, generating the multi-source perception dataset; The construction process of the multi-source fusion positioning model includes: Extracting laser radar geometric features based on the three-dimensional point cloud data, and extracting visual semantic features based on the multiple regional image data; Constructing a deep neural network model, and performing cross-modal feature fusion of the visual semantic features and the lidar geometric features through the deep neural network model to obtain visual-lidar data; Performing motion state estimation on the velocity vector information of the moving object according to an extended Kalman filter algorithm to obtain a motion state estimation result; An attention mechanism is introduced to dynamically adjust the visual and lidar data and the motion state estimation results for weight distribution to construct the multi-source fusion positioning model.

2. The high-precision perception-driven autonomous driving positioning method according to claim 1, characterized in that: Performing motion distortion correction on the fused perception data stream to generate dynamic compensation perception data includes: By performing inertial measurement on the target vehicle, the angular velocity data and linear acceleration data of the target vehicle are obtained; Performing motion compensation based on the angular velocity data and the linear acceleration data to obtain a vehicle posture change; Inverse motion compensation is performed on the three-dimensional point cloud data in combination with the vehicle posture change, and the point cloud trajectory of the dynamic obstacle is denoised to generate the dynamic compensation perception data.

3. The high-precision perception-driven autonomous driving positioning method according to claim 1, characterized in that: The real-time positioning coordinates are iteratively optimized by an adaptive error correction module to generate a vehicle positioning result, including: Extracting the road topology structure in the high-precision map and performing a matching evaluation with the real-time positioning coordinates to obtain a matching evaluation result; Performing particle filtering according to the matching evaluation result to obtain a particle filtering result, and dynamically adjusting the real-time positioning coordinates according to the particle filtering result to obtain a particle distribution density; Obtaining historical positioning trajectories of the target vehicle, performing batch nonlinear optimization on the historical positioning trajectories through a sliding window, and obtaining positioning optimized trajectory data; The positioning optimization trajectory data is combined with the particle distribution density to perform error calculation to obtain positioning error data. When the positioning error data exceeds a preset threshold, a positioning mode is triggered to reset the error and determine the vehicle positioning result.

4. The high-precision perception-driven autonomous driving positioning method according to claim 3, characterized in that: When the positioning error data exceeds a preset threshold, a positioning mode is triggered to reset the error and determine the vehicle positioning result, including: monitoring in real time the degree of matching between the real-time positioning coordinates output by the multi-source fusion positioning model and the high-precision map, determining the positioning error data based on the degree of matching, and judging whether the positioning error data exceeds a preset threshold; When the positioning error data exceeds the preset threshold, an error trigger signal is generated, multi-source data consistency verification is initiated based on the error trigger signal, and a trigger positioning mode is selected according to the verification result, where the positioning mode includes a tight combination positioning mode or a redundant positioning mode; Error correction is performed based on the tightly combined positioning mode or the redundant positioning mode, and the vehicle positioning result is generated according to the correction result combined with the real-time positioning coordinates and the spatial feature constraint relationship of the high-precision map.

5. High-precision perception-driven autonomous driving positioning system, characterized by: The steps for implementing the high-precision perception-driven autonomous driving positioning method according to any one of claims 1 to 4 are as follows: the high-precision perception-driven autonomous driving positioning system comprises: A data acquisition module is used to collect vehicle environment data in real time through a multimodal sensor array to generate a multi-source perception data set; A spatiotemporal synchronization processing module, configured to perform spatiotemporal synchronization processing on the multi-source perception data set to establish a fused perception data stream with a unified spatiotemporal reference; A motion distortion correction module, configured to perform motion distortion correction on the fused perception data stream to generate dynamic compensation perception data; A spatial matching module is used to input the dynamic compensation perception data into a multi-source fusion positioning model, combine it with high-precision map data to perform three-dimensional spatial matching, and obtain real-time positioning coordinates; The vehicle positioning module is used to iteratively optimize the real-time positioning coordinates through the adaptive error correction module to generate a vehicle positioning result.

Images