A map building method and system based on millimeter-wave radar and open-source maps

Abstract

This invention discloses a map construction method and system based on millimeter-wave radar and open-source maps. It utilizes global location information provided by open street maps as a preliminary spatial reference, and combines millimeter-wave radar odometry information with temporal augmentation techniques to achieve robust pose estimation and map construction in dynamic urban environments. This addresses the problems of limited environmental representation and insufficient global consistency in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving, and more specifically, to a map building system based on millimeter-wave radar and open-source maps. Background Technology

[0002] In the field of autonomous driving, continuous mapping technology can provide a globally consistent representation of the environment, making it an important component of autonomous driving systems and robotics. However, currently widely used sensors have many limitations in dynamic urban environments. For example, cameras are sensitive to light and perform poorly under dynamic lighting conditions, making it difficult to cope with complex environmental changes in cities. LiDAR performance deteriorates in feature extraction under adverse weather conditions or when obstructed by other road users. Millimeter-wave radar offers superior long-range detection capabilities and stronger penetration. Millimeter-wave radar exhibits significant stability under adverse weather and dynamic obstruction conditions, providing more reliable feature representations for navigation in complex urban environments, which has led to its increasing attention in recent years. However, millimeter-wave radar also faces challenges, such as complex electrical signal structures, higher noise levels, and lower angular resolution. These characteristics make it prone to accumulating errors during pose estimation and feature map construction, affecting the generation of globally consistent trajectories and feature maps. Summary of the Invention

[0003] To address the limitations of limited environmental representation and insufficient global consistency in existing technologies, this invention proposes a map construction method and system based on millimeter-wave radar and open-source maps. Global location information provided by open street maps is used as a preliminary spatial reference. Then, odometry information from millimeter-wave radar is combined with temporal augmentation techniques to achieve robust pose estimation and map construction in dynamic urban environments.

[0004] The map construction method based on millimeter-wave radar and open-source maps designed in this invention includes the following steps:

[0005] S1, acquires millimeter-wave radar data and open street map data;

[0006] S2, performs Cartesian coordinate transformation on millimeter-wave radar data;

[0007] S3, perform inter-frame pose estimation on the coordinate-transformed data to obtain the local trajectory;

[0008] S4, registers the local trajectory with the open street map;

[0009] S5, combining local trajectories with open street maps for loop closure detection:

[0010] First, align the point clouds of adjacent frames;

[0011] The design uses a temporal density context, replacing the traditional occupancy grid with density information. It takes adjacent frames aligned with the input point cloud, calculates the density grid that each point cloud falls into, and obtains the current keyframe by calculating the row average of the density matrix.

[0012] In the global pose, historical keyframes that are spatially close to the current pose are selected as candidate keyframes. The cosine similarity between the temporal density context of the current keyframe and the candidate keyframes is calculated, and candidate keyframes with high similarity are selected for further loop closure detection.

[0013] S6 combines local trajectory, loop closure detection constraints, and global geometric constraints generated from the open street map to perform global trajectory optimization and generate a globally consistent radar point cloud map.

[0014] Preferably, the millimeter-wave radar data includes local environmental perception of object distribution and distance information in the current frame, and the open street map data includes road networks, building outlines, and terrain information.

[0015] Preferably, step S3 uses a random sampling consensus algorithm to remove noise points, then uses a nearest neighbor iterative point algorithm to align the point clouds of two adjacent frames, extract feature points, and perform inter-frame pose estimation.

[0016] Preferably, the specific process of step S4 includes:

[0017] The system calculates the distance between the trajectory points in each frame and the roads on the open street map, and updates the system pose by minimizing the error. It generates global constraints by matching the local trajectory generated by radar data with the open street map, and uses the constraints to correct the accumulated error. By continuously optimizing the pose constraints, the system keeps the trajectory consistent with the open street map.

[0018] Preferably, the specific process of obtaining the current keyframe pose in step S5 includes:

[0019] For an input point cloud from adjacent frames, first calculate which density grid each point falls into:

[0020]

[0021] in, The coordinates of the center point of the local feature point cloud are set to (0, 0, 0). and The indices represent the rings and sectors of the local feature point cloud; therefore, the maximum radar detection range can be expressed as...

[0022]

[0023] in and These represent the distance of a single ring and the number of rings, respectively, and the azimuth angle covered by each unit. It can be expressed as an equation

[0024]

[0025] in This represents the number of sectors. For each cell, the point density within the sector is encoded, and this encoding is normalized using the radar's range resolution as a weight.

[0026]

[0027] in and These represent the width of the radar ring and the range resolution, respectively. This represents the maximum number of feature points that a single ring can accommodate. Represents environmental weight. This indicates the number of points in each cell. It is the final density weight of each unit. The obtained temporal density context is then used to improve the efficiency of the corresponding search by calculating the row average of the density matrix, thereby obtaining the keyframe.

[0028] Preferably, the process of screening candidate keyframes in step S5 is as follows:

[0029] Set the threshold, select k The most similar historical keyframes are selected as candidate keyframes.

[0030] Before calculating each candidate keyframe n Frames before the current keyframe n The cosine similarity between frames yields a positive similarity sequence; similarly, the cosine similarity is calculated after each candidate keyframe. n Frames before the current frame n The cosine similarity between frames is used to obtain an inverse similarity sequence. The difference between these two similarity sequences is calculated to form a difference sequence. Finally, the difference sequence is weighted and summed using a Gaussian kernel to obtain the final symmetric similarity divergence. The frame with the largest deviation is selected as the final candidate frame, and a set threshold is used to filter out candidate frames with lower similarity scores.

[0031] Preferably, the state X of the robot in the pose optimization algorithm is optimized using odometry factor, loop closure factor, and adaptive map factor:

[0032]

[0033] in Represents the robot's state, framei and frame i-1 The odometry factor of the local pose relationship between them is denoted as ,frame i and frame j The cyclic factor of the long-term pose relationship between them is denoted as ,frame i The globally constrained adaptive map factor is denoted as ;

[0034] Among them, the adaptive map factor information matrix for:

[0035]

[0036] in This indicates the prior accuracy information of the open street map. An indicator representing translation error. Rotational error indicator.

[0037] Based on the same inventive concept, this invention also discloses a system utilizing a map construction method based on millimeter-wave radar and open-source maps, comprising:

[0038] The data acquisition module acquires local environmental perception data through millimeter-wave radar;

[0039] The data fusion module fuses the polar coordinates of the millimeter-wave radar with open street map data.

[0040] The pose estimation module performs inter-frame pose estimation on point cloud data in environmental perception data.

[0041] The map registration module registers the local trajectory with the open street map and ensures that the trajectory is consistent with the map.

[0042] The enhanced environmental feature description module uses the pose estimation results of consecutive radar frames to accurately align point cloud data of adjacent frames;

[0043] The design uses a temporal density context, replacing the traditional occupancy grid with density information. For an input point cloud from an adjacent frame, it first calculates which density grid each point falls into, and then improves the efficiency of the corresponding search by calculating the row average of the density matrix, thereby obtaining the key frame.

[0044] In the global pose, historical frames that are spatially close to the current pose are selected as detection candidates. The cosine similarity between the temporal density context of the current frame and the candidate frames is calculated, and candidate frames with high similarity are selected for further loop closure detection.

[0045] The map building module generates filtered and smoothed trajectories through a pose optimization algorithm on the backend, and outputs a globally consistent radar point cloud map.

[0046] Based on the same inventive concept, the present invention also discloses an electronic device, comprising:

[0047] One or more processors;

[0048] Storage device for storing one or more programs;

[0049] When one or more programs are executed by the one or more processors, the one or more processors implement the map construction method described above.

[0050] Based on the same inventive concept, the present invention also discloses a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the above-described map construction method. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the technical process of the present invention. Specific implementation methods

[0052] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] Terminology Explanation:

[0054] RANSAC (Sampling Consensus Algorithm): A robust model fitting algorithm used to estimate mathematical model parameters from data containing outliers.

[0055] ICP algorithm: A point cloud registration algorithm that achieves spatial alignment by iteratively minimizing the distance error between two point clouds.

[0056] ENU coordinates: A local rectangular coordinate system defined with a reference point as the origin: X-axis: pointing due east; Y-axis: pointing due north; Z-axis: perpendicular to the Earth's surface and upward, consistent with the direction of the normal to the Earth's ellipsoid.

[0057] World Coordinate System (LLA): A global geographic coordinate system based on the Earth ellipsoid model, such as WGS-84, defined as follows: Longitude: The angle east (0°~180°) or west (0°~-180°) of the Prime Meridian; Latitude: The angle north (0°~90°) or south (0°~-90°) of the equatorial plane; Altitude: The vertical elevation of a point relative to the reference ellipsoid (unit: meters).

[0058] Example 1

[0059] This embodiment discloses a map construction method based on millimeter-wave radar and open-source maps, including the following steps:

[0060] Step S1: Acquire millimeter-wave radar polar coordinate images and open street map data as input data. The millimeter-wave radar data provides the system with local environmental perception, including the distribution and distance information of objects in the current frame, which can be used to construct a short-term three-dimensional environmental feature map of the surrounding environment. The open street map data is used in the system to provide open-source map resources for global location information, including detailed road networks, building outlines, terrain and other global geographic information.

[0061] Step S2 involves transforming the radar data into Cartesian coordinates. The transformed data can better integrate radar point cloud data with open street map data, optimize data processing in modules such as pose estimation and map matching, and improve the system's computation and analysis performance in three-dimensional space.

[0062] Step S3: After coordinate transformation, the Random Sample Consensus Algorithm (RANSAC) is used to remove noise points. Then, the Nearest Neighbor Iterative Point Algorithm (ICP) is used to align the point clouds of two adjacent frames, extract feature points, and perform inter-frame pose estimation. The RANSAC algorithm fits the model by randomly selecting samples and removes noise points based on the model's conformity properties. By fitting and validating the model on different point cloud samples, it effectively reduces the occurrence of mismatches and effectively handles point cloud data containing a large amount of noise. After removing noise points, the ICP algorithm is used to further align the point clouds of two adjacent frames. It excels at local registration in environments with low noise, using the geometric relationships (distance, neighborhood) of points in the point cloud for registration. By minimizing the distance error between the point clouds of two frames, it can accurately estimate the pose change between the two frames. After RANSAC processing, ICP can be used for fine registration of the remaining feature points. This process improves the accuracy of inter-frame pose estimation. The formula is as follows:

[0063]

[0064] in, These are points in the source point cloud. It refers to a point in the target point cloud. It is a rotation matrix. It is a translation vector. Through continuous iterative optimization... and Minimizing the distance error between point pairs yields the optimal pose change estimate. Combining the pose estimates results gives the generated local trajectory.

[0065] Step S4 involves matching the local trajectory with the open street map. During operation, the system corrects the trajectory frame by frame and generates constraints. Specifically, it updates the system's pose by calculating the distance between the trajectory points in each frame and the roads in the open street map and minimizing the error. Global constraints are generated by matching the local trajectory generated from radar data with the open street map, and these constraints are used to correct accumulated errors. By continuously optimizing the pose constraints, the system addresses the drift problem that occurs in the trajectory generated after long-term operation, ensuring that the trajectory remains consistent with the open street map.

[0066] Step S5: Perform loop closure detection by combining the local trajectory with the open street map.

[0067] Millimeter-wave radar, due to its strong penetration and relatively low scan frame rate, often exhibits significant feature instability between adjacent frames. To mitigate this problem, this invention utilizes attitude estimation results provided by radar odometry to overlay feature point clouds and construct a temporal density context. This method enhances the ability of descriptors to describe and stabilize environmental features. First, pose estimation results from consecutive radar frames are used to accurately align point cloud data from adjacent frames, helping to mitigate the impact of feature instability. A local feature point cloud representing the current environmental segment is constructed:

[0068] in, and Indicates the current frame Aggregated point cloud and frame Point cloud, from frame to frame The transformation and sliding window size are respectively expressed as: and n When aligning and aggregating adjacent frames, a "short-term sliding window" needs to be selected to capture the relationship between them. This window stacks or accumulates several frames before and after the current frame to form a local subsequence. In this embodiment, the commonly used symbol for several frames is... n Short-term sliding windows are primarily used to stabilize feature extraction within the current time period, addressing the instability of millimeter-wave radar features and ultimately yielding a more robust keyframe. The keyframe generated here is equivalent to a more representative or stable frame on the timeline.

[0069] Due to the penetrating power of radar, multiple points may exist on the same scan line. Occupancy grids divide the environment into grid cells and encode the occupancy state of each cell. This method only considers a single occupancy state and cannot effectively describe the complex structure and multi-layered information in the environment. To address this issue, this invention designs a Short-Term Enhanced Temporal Density Context (TDC). Density information replaces the traditional occupancy grid, providing a more continuous and detailed description of radar data. For an input point cloud from adjacent frames, the first step is to calculate which density grid each point falls into:

[0070]

[0071] in, The coordinates of the center point of the local feature point cloud are set to (0, 0, 0). and The indices represent the rings and sectors of the local feature point cloud. Therefore, the maximum radar detection range can be expressed as...

[0072]

[0073] in and These represent the distance of a single ring and the number of rings, respectively, and the azimuth angle covered by each unit. It can be expressed as an equation

[0074]

[0075] in This indicates the number of sectors. For each cell, the point density within the sector is encoded, and this encoding is normalized using the radar's range resolution as weights.

[0076]

[0077] in and These represent the width of the radar ring and the range resolution, respectively. This represents the maximum number of feature points that a single ring can accommodate. Represents environmental weight. This indicates the number of points in each cell. It is the final density weight of each unit. The obtained temporal density context is then used to improve the efficiency of the corresponding search by calculating the row average of the density matrix, thereby obtaining the keyframe.

[0078] As a vehicle moves, it generates a series of continuous trajectory points in chronological order. Loop closure detection needs to consider this temporal continuity. That is, when a frame's trajectory is similar to a historical trajectory, its adjacent frames should also be consistent with the adjacent frames of the historical trajectory. Therefore, isolated loop closures are incorrect. For example, if the current frame... j With historical frames i Related, then frame j Adjacent frames should be with frame i The adjacent frames are related. Significant similarity is observed along the diagonal or anti-diagonal direction of the similarity matrix. Based on this observation, this invention proposes using long-term feature enhancement to further filter candidate loop closures. After generating the temporal density context using short-term loop closure enhancement, the global pose provided by the map matching algorithm is first used to select historical frames spatially close to the current pose as candidate frames for loop closure detection, which helps avoid the computational overhead problem caused by the increase in the number of historical frames. Subsequently, candidate frames are filtered by comparing the cosine similarity between the temporal density context of the current frame and the temporal density context of the candidate keyframes.

[0079]

[0080] in, For the temporal density context of the current frame (query frame), For the temporal density context of candidate keyframes (historical frames), For the current frame number The feature vector of the sub-segment For candidate frame number The feature vector of the sub-segment This represents the Euclidean norm of a vector.

[0081] If the calculated similarity exceeds a predetermined threshold, the candidate keyframe is retained.

[0082] After processing all candidate keyframes, select the frames most similar to the current frame from the remaining frames. k A number of candidate keyframes are selected to further remove loopback outliers during the long-term enhancement phase. To ensure the consistency of the selected frames across the time dimension, a length of [missing information] is introduced. n A sliding window is used to evaluate the feature similarity of these frames over several time intervals. First, the similarity before selecting candidate keyframes is calculated. n Frames before the current frame n The cosine similarity between frames yields a positive similarity sequence. Similarly, the similarity is then calculated after selecting candidate keyframes. n Frame and the frame after the current frame nCosine similarity between frames yields an inverse similarity sequence. A difference sequence is then obtained by comparing these two similarity sequences. Finally, a Gaussian kernel is applied to this difference sequence for weighted summation, resulting in the final symmetric similarity divergence (SSD). (Gaussian kernel) Use Gaussian function Normalize:

[0083]

[0084] in k This indicates the offset of the current frame. n This indicates the size of the sliding window.

[0085]

[0086] in and These represent the cosine similarity of forward and reverse similarity sequences, respectively. Symmetric similarity divergence. SSD The similarity bias was quantified, among which higher SSD This indicates a higher probability of a true loop closure, whether in the forward or reverse direction, and a lower probability of a true loop closure. SSD This suggests a potential false detection. Finally, based on the calculated... SSD The frame with the largest deviation is selected as the final candidate keyframe, and a threshold is used to filter out candidate keyframes with low similarity scores to further reduce the false detection rate of loop closure.

[0087] Once a loop closure is confirmed between a keyframe and candidate keyframes, it means that the current frame and its historical counterpart represent the same or similar geographical / environmental location. The system will input the relative transformation or pose error between them into the subsequent optimization process, that is: find similar frame pairs, treat them as loop closure pairs, add "loop closure edges / loop closure factors", and perform global pose graph or factor graph optimization.

[0088] In urban road scenarios, loop closure constraints are limited by the number of repeated regions along the trajectory, making it challenging to continuously correct accumulated odometry errors. Road networks in open street maps, due to their ease of acquisition and simple structure, can provide a coarse pose reference for the system. Therefore, this invention proposes an adaptive map constraint method. By treating the state estimation problem as a maximum a posteriori (MAP) estimation problem, three types of factors are defined to optimize the state X: odometry factor, loop closure factor, and adaptive map factor.

[0089]

[0090] in Represents the robot's state, frame i and frame i-1The odometry factor of the local pose relationship between them is denoted as ,frame i and frame j The cyclic factor of the long-term pose relationship between them is denoted as ,frame i The globally constrained adaptive map factor is denoted as .

[0091] Odometer factor

[0092] Feature extraction: Feature points are extracted from radar data using signal processing methods based on radar images.

[0093] Pose estimation: RANSAC and ICP are used to match feature points in adjacent frames to estimate their relative pose.

[0094] Establish odometry factor: Obtain odometry factor by calculating the relative pose between two frames.

[0095] Odometer factor The calculation is as follows:

[0096]

[0097] and It refers to the pose of the current frame and the previous frame. It is the pose transformation between two frames calculated using RANSAC and ICP.

[0098] Loop factor

[0099] Short-term enhancement: By constructing stable feature descriptions within a short time window using Temporal Density Context (TDC), the accuracy of loop closure detection is improved.

[0100] Long-term enhancement: Based on the temporal continuity of the trajectory, incorrect loop closure candidate frames are eliminated.

[0101] Loop closure detection: Calculate the similarity between the current frame and historical frames, determine whether loops exist, and establish constraints.

[0102] Loop factor calculation

[0103]

[0104] and The current frame and historical frames The position; It is the relative pose transformation obtained from loop closure detection.

[0105] Adaptive map factors

[0106] Map matching: Hidden Markov Model (HMM) is used to match radar trajectories with OSM to obtain a rough global pose.

[0107] Global coordinate transformation: Projecting the local trajectory onto the global coordinate system and aligning it with map information.

[0108] Adaptive map factor construction: Calculate the error between the current pose and the map matching pose, and generate map constraints.

[0109] Map matching error calculation

[0110]

[0111] It is the current pose calculated by the radar;

[0112] The matching pose is obtained through map matching.

[0113] Map factor residual calculation

[0114]

[0115] in: It is the information matrix, i.e., the inverse of the covariance matrix, used to measure error weights.

[0116] Open street map data is typically readily available in urban road scenarios. These maps provide comprehensive information on road structure and topological relationships, offering stable and reliable geometric constraints as effective coarse pose priors for continuous mapping. First, the starting point and orientation on the open street map are manually determined, and a coarse mapping between local and global coordinates is constructed based on the WGS84 ellipsoid. Next, pose transformation is used to convert the optimized trajectory within the time window to global coordinates. Then, a map matching algorithm based on a Hidden Markov Model is employed to establish the association between the local trajectory and the map. Finally, the trajectory is transformed back to local coordinates to establish pose constraints.

[0117]

[0118] in, This indicates the rotation of the odometer origin relative to the ENU coordinate system. and These represent the frames provided by the odometer in the local coordinate system. i The pose of the object and its corresponding position in the world coordinate system LLA. This represents the transformation from the world coordinate system (LLA) to the local coordinate system. The translational offset between the odometer origin and the ENU coordinates in the local coordinate system is used for coordinate system projection.

[0119]

[0120] Typically, the first frame of an odometer is considered the origin of the local coordinate system. Because... It relates only to the ellipsoidal surface and the projection reference line, and can be determined manually. and To determine the offset between the reference point and the odometer origin:

[0121]

[0122] Therefore, the world coordinate mapping for each subsequent odometry pose can be obtained in the following way:

[0123]

[0124] Current and historical odometry poses are input into the map matching module in world coordinates to obtain the matched poses:

[0125]

[0126] in This represents the map matching function. This indicates the number of historical frames that participated in the matching. This represents input from an open street map. The results are then converted to a local coordinate system for residual calculation and factorization.

[0127]

[0128] Error function for:

[0129]

[0130] in and Representing frames respectively i Pose in a local coordinate system and its corresponding map-matching pose, symbol This represents the Euclidean distance calculation between two poses. The residual function is:

[0131]

[0132] The level of observation error is represented using the inverse of the covariance matrix. Considering the limited accuracy of open street maps, and the additional errors introduced by map matching algorithms, communication delays, and initial point selection, a constant error model for open street maps is first established based on prior knowledge. Then, the distance between the current pose and the corresponding matching point, and the line from the current pose to the two nearest historical matching points are used... lThe distance is used as a reference for translation error. For rotation error, considering the orientation difference between the current pose and the corresponding matched pose, an adaptive error estimate is constructed:

[0133]

[0134] in Indicates the frame in the local coordinate system i and frame i-1 The line segment formed by the pose. Function Calculate the distance from a point to a line. An indicator representing translation error. The rotation error indicator is:

[0135] in A function that calculates the angle between two lines. An indicator representing rotation error. Open street map data is two-dimensional, in the adaptive map factor information matrix. Only local coordinates are calculated. Elements related to yaw angle are padded with zeros, while all other elements are padded with zeros. Therefore, the adaptive map factor information matrix... for:

[0136]

[0137] in This represents the prior accuracy information of the open street map. Based on the distance and orientation between trajectory points and matching points, a method is provided to evaluate the planar and heading errors of the matching points. It can adaptively evaluate the confidence of map constraints, thereby mitigating the impact of incorrect matches on the pose graph. As the output pose is continuously updated, map matching is performed on the updated pose to improve the accuracy of the coarse prior pose provided by the open street map.

[0138] Example 2

[0139] Based on the same inventive concept, this embodiment also discloses a system utilizing a map construction method based on millimeter-wave radar and open-source maps. The map construction method is detailed in Embodiment 1 and will not be repeated here. The system disclosed in this embodiment includes:

[0140] The data acquisition module acquires local environmental perception data through millimeter-wave radar;

[0141] The data fusion module fuses the polar coordinates of the millimeter-wave radar with open street map data.

[0142] The pose estimation module performs inter-frame pose estimation on point cloud data in environmental perception data.

[0143] The map registration module registers the local trajectory with the open street map and ensures that the trajectory is consistent with the map.

[0144] The enhanced environmental feature description module uses the pose estimation results of consecutive radar frames to accurately align point cloud data of adjacent frames;

[0145] The design uses a temporal density context, replacing the traditional occupancy grid with density information. For an input point cloud from an adjacent frame, it first calculates which density grid each point falls into, and then improves the efficiency of the corresponding search by calculating the row average of the density matrix, thereby obtaining the key frame.

[0146] In the global pose, historical frames that are spatially close to the current pose are selected as detection candidates. The cosine similarity between the temporal density context of the current frame and the candidate frames is calculated, and candidate frames with high similarity are selected for further loop closure detection.

[0147] The map building module generates filtered and smoothed trajectories through a pose optimization algorithm on the backend, and outputs a globally consistent radar point cloud map.

[0148] Example 3

[0149] Based on the same inventive concept, the present invention also provides an electronic device, including one or more processors; a storage device for storing one or more programs; and when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in Embodiment 1.

[0150] Since the device described in Embodiment 3 of this invention is the same electronic device used in implementing the map construction method based on millimeter-wave radar and open-source maps in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this electronic device based on the method described in Embodiment 1 of this invention, and therefore will not be described again here. All electronic devices used in any method of this invention's embodiments fall within the scope of protection of this invention.

[0151] Example 4

[0152] Based on the same inventive concept, the present invention also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the method described in Embodiment 1.

[0153] Since the device described in Embodiment 4 of this invention is a computer-readable medium used in implementing the map construction method based on millimeter-wave radar and open-source maps in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this electronic device based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All electronic devices used in any method of this invention's embodiments fall within the scope of protection of this invention.

[0154] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the content of this specification should be included within the protection scope of the present invention.

Claims

1. A map construction method based on millimeter wave radar and open source map, characterized by, The method comprises the following steps: S1, obtaining millimeter wave radar data and open street map data; S2, performing Cartesian coordinate conversion on the millimeter wave radar data; S3, performing inter-frame pose estimation on the data after coordinate conversion to obtain a local trajectory; S4, registering the local trajectory with the open street map; S5, combining the local trajectory with the open street map for loop closure detection: aligning point clouds of adjacent frames first; designing a time domain density context, using density information to replace the traditional occupancy grid, inputting adjacent frames of point cloud alignment, calculating the density grid into which each point cloud falls, and obtaining the current key frame by calculating the row average of the density matrix; selecting a historical key frame close to the current pose in space as a candidate key frame in the global pose, calculating the cosine similarity between the time domain density context of the current key frame and the candidate key frame, and screening out candidate key frames with higher similarity for loop closure detection; S6, combining the local trajectory, the loop closure detection constraint, and the global geometric constraint generated from the open street map to perform global trajectory optimization and generate a globally consistent radar point cloud map. 2.The map building method based on millimeter wave radar and open source map according to claim 1, characterized in that: The millimeter wave radar data includes local environment perception of object distribution and distance information in the current frame, and the open street map data includes road network, building outline, and terrain information. 3.The map building method based on millimeter wave radar and open source map according to claim 1, characterized in that: Step S3 uses a random sample consensus algorithm to remove noise points, and then aligns the point clouds of two adjacent frames through a nearest neighbor iterative point algorithm, extracts feature points, and performs inter-frame pose estimation. 4.The map building method based on millimeter wave radar and open source map according to claim 1, characterized in that: The specific process of step S4 includes: calculating the distance between the trajectory points of each frame and the road on the open street map, minimizing the error to update the pose of the system, generating global constraints by matching the local trajectory generated by the radar data with the open street map, and correcting the accumulated error using the constraints, and continuously optimizing the pose constraints to keep the trajectory consistent with the open street map. 5.The map building method based on millimeter wave radar and open source map according to claim 1, characterized in that: The specific process of obtaining the current key frame pose in step S5 includes: For input point clouds from adjacent frames, first calculate which density grid each point falls into: wherein, represents the coordinate of the center point of the local feature point cloud, which is set as (0, 0, 0), and represents the index of the ring and sector of the local feature point cloud, thus, the maximum radar detection distance can be represented as where and represent the width of the radar ring and the number of rings, respectively, covered by each cell in azimuth is expressed as the equation where represents the number of sectors, for each bin, encodes the point density within the sector, and normalizes this encoding using the range resolution of the radar as a weight: where and denote the width and range resolution of the radar ring, respectively, denotes the maximum number of feature points accommodated by a single ring, represents the environment weight, denotes the number of points in each cell, is the final density weight of each cell, is the resulting time-domain density context, finally, the efficiency of the corresponding search is enhanced by calculating the row average of the density matrix, thus obtaining the key frame. 6.The map building method based on millimeter wave radar and open source map according to claim 1, characterized in that: The process of screening candidate key frames in step S5 is as follows: setting a threshold, selecting k a most similar historical key frame as a candidate key frame; Before calculating each candidate keyframe n Frames before the current keyframe n The cosine similarity between frames yields a positive similarity sequence; similarly, the cosine similarity is calculated after each candidate keyframe. n Frames before the current frame n The cosine similarity between frames yields an inverse similarity sequence; calculate the difference between the two similarity sequences to form a difference sequence, and finally, use a Gaussian kernel to perform weighted summation on the difference sequence to obtain the final symmetric similarity divergence, select the frame with the largest deviation as the final candidate frame, and filter out candidate frames with lower similarity scores using a set threshold.

7. The map building method based on millimeter wave radar and open source map according to claim 1, characterized in that: The robot state X in the pose algorithm is optimized using an odometer factor, a loop factor, and an adaptive map factor: wherein represents the state of the robot, frame i and the local pose relationship between frame i-1 is denoted as , frame i and frame j is denoted as , frame i is denoted as ; wherein the adaptive map factor information matrix is: wherein represents a prior accuracy information of the open street map, represents an indicator of translation error, a rotation error indicator.

8. A system for map construction using the method of any one of claims 1-7, wherein: It includes: a data acquisition module that acquires local environment perception data through a millimeter wave radar; a data fusion module that fuses the polar coordinates of the millimeter wave radar with the open street map data; a pose estimation module that performs inter-frame pose estimation on point cloud data in the environment perception data; a map registration module that registers the local trajectory with the open street map and keeps the trajectory consistent with the map; an enhanced environment feature description module that uses the pose estimation results of consecutive frames of the radar to accurately align adjacent frame point cloud data; A time domain density context is designed to replace the traditional occupancy grid with density information. For the input point cloud from the adjacent frame, it is calculated that each point falls into which density grid. The efficiency of the corresponding search is enhanced by calculating the row average of the density matrix, so as to obtain the key frame; In the global pose, a historical frame close to the current pose in space is selected as a detection candidate. The cosine similarity between the time domain density context of the current frame and the candidate frame is calculated, and the candidate frame with high similarity is selected for loop detection; The map construction module generates a trajectory filtered and smoothed by a pose optimization algorithm of the back end, and outputs a radar point cloud map with global consistency.

9. An electronic device, comprising: Comprise: One or more processors; Storage devices for storing one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the map construction method as claimed in any one of claims 1-7.

10. A computer readable medium having stored thereon a computer program, characterized in that: The program is executed by the processor to implement the map construction method as claimed in any one of claims 1-7.

Images