Unmanned environment sensing accurate positioning system and method

Abstract

The invention discloses an unmanned environment perception accurate positioning system and method, and relates to the technical field of positioning. Comprising the steps of segmenting a point cloud into a static layer point cloud set and a dynamic layer point cloud set; performing ground projection on the static layer point cloud set, extracting road marking geometric features and generating topological coding vectors; gNSS positioning data, IMU inertial data and the topological coding vector are synchronously received, and a confidence coefficient matrix is generated according to the signal-to-noise ratio of each data source; and performing projection matching on the topological coding vector and a marking coding matrix in a high-precision map, activating a visual weight enhancement mode of the confidence coefficient matrix when the matching similarity is greater than a matching threshold, and outputting a fusion positioning coordinate. According to the invention, matching verification is carried out on the topological coding vector and the marking information in the high-precision map, so that the positioning accuracy is further improved, the problem of poor adaptability of a traditional map is solved, and the positioning precision and robustness of the unmanned vehicle in a complex environment are improved.

Description

Technical Field

[0001] The present invention relates to the field of positioning technology, and in particular to an unmanned driving environment perception and precise positioning system and method. Background Art

[0002] With the continuous development of unmanned driving technology, the environmental perception and precise positioning of the autonomous driving system have become one of the key technologies to achieve safe and stable driving of unmanned vehicles. In the past few decades, with the advancement of sensor technology, technologies such as LiDAR, visual sensors, and GNSS (Global Navigation Satellite System) have been gradually applied to autonomous vehicles, providing a rich source of data for environmental perception. LiDAR can effectively perceive the three-dimensional structure of the surrounding environment through its high-precision point cloud data, especially in complex urban environments.

[0003] However, there are still some shortcomings in the existing technology. For example, although multi-sensor fusion technology can improve positioning accuracy, most of the existing methods rely on simple weighted fusion strategies and fail to effectively consider the signal-to-noise ratio and reliability of different sensor data in different environments, resulting in the accuracy of sensor data fusion being affected in some complex scenarios (such as high-dynamic scenarios or environmental occlusion). In addition, the existing lidar point cloud processing methods often only focus on static environment modeling, but lack effective distinction in the detection and processing of dynamic objects, making it difficult to accurately identify dynamic obstacles and static environmental features. Summary of the invention

[0004] In view of the fact that existing unmanned driving environment perception systems usually have problems such as unstable positioning accuracy and poor environmental adaptability, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to improve the positioning accuracy and robustness of unmanned vehicles in complex environments.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In the first aspect, the present invention provides an unmanned driving environment perception and precise positioning system, which includes: a dynamic point cloud stratification module, which receives the original point cloud data of the laser radar, divides the point cloud into a static layer point cloud set and a dynamic layer point cloud set, and outputs the static layer point cloud set; a marking feature encoding module, which performs ground projection on the static layer point cloud set, extracts the geometric features of the road markings and generates a topological coding vector, wherein the topological coding vector includes a triplet coding sequence of lane line type, curvature radius and virtual-to-real ratio; a confidence matrix generation module, which synchronously receives GNSS positioning data, IMU inertial data and the topological coding vector, and generates a confidence matrix according to the signal-to-noise ratio of each data source; a cross-modal verification module, which performs projection matching on the topological coding vector and the marking coding matrix in the high-precision map, activates the visual weight enhancement mode of the confidence matrix when the matching similarity is greater than the matching threshold, and outputs the fused positioning coordinates.

[0007] As a preferred solution of the unmanned driving environment perception and precise positioning system described in the present invention, the dynamic point cloud stratification module includes: receiving the original point cloud data of the laser radar, and performing time series segmentation on the original point cloud data of the laser radar; calculating the dynamic changes of the point cloud based on the difference analysis of the point cloud trajectory, and judging whether it belongs to a static scene or a dynamic object; if it belongs to a static scene, performing cluster analysis to identify the fixed structure in the environment, and outputting the distribution and boundary information of each object in the static point cloud; and outputting a static layer point cloud set.

[0008] As a preferred solution of the unmanned driving environment perception and precise positioning system described in the present invention, the marking feature encoding module includes: using a ground detection algorithm to identify ground points in a static point cloud; projecting all points in the static layer point cloud set onto the ground plane according to their spatial coordinates to obtain corresponding ground coordinates, thereby obtaining a ground projection point cloud set; extracting road markings from the ground projection point cloud set, and performing geometric feature analysis on the extracted road markings; generating a topological coding vector based on the extracted road marking geometric features, wherein the topological coding vector is composed of a triplet coding sequence of three features: lane line type, curvature radius, and virtual-to-real ratio.

[0009] As a preferred solution of the unmanned driving environment perception and precise positioning system described in the present invention, the geometric feature analysis of the extracted road markings includes: analyzing the shape of the markings to determine the lane line type; calculating the radius of curvature according to the degree of curvature of the markings; and calculating the ratio of the solid and dotted parts of the markings, that is, the ratio of the length of the continuous solid line to the dotted line in the markings.

[0010] As a preferred solution of the unmanned driving environment perception and precise positioning system described in the present invention, the confidence matrix generation module includes: a signal-to-noise ratio calculation unit, which is used to calculate the signal-to-noise ratio of each data source according to GNSS positioning data, IMU inertial data and visual data; an initial confidence matrix generation unit, which is used to generate an initial confidence matrix according to the calculated signal-to-noise ratio; an environmental correction unit, which is used to generate an environmental correction matrix based on a vehicle flow density factor and an environmental visibility factor, correct the initial confidence matrix, and obtain a confidence matrix.

[0011] As a preferred solution of the unmanned driving environment perception and precise positioning system described in the present invention, wherein: the initial confidence matrix is ​​weightedly calculated based on the signal-to-noise ratio and the importance of different data sources; according to the weight values, an initial confidence matrix is ​​generated, and the initial confidence matrix contains the confidence values ​​of each data source.

[0012] As a preferred solution of the unmanned driving environment perception and precise positioning system described in the present invention, the environment correction unit includes: the traffic flow density factor is the number of vehicles detected by a camera or radar / unit time; the environmental visibility factor is obtained through a light intensity sensor or meteorological data; an environment correction matrix is ​​generated according to a pre-stored environmental impact coefficient table; a Hadamard product operation is performed on the initial confidence matrix and the environment correction matrix, wherein the non-diagonal elements are updated to the coupling weights between sensors to obtain a corrected confidence matrix; the dynamic layer point cloud density is used as a real-time attenuation factor of the visual confidence, the confidence of the visual positioning is adjusted secondary, and the final confidence matrix is ​​output.

[0013] As a preferred solution of the unmanned driving environment perception and precise positioning system described in the present invention, the cross-modal verification module includes: according to the current position, direction and geometric characteristics of the vehicle's road markings, the topological coding vector is spatially projected and aligned with the marking coding matrix in the high-precision map, and the similarity is calculated; the calculated similarity is compared with the matching threshold, and when the matching similarity is greater than the matching threshold, the visual weight enhancement mode of the confidence matrix is ​​activated to enhance the visual data weight in the confidence matrix; using the enhanced visual weight, the positioning results of each data source are weightedly fused in the confidence matrix to obtain the final fused positioning coordinates as the current position of the vehicle.

[0014] In the second aspect, the present invention provides a method for precise positioning of unmanned driving environment perception, which includes: receiving raw point cloud data of a laser radar, dividing the point cloud into a static layer point cloud set and a dynamic layer point cloud set, and outputting the static layer point cloud set; performing ground projection on the static layer point cloud set, extracting geometric features of road markings and generating a topological coding vector, wherein the topological coding vector includes a triplet coding sequence of lane line type, curvature radius and virtual-to-real ratio; synchronously receiving GNSS positioning data, IMU inertial data and the topological coding vector, and generating a confidence matrix according to the signal-to-noise ratio of each data source; projecting and matching the topological coding vector with the marking coding matrix in the high-precision map, activating the visual weight enhancement mode of the confidence matrix when the matching similarity is greater than the matching threshold, and outputting fused positioning coordinates.

[0015] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the unmanned driving environment perception and precise positioning system as described in the first aspect of the present invention are implemented.

[0016] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, the steps of the unmanned driving environment perception and precise positioning system as described in the first aspect of the present invention are implemented.

[0017] The beneficial effects of the present invention are as follows: the present invention not only improves the system's modeling capability for static environments, but also makes the identification of dynamic obstacles more accurate by separating the static layer and the dynamic layer of the dynamic point cloud. In addition, by introducing a confidence matrix generation method with signal-to-noise ratio correction, the system can adaptively adjust weights according to the reliability of different sensor data, thereby avoiding overfitting in highly dynamic or complex environments. And through a cross-modal verification module, the topological coding vector is matched and verified with the marking information in the high-precision map, which further improves the accuracy of positioning, solves the problem of poor adaptability of traditional maps, and improves the positioning accuracy and robustness of unmanned vehicles in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 This is a structural diagram of the unmanned driving environment perception and precise positioning system.

[0020] Figure 2Structural diagram of the confidence matrix generation module for the unmanned driving environment perception and precise positioning system.

[0021] Figure 3 Schematic diagram of the precise positioning method for unmanned driving environment perception. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0025] As mentioned in the background technology above, although multi-sensor fusion technology can improve positioning accuracy, most existing methods rely on simple weighted fusion strategies and fail to effectively consider the signal-to-noise ratio and reliability of different sensor data in different environments, resulting in the accuracy of sensor data fusion being affected in some complex scenarios (such as high-dynamic scenarios or environmental occlusion). In addition, existing lidar point cloud processing methods often only focus on static environment modeling, but lack effective distinction in the detection and processing of dynamic objects, making it difficult to accurately identify dynamic obstacles and static environmental features.

[0026] In response to the above problems, the present invention improves the positioning accuracy and robustness of unmanned vehicles in complex environments through dynamic point cloud stratification, marking feature encoding and cross-modal verification, as follows.

[0027] Figure 1 : is a system block diagram of an unmanned driving environment perception and precise positioning system according to an embodiment of the present application. Figure 1As shown, in the unmanned driving environment perception and precise positioning system, it includes: a dynamic point cloud stratification module 100, which receives the original point cloud data of the laser radar, divides the point cloud into a static layer point cloud set and a dynamic layer point cloud set, and outputs the static layer point cloud set; a marking feature encoding module 200, which performs ground projection on the static layer point cloud set, extracts the geometric features of the road markings and generates a topological coding vector, wherein the topological coding vector includes a triplet coding sequence of lane type, curvature radius and virtual-to-real ratio; a confidence matrix generation module 300, which synchronously receives GNSS positioning data, IMU inertial data and the topological coding vector, and generates a confidence matrix according to the signal-to-noise ratio of each data source; a cross-modal verification module 400, which performs projection matching on the topological coding vector and the marking coding matrix in the high-precision map, activates the visual weight enhancement mode of the confidence matrix when the matching similarity is greater than the matching threshold, and outputs the fused positioning coordinates.

[0028] In an embodiment of the present application, the dynamic point cloud stratification module 100 is used to receive the original point cloud data of the laser radar, divide the point cloud into a static layer point cloud set and a dynamic layer point cloud set, and output the static layer point cloud set. It should be understood that the difference analysis method based on the point cloud trajectory is one of the key algorithms for environmental perception in unmanned driving systems. This method can not only distinguish between static and dynamic objects, but also provide stable dynamic object detection in complex real-time driving environments; the effectiveness and real-time performance of this method directly affect the safety and reliability of the unmanned driving system. Therefore, in the system implementation, the dynamic point cloud stratification module often needs to be fused with other sensor data (such as cameras, radars, etc.) to jointly improve the overall perception capability.

[0029] The original LiDAR point cloud data contains information about all three-dimensional points obtained by LiDAR scanning in an unmanned driving environment, and each point contains position coordinates and reflection intensity. According to the sensor timestamp information, the original point cloud data is divided into multiple time segments, and each time segment represents all point cloud data collected by LiDAR scanning within a time window.

[0030] Optionally, in order to improve the accuracy and reliability of point cloud data, point cloud data is usually fused with other sensors (such as cameras, radars, etc.), which can make up for the blind spots or errors of a single sensor (such as lidar).

[0031] The dynamic changes of the point cloud are calculated using motion detection algorithms (such as difference analysis based on point cloud trajectories). For each point cloud in a time segment, the motion differences between consecutive time segments are compared to determine whether the point belongs to a static scene or a dynamic object. Among them, the position changes of dynamic objects are obvious, reflecting the movement of objects such as vehicles and pedestrians; while static scenes have no obvious position changes, reflecting fixed structures such as buildings, trees, and roads in the environment. Based on this judgment, the point cloud data is divided into a static layer point cloud set and a dynamic layer point cloud set.

[0032] Exemplarily, using a motion detection algorithm (such as difference analysis based on point cloud trajectories) to calculate the dynamic changes of the point cloud includes the following steps: The original point cloud data of the LiDAR is preprocessed, including denoising, filtering and other operations. The denoising step can use methods such as voxel grid filtering to reduce redundant point cloud data to an accuracy suitable for subsequent processing; according to the data acquisition frequency of the LiDAR and the timestamp of the sensor, the original point cloud data is divided into multiple time segments (time windows), each of which contains all the point cloud data scanned at a certain moment; the point cloud data in each time segment is registered to ensure that the point clouds of each time segment are aligned in the same coordinate system.

[0033] Preferably, the ICP algorithm or other point cloud registration algorithms can be used for this processing to obtain globally consistent point cloud data.

[0034] Furthermore, after completing the point cloud registration, the difference analysis algorithm is used to compare the point cloud data between consecutive time segments. The position change can be measured by calculating the Euclidean distance between point clouds or using the point-to-point distance between point clouds. For the point cloud in each time segment, by comparing its position change with the point cloud in the previous time segment, it is determined whether the point cloud has moved significantly. If the position difference between a point cloud and consecutive time segments is determined, it is determined whether the point belongs to a dynamic object. Based on the difference analysis, dynamic objects and static objects are judged according to the set threshold. The point cloud position of dynamic objects changes greatly, reflecting the movement trajectory of the object, such as a moving vehicle or pedestrian; while the point cloud position of static objects changes little, reflecting the fixed structure in the environment, such as buildings, roads, etc. Using the dynamic detection results, the point cloud data is divided into dynamic layer point cloud set and static layer point cloud set.

[0035] The above methods have good adaptability to complex scenes (such as dynamic traffic, pedestrian-dense areas, etc.).

[0036] Furthermore, cluster analysis is performed on the static layer point cloud to identify the fixed structures in the environment.

[0037] For example, the purpose of cluster analysis is to group static point clouds by spatial position, identify individual static objects, such as buildings, roads, obstacles, etc., output the distribution and boundary information of each object in the static point cloud, and help the unmanned driving system better understand the static objects in the surrounding environment.

[0038] In the embodiment of the present application, the road marking feature encoding module 200 is used to perform ground projection on the static layer point cloud set, extract road marking geometric features and generate a topological encoding vector, wherein the topological encoding vector includes a triple encoding sequence of lane line type, curvature radius and virtual-real ratio, specifically including the following contents: Receive a static layer point cloud set, including static point cloud data output from the dynamic point cloud stratification module, representing the non-moving parts of the road environment (such as roads, buildings and other fixed objects); use a ground detection algorithm to identify ground points in the static point cloud; project all points in the static layer point cloud set onto the ground plane according to their spatial coordinates, that is, perform vertical projection, obtain points on the ground to eliminate the influence of height differences, and obtain a ground projection point cloud set; The road markings are extracted from the ground projection point cloud set, and the markings on the road are identified using a straight line fitting algorithm (such as the least squares method, Hough transform, etc.). The extracted markings are subjected to geometric feature analysis, mainly including: lane line type: the type is determined by analyzing the shape of the marking (solid line or dashed line); curvature radius: the curvature radius R is calculated according to the curvature of the marking, and the curvature radius can be obtained by fitting a quadratic curve to the fitted marking; virtual-real ratio: the virtual-real ratio is calculated according to the ratio of the solid and dashed parts of the marking, that is, the ratio of the length of the continuous solid line to the dashed line in the marking.

[0039] Furthermore, based on the extracted geometric features of the lane markings, a topological coding vector is generated. The vector is composed of a triplet coding sequence of three features: lane line type, curvature radius, and virtual-real ratio. Each feature coding can be performed in the following way: The lane line type is represented by a binary value (such as 1 for solid line and 0 for dotted line) or a classification number; the curvature radius is discretized into a value within a limited range, and a code is generated based on the specific value. For example, the curvature can be divided into different levels (such as large curvature, medium curvature, small curvature, etc.) by a certain threshold; the virtual-real ratio is discretized into several levels (such as high ratio, moderate ratio, low ratio), and the corresponding code is generated.

[0040] Exemplarily, the lane line type is represented by a binary value or a classification number: a solid line can be represented by a binary value of 1; a dotted line can be represented by a binary value of 0. Classification numbers can also be used to represent different lane line types, for example: a solid line is coded as 1, and a dotted line is coded as 2; the radius of curvature can be discretized by a certain threshold and divided into different levels. For example: a large curvature, that is, a large radius of curvature, indicates a relatively straight road, which can be represented by code 1; a medium curvature, that is, a moderate radius of curvature, indicates that the road has a certain bend, which can be represented by code 2; a small curvature, that is, a small radius of curvature, indicates that the road has a sharp bend, which can be represented by code 3. Among them, the discretization interval and coding of the curvature can be adjusted according to the actual curvature range and needs. The virtual-real ratio reflects the ratio of the dotted line to the solid line in the lane line, which can be discretized by dividing the ratio into several levels. The classification levels can be: high ratio, indicating that most lane lines are dotted lines, which is a high ratio and can be represented by code 1; moderate ratio, indicating that the ratio of dotted lines to solid lines in the lane lines is moderate, which can be represented by code 2; low ratio, indicating that most lane lines are solid lines, and the proportion of dotted lines is low, which can be represented by code 3.

[0041] Suppose the geometric features of lane lines are as follows: Lane line type: solid line → coded as 1; Curvature radius: medium curvature → coded as 2; Virtual-real ratio: moderate ratio → coded as 2; then the generated triplet coding sequence is: .

[0042] The above triplet encoding sequence provides a structured representation for the geometric features of each lane line. The specific discretization and encoding methods can be adjusted according to actual scenarios and requirements.

[0043] In this embodiment, using a ground detection algorithm to identify ground points in a static point cloud includes: randomly selecting point pairs from the point cloud and performing plane fitting, using a RANSAC algorithm to identify and extract a maximum number of ground points, and the specific steps are as follows: randomly selecting a certain number of points from the static point cloud set, assuming that these points belong to the ground plane; calculating the plane equation, and counting the points with the smallest variance on the plane equation (i.e., the points with the smallest vertical distance from the plane); updating the plane model through multiple iterations, and determining which points belong to the ground and which points do not according to a threshold.

[0044] Once the ground points are identified, the next step is to project all the points in the static point cloud onto the ground plane. This operation maps the spatial coordinates of each point onto the ground plane through vertical projection, eliminating the effects of height differences. This will yield point cloud data that is parallel to the ground plane, making it easier to analyze and process further.

[0045] In this embodiment, projecting all points in the static layer point cloud set onto the ground plane according to their spatial coordinates to obtain the ground projection point cloud set includes the following operations: For each point in the static point cloud Project it onto the plane by its distance from the ground plane (i.e. the perpendicular distance from the point to the plane); suppose the equation of the ground plane is ,in, is the normal vector of the plane, D is the offset; for each point New point projected onto the ground plane It can be calculated by the following formula: in, is the vertical coordinate of the projection point on the ground plane, and By keeping and The coordinates remain unchanged and projection is performed.

[0046] This process will eliminate the effects of different heights in the static point cloud (such as the height difference of buildings or other objects), ensuring that the point cloud data on the ground is more accurate and consistent. After vertical projection, all the point cloud data obtained is the ground projection point cloud set, which represents all points parallel to the ground plane in the static environment, eliminating the influence of height differences, and can be used for further analysis and processing.

[0047] It should be noted that in the embodiment of the present application, the road marking feature encoding module extracts the geometric features of the road markings and generates a topological coding vector by performing ground projection on the static layer point cloud set, thereby effectively improving the accuracy and robustness of the unmanned driving environment perception system. The static layer point cloud set contains the static point cloud data output from the dynamic point cloud layering module, which represents the immovable parts of the road environment (such as roads, buildings and other fixed objects). The road marking feature encoding module first uses the ground detection algorithm to identify the ground points in the static point cloud. The ground plane is fitted by the RANSAC algorithm, and the maximum number of ground points are identified, thereby eliminating the impact caused by height differences. Subsequently, all static point cloud data are projected onto the ground plane through vertical projection to obtain a ground projection point cloud set. This step provides accurate point cloud data for further analysis of the geometric features of road markings.

[0048] The topological coding vector generation method in this process has significant advantages: First, by discretizing the geometric features of the markings (such as lane type, curvature radius, virtual-to-real ratio, etc.), the complexity of data processing is effectively reduced, and it can be flexibly adjusted for different scenarios. The discretization of lane type, curvature radius, and virtual-to-real ratio not only makes the data format adapt to the system's processing capabilities, but also can encode the markings within a specific threshold range according to actual needs, thus providing a reliable and scalable representation method. Through this structured topological coding vector, the system can efficiently identify and match road markings.

[0049] In the embodiment of the present application, the confidence matrix generation module 300 is used to synchronously receive GNSS positioning data, IMU inertial data and the topological coding vector, and generate a confidence matrix according to the signal-to-noise ratio of each data source, such as Figure 2 As shown, it includes a signal-to-noise ratio calculation unit 301, which is used to calculate the signal-to-noise ratio of each data source according to the GNSS positioning data, the IMU inertial data and the visual data; an initial confidence matrix generation unit 302, which is used to generate an initial confidence matrix according to the calculated signal-to-noise ratio; and an environment correction unit 303, which is used to generate an environment correction matrix based on a vehicle flow density factor and an environment visibility factor, and correct the initial confidence matrix to obtain a confidence matrix.

[0050] In this embodiment, the signal-to-noise ratio calculation unit 301 specifically includes: GNSS positioning data provides the current vehicle position in the global coordinate system, including longitude, latitude and altitude information; IMU inertial data provides the vehicle's acceleration and angular velocity data for estimating the vehicle's motion state; topological coding vectors include lane line type, curvature radius and virtual-to-real ratio, providing geometric feature information of road markings.

[0051] Specifically, the calculation of GNSS signal-to-noise ratio is based on GNSS positioning accuracy, which is usually expressed as positioning error. The higher the positioning accuracy, the higher the signal-to-noise ratio. Generally, the calculation of GNSS signal-to-noise ratio can be performed according to the following steps: First, you need to obtain the accuracy of GNSS positioning, usually in express, The smaller it is, the smaller the positioning error is and the higher the accuracy is: in, is the error range of GNSS positioning.

[0052] IMU accuracy is generally determined by estimating the error range of the accelerometer and gyroscope. For example, the inertial measurement error may be expressed as It represents the standard deviation or deviation of the IMU output data. The signal-to-noise ratio formula can be expressed as: Furthermore, the visual signal-to-noise ratio is estimated by analyzing the stability and availability of the marking feature encoding vector in the image. The visual signal-to-noise ratio is usually adjusted based on the marking information in the image (such as lane line features). Specifically, the visual signal-to-noise ratio can be calculated based on the stability of the marking information in the image. For example, whether the marking is clearly visible, whether it is continuously stable in the image, etc. In practical applications, the signal-to-noise ratio can be adjusted based on the quality of the marking features. For example, a simple quality factor (such as the clarity index of the marking) can be used to adjust the visual signal-to-noise ratio, which can be expressed as: in, is the error range of line recognition in the visual image, is the image quality factor, which is used to indicate the stability and visibility of the marking features in the image.

[0053] The above three signal-to-noise ratios can be dynamically calculated through algorithms to adjust the system's perception and positioning capabilities according to different input data sources and accuracies.

[0054] In this embodiment, the initial confidence matrix generating unit 302 specifically includes: The initial confidence matrix is ​​calculated based on the signal-to-noise ratio and the importance weighting of different data sources, as follows: For each data source, the weight is determined according to the inverse of its signal-to-noise ratio; based on the determined weight value, an initial confidence matrix is ​​generated, which contains the confidence value of each data source, and is calculated as follows: in, , , are the weights of the corresponding data sources, i.e., the inverse of the corresponding signal-to-noise ratio; , , They are the values ​​of the corresponding data sources. The initial value of the non-diagonal elements is 0, indicating that there is no preset correlation between sensors.

[0055] In this embodiment, by combining the signal-to-noise ratio calculation and confidence matrix generation technology of multiple sensor data, the environmental perception accuracy and positioning accuracy of the unmanned driving system can be significantly improved. Specifically, the signal-to-noise ratio calculation unit dynamically adjusts the weights of each data source according to the signal quality of GNSS positioning data, IMU inertial data and topological coding vector. The calculation of the GNSS signal-to-noise ratio is based on the GNSS positioning error. The higher the positioning accuracy, the higher the signal-to-noise ratio, which can ensure the reliability of the GNSS data. The IMU data determines its signal-to-noise ratio by estimating the range of inertial error, so that the inertial navigation system can more accurately estimate the vehicle motion state during positioning. The signal-to-noise ratio of the visual signal is evaluated by analyzing the stability and availability of the marking features in the image, which improves the effectiveness of the image data in harsh environments. Therefore, the signal-to-noise ratio calculation unit can dynamically adjust the weight of different sensor data sources in positioning according to their stability.

[0056] In this embodiment, the environment correction unit 303 specifically includes: Matrix correction is performed based on the traffic density factor and the environmental visibility factor, including: The traffic density factor adjusts the lane-related part of the confidence matrix according to the changes in real-time traffic flow. When the traffic density is high, the confidence of the marking is low. The environmental visibility factor adjusts the overall confidence according to the current environmental visibility (such as haze, rain, snow and other weather conditions). The lower the visibility, the lower the confidence should be.

[0057] Set the traffic density factor and ambient visibility factor And generate the environmental correction matrix , whose value is adjusted according to the above two factors: ; The initial confidence matrix Environmental Correction Matrix Perform Hadamard product operation (i.e. element-by-element multiplication) to obtain the modified confidence matrix , where the off-diagonal elements are updated as the coupling weights between sensors, such as the collaborative credibility of GNSS and lidar or the error compensation relationship between vision and inertial navigation.

[0058] Preferably, when the point cloud density of the dynamic layer exceeds a threshold (e.g. >30%), the point cloud density of the dynamic layer is superimposed as a visual confidence attenuation factor to further adjust the confidence of visual positioning, for example: in, is the attenuation factor, and outputs the final confidence matrix.

[0059] It should be noted that the confidence matrix generation unit determines the importance of each data source in positioning through weighted calculation based on the signal-to-noise ratio of each sensor. Specifically, the generation of the initial confidence matrix assigns weights according to the inverse of the signal-to-noise ratio of each sensor, which enables high-precision data sources (such as high-quality GNSS positioning data) to play a greater role in positioning, while low-precision data sources (such as visual data with large noise) are appropriately suppressed.

[0060] The environmental correction unit further corrects the confidence matrix and adjusts the confidence of the relevant parts of the lane through the traffic density factor and the environmental visibility factor. When the traffic density is high, the confidence of the marking will be reduced, and when the visibility is poor, the overall positioning confidence will be reduced. This dynamic adjustment mechanism can effectively cope with environmental influences such as traffic complexity and bad weather, thereby maintaining the stability and accuracy of the system, and the Hadamard product operation further optimizes the calculation of confidence by multiplying the initial confidence matrix with the environmental correction matrix element by element.

[0061] In the embodiment of the present application, the cross-modal verification module 400 is used to perform projection matching on the topological coding vector and the marking coding matrix in the high-precision map, activate the visual weight enhancement mode of the confidence matrix when the matching similarity is greater than the matching threshold, and output the fused positioning coordinates, including the following contents: Receive the topological coding vector and the HD map marking coding matrix, where the topological coding vector contains information such as lane line type, curvature radius, and virtual-to-real ratio; the HD map marking coding matrix contains the geometric features and topological information of the road markings in the HD map, such as lane line type, position, curvature, etc. Perform projection matching on the topological coding vector and the marking coding matrix in the HD map, and align the two spatially based on the current position and direction of the vehicle and the geometric features of the road markings. During the projection matching process, consider the similarity of the topological information, and use matching metrics such as cosine similarity and Euclidean distance to calculate the similarity between the two.

[0062] The calculated similarity is compared with the matching threshold. If the similarity is greater than the matching threshold, the topological coding vector is considered to be successfully matched with the high-precision map line information, and the visual weight enhancement mode of the visual confidence matrix is ​​activated. When the match is successful, the visual weight enhancement mode in the confidence matrix is ​​activated, and the visual data weight in the confidence matrix is ​​enhanced according to the preset enhancement coefficient: in, is the original visual confidence; is the enhanced visual confidence, is the visual enhancement coefficient. The enhanced visual confidence enhances the importance of visual information in fusion positioning and improves the accuracy of visual positioning.

[0063] The enhanced visual weights are used to weight the positioning results of each data source in the confidence matrix. The final fused positioning coordinates are obtained as the current position of the vehicle through weighted averaging, Kalman filtering and other methods.

[0064] In an embodiment of the present application, the cross-modal verification module significantly improves the positioning accuracy of the unmanned driving system in a complex environment by projecting and matching the topological coding vector with the marking coding matrix in the high-precision map. The topological coding vector contains key information such as lane type, curvature radius and virtual-real ratio, while the marking coding matrix in the high-precision map contains the geometric features and topological information of the road markings, such as the position, shape and curvature of the lane lines. By accurately aligning the topological coding vector with the high-precision map marking coding matrix, the system can consider the similarity of topological information such as lane type and position, and quantify the matching degree between the two using methods such as cosine similarity and Euclidean distance. The calculation of matching similarity can compare the road marking features in the high-precision map according to the current position and direction of the vehicle, thereby ensuring that the topological coding vector and the marking information of the high-precision map are highly consistent. When the calculated similarity is greater than the preset matching threshold, the cross-modal verification module automatically activates the visual weight enhancement mode in the confidence matrix. This process enhances the importance of visual information in overall positioning by strengthening the weight of visual data. The enhanced visual confidence further enhances the role of visual data in multi-sensor fusion, especially when the HD map is successfully matched, the data of the visual sensor can better support the fusion positioning calculation. Through this mechanism, the system can improve the signal-to-noise ratio of the visual signal based on the matching results of the HD map and the topological coding vector, thereby effectively improving the accuracy of visual positioning.

[0065] Figure 3 Flow chart of the method for accurate positioning of unmanned driving environment perception according to an embodiment of the present application. Figure 3 As shown, the unmanned driving environment perception and precise positioning method includes: Receiving original point cloud data from a laser radar, dividing the point cloud into a static layer point cloud set and a dynamic layer point cloud set, and outputting the static layer point cloud set; Performing ground projection on the static layer point cloud set, extracting geometric features of road markings and generating a topological coding vector, wherein the topological coding vector includes a triple coding sequence of lane line type, curvature radius and virtual-real ratio; Synchronously receiving GNSS positioning data, IMU inertial data and the topological coding vector, and generating a confidence matrix according to the signal-to-noise ratio of each data source; The topological coding vector is projected and matched with the marking coding matrix in the high-precision map. When the matching similarity is greater than the matching threshold, the visual weight enhancement mode of the confidence matrix is ​​activated, and the fused positioning coordinates are output.

[0066] Here, those skilled in the art can understand that the specific operations of each step in the above-mentioned unmanned driving environment perception and precise positioning method have been referred to above. Figure 1 to Figure 2 The description of the unmanned driving environment perception precise positioning system has been introduced in detail, and therefore, its repeated description will be omitted.

[0067] This embodiment also provides a computer device, which is suitable for the unmanned driving environment perception and precise positioning system, including a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to realize the unmanned driving environment perception and precise positioning system proposed in the above embodiment.

[0068] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covered on the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0069] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, an unmanned driving environment perception and precise positioning system as proposed in the above embodiment is implemented.

[0070] In summary, the present invention not only improves the system's modeling capability for static environments, but also makes the identification of dynamic obstacles more accurate by separating the static layer and the dynamic layer of the dynamic point cloud. In addition, by introducing a confidence matrix generation method with signal-to-noise ratio correction, the system can adaptively adjust weights according to the reliability of different sensor data, thereby avoiding overfitting in highly dynamic or complex environments. And through a cross-modal verification module, the topological coding vector is matched and verified with the marking information in the high-precision map, which further improves the accuracy of positioning, solves the problem of poor adaptability of traditional maps, and improves the positioning accuracy and robustness of unmanned vehicles in complex environments.

[0071] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An unmanned driving environment perception and precise positioning system, characterized by: include: The dynamic point cloud stratification module receives the original point cloud data of the laser radar, divides the point cloud into a static layer point cloud set and a dynamic layer point cloud set, and outputs the static layer point cloud set; A road marking feature coding module is used to perform ground projection on the static layer point cloud set, extract road marking geometric features and generate a topological coding vector, wherein the topological coding vector includes a triple coding sequence of lane line type, curvature radius and virtual-real ratio; A confidence matrix generation module synchronously receives GNSS positioning data, IMU inertial data and the topological coding vector, and generates a confidence matrix according to the signal-to-noise ratio of each data source; The cross-modal verification module performs projection matching on the topological coding vector and the marking coding matrix in the high-precision map, activates the visual weight enhancement mode of the confidence matrix when the matching similarity is greater than the matching threshold, and outputs the fused positioning coordinates.

2. The unmanned driving environment perception and precise positioning system according to claim 1, characterized in that: The dynamic point cloud layering module includes: Receiving raw laser radar point cloud data, and performing time series segmentation on the raw laser radar point cloud data; Based on the difference analysis of point cloud trajectories, the dynamic changes of point clouds are calculated to determine whether it belongs to a static scene or a dynamic object. If it belongs to a static scene, cluster analysis is performed to identify the fixed structure in the environment and output the distribution and boundary information of each object in the static point cloud. Output static layer point cloud set.

3. The unmanned driving environment perception and precise positioning system according to claim 2, characterized in that: The marking feature encoding module includes: Use ground detection algorithms to identify ground points in static point clouds; Projecting all points in the static layer point cloud set onto the ground plane according to their spatial coordinates to obtain corresponding ground coordinates, thereby obtaining a ground projection point cloud set; Extracting road markings from the ground projection point cloud set, and performing geometric feature analysis on the extracted road markings; A topological coding vector is generated based on the extracted geometric features of the road markings. The topological coding vector is composed of a triplet coding sequence of three features: lane line type, curvature radius, and virtual-real ratio.

4. The unmanned driving environment perception and precise positioning system according to claim 3, characterized in that: The geometric feature analysis of the extracted road markings includes: Analyze the shape of the lane marking to determine the lane line type; Calculate the radius of curvature according to the curvature of the marking line; According to the ratio of the solid and dotted parts of the marking line, calculate the real-dotted ratio, that is, the ratio of the length of the continuous solid line to the dotted line in the marking line.

5. The unmanned driving environment perception and precise positioning system according to claim 4, characterized in that: The confidence matrix generation module includes: A signal-to-noise ratio calculation unit, used to calculate the signal-to-noise ratio of each data source based on GNSS positioning data, IMU inertial data and visual data; An initial confidence matrix generating unit, used for generating an initial confidence matrix according to the calculated signal-to-noise ratio; The environment correction unit is used to generate an environment correction matrix based on the vehicle flow density factor and the environment visibility factor, and to correct the initial confidence matrix to obtain the confidence matrix.

6. The unmanned driving environment perception and precise positioning system according to claim 5, characterized in that: The initial confidence matrix is ​​weightedly calculated based on the signal-to-noise ratio and the importance of different data sources; An initial confidence matrix is ​​generated according to the weight values, wherein the initial confidence matrix includes the confidence value of each data source.

7. The unmanned driving environment perception and precise positioning system according to claim 6, characterized in that: The environment correction unit comprises: The vehicle flow density factor is the number of vehicles detected by the camera or radar per unit time; The environmental visibility factor is obtained through a light intensity sensor or meteorological data; Generate an environmental correction matrix according to a pre-stored environmental impact coefficient table; Performing a Hadamard product operation on the initial confidence matrix and the environmental correction matrix, wherein the non-diagonal elements are updated to the inter-sensor coupling weights, to obtain a corrected confidence matrix; The density of the dynamic layer point cloud is used as a real-time attenuation factor of the visual confidence, the confidence of the visual positioning is adjusted secondary, and the final confidence matrix is ​​output.

8. The unmanned driving environment perception and precise positioning system according to claim 1, characterized in that: The cross-modal verification module includes: According to the current position and direction of the vehicle and the geometric features of the road markings, the topological coding vector is spatially aligned with the road marking coding matrix in the high-precision map, and the similarity is calculated; The calculated similarity is compared with a matching threshold, and when the matching similarity is greater than the matching threshold, a visual weight enhancement mode of the confidence matrix is ​​activated to enhance the visual data weight in the confidence matrix; Using the enhanced visual weights, the positioning results of each data source are weighted and fused in the confidence matrix to obtain the final fused positioning coordinates as the current position of the vehicle.

9. A method for accurate positioning of an unmanned driving environment perception, based on the unmanned driving environment perception accurate positioning system according to any one of claims 1 to 7, characterized in that: Also includes: Receiving original point cloud data from a laser radar, dividing the point cloud into a static layer point cloud set and a dynamic layer point cloud set, and outputting the static layer point cloud set; Performing ground projection on the static layer point cloud set, extracting geometric features of road markings and generating a topological coding vector, wherein the topological coding vector includes a triple coding sequence of lane line type, curvature radius and virtual-real ratio; Synchronously receiving GNSS positioning data, IMU inertial data and the topological coding vector, and generating a confidence matrix according to the signal-to-noise ratio of each data source; The topological coding vector is projected and matched with the marking coding matrix in the high-precision map. When the matching similarity is greater than the matching threshold, the visual weight enhancement mode of the confidence matrix is ​​activated, and the fused positioning coordinates are output.

Images