Laser target spot identification method, equipment and medium

Abstract

The embodiment of the invention discloses a laser target spot recognition method and device and a medium, belongs to the technical field of laser positioning, and solves the problem that the stability and accuracy are low when a laser target spot is recognized through a traditional method under complex working conditions. Comprising the following steps: acquiring laser data transmitted by a non-repeated scanning radar, and obtaining a local prediction initial value corresponding to the laser data based on imu data or wheel speed data; determining a local region point based on the local prediction initial value, and determining a local gradient value between the local prediction initial value and the local region point, so as to perform point cloud feature point extraction according to the local gradient value; determining neighbor feature points corresponding to the point cloud feature points, and constructing a point pair structure based on the point cloud feature points and the neighbor feature points; performing least square optimization on the point pair structure to obtain a transformation pose of the current point cloud relative to the historical point cloud; and constructing a local dense point cloud based on the preset time window and the transformation pose, and obtaining a target point coordinate based on the local dense point cloud.

Description

Technical Field

[0001] The present application relates to the field of laser positioning technology, and in particular to a laser target recognition method, device and medium. Background Art

[0002] In the field of multi-line LiDAR positioning, solutions that rely on prior maps are prone to failure or a sharp drop in accuracy in complex working conditions, such as degraded environments like long corridors, large scene changes, and severe occlusion. Therefore, methods such as placing landmarks, QR codes, magnetic tracks, and reflective columns are often used to enhance positioning reliability and accuracy. Reflective target positioning is a common method due to its highly reflective properties.

[0003] Traditional single-line LiDAR positioning requires reflective tape to be placed on the same plane as the radar, and the target coordinates are determined by taking the midpoint of the highest reflective point. However, multi-line LiDAR, due to its low vertical resolution, small number of horizontal points, and incident reflection angle loss, requires increasing the reflective tape size to ensure the number of scan points, and then obtaining the center coordinates through clustering.

[0004] However, the traditional single-line laser radar positioning method has strict requirements on the size and installation position of the reflective part, which increases the difficulty and cost of deployment. The target center coordinates of the multi-line laser radar positioning method are greatly affected by the calculation method and rely on subsequent optimization algorithm correction. This not only prolongs the calibration cycle, but also makes it difficult to ensure repeated positioning accuracy. Therefore, the existing technology has high cost and low stability and accuracy when identifying laser targets under complex working conditions. Summary of the Invention

[0005] The embodiments of the present application provide a laser target recognition method, device, and medium for solving the following technical problems: the existing technology has high costs and low stability and accuracy when recognizing laser targets under complex working conditions.

[0006] The embodiments of this application adopt the following technical solutions:

[0007] The present application provides a laser target recognition method, comprising: acquiring laser data transmitted by a non-repeating scanning radar, and obtaining a local prediction initial value corresponding to the laser data based on IMU data or wheel speed data; determining a local area point based on the local prediction initial value, and determining a local gradient value between the local prediction initial value and the local area point, so as to extract point cloud feature points based on the local gradient value; determining the nearest neighbor feature points corresponding to each point cloud feature point, and constructing a point pair structure based on the point cloud feature points and the nearest neighbor feature points; performing least squares optimization on the point pair structure to obtain a transformed pose of the current point cloud relative to the historical point cloud; and constructing a local dense point cloud based on a preset time window and the transformed pose, so as to obtain target coordinates based on the local dense point cloud.

[0008] The embodiment of the present application can accumulate point cloud density in a stationary state through a non-repetitive scanning radar, calculate the local prediction initial value in combination with the imu or wheel speed data, compensate for the point cloud distortion caused by the radar movement, and provide an initial pose close to the true value for the point cloud registration, reduce the number of registration iterations, and improve the calculation efficiency and accuracy. Secondly, a point pair structure is constructed based on the neighboring feature points of the point cloud feature points, and the local geometric similarity of the feature points is used to narrow the matching range and reduce the probability of mismatching. The structured point pairs provide a clear correspondence for subsequent pose optimization, making the calculation more targeted and improving the accuracy and stability of point cloud matching. Multi-frame point cloud data is collected through a preset time window, and combined with the transformed pose to unify them into the same coordinate system and superimpose them to form a local dense point cloud. The target coordinates are extracted based on this, which can improve the accuracy and reliability of target positioning.

[0009] In one implementation of the present application, the imu data or wheel speed data is used to obtain the local prediction initial value corresponding to the laser data, specifically including: aligning the time stamp with the imu data and wheel speed data based on the point cloud data corresponding to the laser data; performing rotation de-distortion processing on the point cloud data based on the imu data, and / or performing translation de-distortion processing on the point cloud data based on the wheel speed data; obtaining the posture transformation matrix by pre-integrating the imu data within the time interval of two frames of point clouds, and / or performing track calculation based on the wheel speed data to construct a posture transformation matrix to obtain the local prediction initial value according to the posture transformation matrix.

[0010] In one implementation of the present application, the local gradient value between the local prediction initial value and the local area point is determined to extract the point cloud feature points based on the local gradient value, specifically including: constructing a local neighborhood for each local prediction initial value, and determining the covariance matrix corresponding to the local neighborhood; performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues; determining the geometric feature description information corresponding to the local prediction value based on the eigenvalues; comparing the geometric feature description information with the preset threshold condition, and classifying the feature points based on the comparison result to obtain point cloud feature points; wherein, the point cloud feature points include plane points, line points and corner points.

[0011] In one implementation of the present application, the neighboring feature points corresponding to each point cloud feature point are determined, and a point pair structure is constructed based on the point cloud feature points and the neighboring feature points, specifically including: dividing the point cloud feature points into different feature point sets according to type, and constructing an ivox point cloud data structure based on each feature point set; based on the feature points corresponding to the current frame, selecting the corresponding ivox point cloud data structure for query to determine the neighboring feature points; and constructing a point pair structure based on the point cloud feature points and the neighboring feature points.

[0012] In one implementation of the present application, the point pair structure is least squares optimized to obtain the transformation posture of the current point cloud relative to the historical point cloud, specifically including: constructing a weighted error square sum objective function based on the point pair structure; parameterizing the transformation matrix to perform linear approximation processing on the objective function, solving the increment through an iterative optimization algorithm to update the estimated value; obtaining the initial transformation matrix through a random sampling consistency algorithm, and screening the internal points; imposing constraints on the transformation matrix; wherein the constraints include at least rotation matrix orthogonality constraints, translation range constraints, and time continuity constraints; when the preset threshold conditions are reached, terminating the iterative optimization to obtain the transformation posture of the current point cloud relative to the historical point cloud.

[0013] In one implementation of the present application, a local dense point cloud is constructed based on a preset time window and a transformed posture, specifically including: synchronously collecting original point cloud data and the local transformed posture at the corresponding moment based on the preset time window; determining the transformation matrix of the posture at other moments relative to the reference posture according to the reference posture corresponding to the last moment; converting all point clouds to the reference coordinate system through the corresponding transformation matrix for each point cloud; and superimposing the converted point clouds corresponding to the preset time window to obtain a local dense point cloud.

[0014] In one implementation of the present application, before constructing a local dense point cloud based on a preset time window and transformed posture, the method also includes: setting a time window threshold; when the amount of window data is greater than the time window threshold, triggering a batch deletion mechanism for old data to keep the time window length constant.

[0015] In one implementation of the present application, the target coordinates are obtained based on a local dense point cloud, specifically including: extracting a set of high-reflection points in the local dense point cloud based on preset high-reflection point screening conditions; obtaining the target coordinates by clustering based on the high-reflection point set; or performing edge segmentation based on the high-reflection point set to obtain the target coordinates.

[0016] The embodiment of the application provides a laser target point recognition device, comprising: at least one processor; and a memory connected with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: acquire laser data transmitted by a non-repeated scanning radar, and obtain a local prediction initial value corresponding to the laser data based on IMU data or wheel speed data; determine a local area point based on the local prediction initial value, and determine a local gradient value between the local prediction initial value and the local area point, to perform point cloud feature point extraction according to the local gradient value; determine respective neighboring feature points corresponding to each point cloud feature point, and construct a point pair structure based on the point cloud feature points and the neighboring feature points; perform least square optimization on the point pair structure to obtain a transformation pose of a current point cloud relative to a historical point cloud; and construct a local dense point cloud based on a preset time window and the transformation pose, to obtain a target point coordinate based on the local dense point cloud.

[0017] The embodiment of the application provides a non-volatile computer storage medium, which stores computer executable instructions, and the computer executable instructions are configured to: acquire laser data transmitted by a non-repeated scanning radar, and obtain a local prediction initial value corresponding to the laser data based on IMU data or wheel speed data; determine a local area point based on the local prediction initial value, and determine a local gradient value between the local prediction initial value and the local area point, to perform point cloud feature point extraction according to the local gradient value; determine respective neighboring feature points corresponding to each point cloud feature point, and construct a point pair structure based on the point cloud feature points and the neighboring feature points; perform least square optimization on the point pair structure to obtain a transformation pose of a current point cloud relative to a historical point cloud; and construct a local dense point cloud based on a preset time window and the transformation pose, to obtain a target point coordinate based on the local dense point cloud.

[0018] The above at least one technical scheme adopted by the embodiment of the application can achieve the following beneficial effects: the non-repeated scanning radar can accumulate point cloud density in a static state, the local prediction initial value is calculated in combination with IMU or wheel speed data, point cloud distortion caused by radar motion is compensated, an initial pose close to a true value is provided for point cloud registration, the number of registration iterations is reduced, and the calculation efficiency and accuracy are improved. Secondly, the point pair structure is constructed according to the neighboring feature points of the point cloud feature points, the matching range is reduced by using the local geometric similarity of the feature points, and the probability of false matching is reduced. The structured point pair provides an explicit corresponding relationship for subsequent pose optimization, the calculation is more targeted, and the accuracy and stability of point cloud matching are improved. Multi-frame point cloud data is collected by using the preset time window, and the multi-frame point cloud data is superimposed in the same coordinate system in combination with the transformation pose to form a local dense point cloud, and the target point coordinate is extracted based on the local dense point cloud, so that the accuracy and reliability of target point positioning are improved. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments described in the present application. For those skilled in the art, other drawings can be obtained based on these drawings without inventive work. In the drawings:

[0020] Figure 1 A flow chart of a laser target recognition method provided in an embodiment of the present application;

[0021] Figure 2 A schematic diagram of a laser target recognition process provided in an embodiment of the present application;

[0022] Figure 3 A schematic structural diagram of a laser target recognition device provided in an embodiment of the present application.

[0023] Reference numerals:

[0024] 200: Laser target recognition device, 201: Processor, 202: Memory. DETAILED DESCRIPTION

[0025] The embodiments of the present application provide a laser target recognition method, device, and medium.

[0026] In order to enable those skilled in the art to better understand the technical solutions in this application, the following will clearly and completely describe the technical solutions in the embodiments of this application in conjunction with the drawings in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0027] The technical solutions proposed in the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0028] Figure 1 A flow chart of a laser target recognition method provided in an embodiment of the present application is shown as follows: Figure 1 As shown, the laser target recognition method includes the following steps:

[0029] S101. Obtain laser data transmitted by a non-repeated scanning radar, and obtain a local prediction initial value corresponding to the laser data based on imu data or wheel speed data.

[0030] In one implementation of the present application, the surrounding environment is scanned by a non-repetitive scanning radar, and the laser data obtained by the scan is input into a local point cloud estimator.

[0031] Specifically, the embodiment of the present application uses a non-repetitive scanning radar that can scan the surrounding environment while stationary to form a local point cloud, and the density of the point cloud accumulates and increases over time. Traditional repetitive scanning radars are limited by their hardware structure, and the point cloud scanned when stationary does not change. Therefore, the accumulated point cloud under stationary conditions will not become a dense point cloud. However, processing based on dense point clouds can more accurately extract target features. Therefore, the embodiment of the present application chooses to use a non-repetitive scanning radar, uses a point cloud estimator to estimate local motion changes, and combines the point cloud data scanned during a period of time based on the results of the local motion estimation over a period of time to ultimately form a dense point cloud.

[0032] In one implementation of the present application, the point cloud data corresponding to the laser data is timestamped with the IMU data and wheel speed data. The point cloud data is subjected to rotational dedistortion based on the IMU data, and / or translational dedistortion based on the wheel speed data. A pose transformation matrix is ​​obtained by pre-integrating the IMU data within the time interval of two point cloud frames, and / or dead reckoning is performed based on the wheel speed data to construct a pose transformation matrix, thereby obtaining a local prediction initial value based on the pose transformation matrix.

[0033] Specifically, the embodiment of the present application first performs fine preprocessing on the point cloud, and the point cloud registration requires a predicted change initial value to achieve faster registration accuracy and registration speed. The embodiment of the present application can use imu data or wheel speed data to perform dedistortion processing of the point cloud rotational motion. In addition, the local predicted initial value before point cloud matching can be given by short-range imu pre-integration or wheel speed track calculation.

[0034] Furthermore, the embodiment of the present application first aligns the time stamps of the point cloud data corresponding to the laser data with the IMU data and wheel speed data to ensure the consistency of the sensor data in the time dimension. Subsequently, the angular velocity information in the IMU data is used to calculate the rotation matrix through integration to achieve rotational dedistortion of the point cloud data, or the translation amount and rotation angle are calculated based on the wheel speed data to complete translational dedistortion of the point cloud data and eliminate the point cloud distortion caused by device motion.

[0035] Furthermore, the IMU data within the time interval between two frames of point cloud is pre-integrated to obtain the rotation increment, velocity increment, and position increment, and to construct the pose transformation matrix. Alternatively, dead reckoning is performed based on the wheel speed data to obtain the cumulative translation and rotation, and then the pose transformation matrix is ​​constructed to determine the local prediction initial value, providing reliable initial parameters for subsequent point cloud registration.

[0036] The embodiments of the present application reduce the motion distortion of point cloud data through preprocessing, improve the accuracy of initial value prediction, and enhance the precision and efficiency of the point cloud processing algorithm.

[0037] S102 : determining a local area point based on the local prediction initial value, and determining a local gradient value between the local prediction initial value and the local area point, so as to extract feature points of the point cloud according to the local gradient value.

[0038] In one implementation of the present application, a local neighborhood is constructed for each local prediction initial value, and a covariance matrix corresponding to the local neighborhood is determined. The covariance matrix is ​​subjected to eigenvalue decomposition to obtain eigenvalues. Based on the eigenvalues, geometric feature description information corresponding to the local prediction value is determined. The geometric feature description information is compared with a preset threshold condition, and the feature points are classified based on the comparison results to obtain point cloud feature points; wherein, the point cloud feature points include plane points, line points, and corner points.

[0039] Specifically, for each local initial prediction value, a certain range is defined in the point cloud space with its corresponding point as the center, and a local neighborhood is constructed that includes the point and its surrounding adjacent points. Within the constructed local neighborhood, the covariance matrix of the local neighborhood point cloud is calculated based on the 3D coordinate information of each point in the neighborhood and the calculation formula of the covariance matrix. The covariance matrix can reflect the degree of dispersion of points in the local neighborhood in each dimension and the correlation between dimensions.

[0040] Furthermore, the calculated covariance matrix is ​​decomposed using an eigenvalue decomposition algorithm to obtain its corresponding eigenvalues. Based on the obtained eigenvalues, geometric feature description information such as surface change rate, linearity, flatness, and sphericity is determined. The description information in the embodiments of this application depicts the geometric morphology of the local point cloud from multiple dimensions, converting abstract eigenvalues ​​into concrete, judgment-enabling geometric information.

[0041] Furthermore, the calculated geometric feature description information is compared one by one with pre-set threshold conditions. If the geometric feature descriptor of a point meets the threshold conditions corresponding to a plane point, such as high flatness and low linearity, the point is classified as a plane point; if it meets the threshold conditions for a line point, such as high linearity and low flatness, it is classified as a line point; if it meets the threshold conditions for a corner point, such as large eigenvalues ​​and a high surface change rate, it is classified as a corner point. In this way, automatic classification of feature points is achieved, resulting in a point cloud feature point set containing plane points, line points, and corner points.

[0042] S103: Determine the neighboring feature points corresponding to each point cloud feature point, and construct a point pair structure based on the point cloud feature points and the neighboring feature points.

[0043] In one implementation of the present application, point cloud feature points are divided into different feature point sets by type, and an ivox point cloud data structure is constructed based on each feature point set. Based on the feature points corresponding to the current frame, the corresponding ivox point cloud data structure is selected for query to determine the neighboring feature points. A point pair structure is constructed based on the point cloud feature points and the neighboring feature points.

[0044] Specifically, based on the obtained point cloud feature point classification results, all feature points are classified one by one. For example, plane points are classified into one set, line points are classified into another set, and corner points are classified into another set.

[0045] Furthermore, for each partitioned feature point set, a separate ivox (Incremental Voxel Grid) point cloud data structure is constructed. During the construction process, the points in the feature point set are mapped to the corresponding voxel grid according to their three-dimensional coordinates. Each voxel records the feature point information falling within it, and an index relationship between voxels and points is established through a hash table to facilitate fast access and query. This structure can efficiently manage different types of feature points, supports dynamic updates, fast insertion and deletion operations on point cloud data, and can significantly reduce the search range when performing nearest neighbor queries, improving query efficiency.

[0046] Furthermore, for each feature point in the current frame, the feature type to which it belongs is first determined, and then the corresponding ivox point cloud data structure is selected based on the type. The hash table index and voxel grid characteristics of the ivox structure are used to quickly find the neighboring feature points that are closer to the feature point in the selected ivox structure. After determining the feature points of the current frame and their corresponding neighboring feature points, each feature point is paired one-to-one with its found neighboring feature points to construct a point pair structure. During the construction process, relevant information is recorded for each point pair, such as the coordinates of the two points, the feature type, the distance between the two points, etc. These point pair structures will serve as the basic data for subsequent operations such as point cloud registration and pose optimization. By analyzing the relationship between the point pairs, the transformed pose of the point cloud can be calculated more accurately, and accurate matching and fusion of the point cloud can be achieved.

[0047] S104. Perform least squares optimization on the point pair structure to obtain the transformed pose of the current point cloud relative to the historical point cloud.

[0048] In one implementation of the present application, a weighted square error sum objective function is constructed based on a point pair structure. The transformation matrix is ​​parameterized to perform linear approximation processing on the objective function, and the increment is solved by an iterative optimization algorithm to update the estimated value. The initial transformation matrix is ​​obtained by a random sampling consistency algorithm, and the inliers are screened. Constraints are imposed on the transformation matrix; wherein the constraints include at least rotation matrix orthogonality constraints, translation range constraints, and time continuity constraints. When a preset threshold condition is reached, the iterative optimization is terminated to obtain the transformed pose of the current point cloud relative to the historical point cloud.

[0049] Specifically, for the constructed point pair structure, a corresponding weight is assigned to each point pair. The weight can be determined based on the matching quality of the point pair, the point cloud density, or other relevant factors. Based on the distance error between two points in the point pair, a weighted error square sum objective function is constructed. To facilitate the optimization and solution of the objective function, the transformation matrix describing the point cloud transformation is represented using a specific parameterization method, such as Lie algebra, quaternion, etc. Through parameterization, the objective function can be linearly approximated, converting the originally complex nonlinear optimization problem into an approximate linear problem, thereby reducing the difficulty of solving it.

[0050] Furthermore, based on the linearized objective function, the transformation matrix increment is calculated through continuous iteration. After each iteration, the increment is used to update the current transformation matrix estimate, gradually adjusting the transformation matrix to continuously reduce the value of the objective function, thereby approaching the optimal transformation pose.

[0051] Furthermore, to improve the robustness of the optimization and avoid falling into local optima, a random sampling consensus algorithm is employed. A small number of point pairs are randomly selected from the point pair structure, and an initial transformation matrix is ​​calculated based on these points. This serves as the starting point for iterative optimization. This initial transformation matrix is ​​then used to evaluate all point pairs. Point pairs with errors within a certain range are identified as inliers, while outliers with larger errors are excluded. This reduces the interference of outliers on the optimization results and ensures that subsequent optimization processes are based on reliable data.

[0052] Further, in the optimization process, various constraint conditions are imposed on the transformation matrix to ensure its physical meaning and rationality. By constraining the rotation matrix to satisfy orthogonality, the correctness and stability of the rotation operation are ensured; by limiting the translation range, unreasonable large displacement is avoided; by introducing time continuity constraint, the transformation between adjacent frames is kept smooth, which conforms to the actual motion law. These constraint conditions provide a limiting framework for the optimization process, guiding the algorithm to search for the optimal solution in the direction consistent with the actual situation. The threshold conditions for terminating iteration are preset, such as the change of the objective function being less than a certain minimum value, the number of iterations reaching the upper limit, or the increment of the transformation matrix being less than a certain threshold. During the iterative optimization process, these conditions are monitored in real time, and when any of the preset threshold conditions is met, the iteration process is stopped, and the transformation matrix at this time is taken as the final result, i.e. the transformation pose of the current point cloud relative to the historical point cloud is obtained.

[0053] In S105, a local dense point cloud is constructed based on the preset time window and the transformation pose, so as to obtain the target point coordinates based on the local dense point cloud.

[0054] In an implementation manner of the present application, based on the preset time window, the original point cloud data and the local transformation pose at the corresponding time are synchronously acquired. According to the reference pose corresponding to the last time, the transformation matrix of the poses at other times relative to the reference pose is determined. Through the corresponding transformation matrix, all the point clouds are converted to the reference coordinate system. The converted point clouds corresponding to the preset time window are superimposed to obtain the local dense point cloud.

[0055] Specifically, according to the preset time window parameters, in the continuous acquisition process, the original point cloud data and the local transformation pose output by the point cloud estimator at each time point are synchronously recorded, and the time sequence consistency is ensured through timestamp alignment, thereby providing a time reference and a pose reference for subsequent multi-frame point cloud fusion.

[0056] Further, the pose at the last time in the time window is set as the global reference pose, the inverse operation and matrix multiplication are performed on the pose matrices at other times in the window, the transformation matrix of the pose at each time relative to the reference pose is calculated, the coordinate conversion relationship of the point clouds at each time to the reference coordinate system is established, and the spatial misalignment caused by the pose difference at different times is eliminated. For each frame of point cloud in the time window, each three-dimensional point in the point cloud is traversed, the transformation matrix at the corresponding time is applied for coordinate transformation, the point cloud is mapped from the original acquisition coordinate system to the reference coordinate system, the multi-frame point clouds are aligned under the same spatial reference, and the spatial consistency foundation is laid for subsequent superposition operation. The multi-frame point clouds converted to the reference coordinate system are superimposed, the points at the same spatial position are increased in density due to multi-frame accumulation, and finally the local dense point cloud containing rich environmental details is generated, thereby improving the description accuracy of the point cloud to the scene structure.

[0057] In one implementation of the present application, before constructing a local dense point cloud, the embodiment of the present application sets a time window threshold. When the amount of window data exceeds the time window threshold, a batch deletion mechanism for old data is triggered to keep the time window length constant.

[0058] Specifically, a time window threshold is pre-determined based on point cloud processing requirements and system resource constraints. This threshold can be defined based on time length or data volume. During data acquisition and processing, the data volume within the time window is monitored in real time, continuously tracking the amount or duration of raw point cloud data and corresponding pose data collected within the current window. By comparing the current window data volume with the pre-determined time window threshold, a decision is made as to whether to trigger data cleanup.

[0059] Furthermore, if the amount of data in a window exceeds the time window threshold, a batch deletion mechanism for old data is initiated. The oldest data collected in the time window is prioritized and deleted. After deletion, the amount of data in the time window is ensured to return to the threshold range, maintaining a constant window length to ensure the stability and real-time performance of the subsequent point cloud processing process.

[0060] In one implementation of the present application, a set of high-reflection points is extracted from a local dense point cloud based on preset high-reflection point screening conditions. Target coordinates are obtained by clustering the set of high-reflection points, or by performing edge segmentation on the set of high-reflection points to obtain target coordinates.

[0061] Specifically, based on the reflection intensity, geometric features and other attributes of each point in the point cloud data, combined with the pre-set high-reflection point screening conditions, all points in the local dense point cloud are screened one by one. Points that meet the screening conditions are selected to form a set of high-reflection points. In the clustering process, high-reflection points with similar spatial positions and similar features are divided into different clusters based on the set parameters such as the distance threshold and the minimum number of points. For each cluster, its centroid coordinates or weighted centroid coordinates are calculated as the coordinates of the target represented by the cluster. By clustering, discrete high-reflection points can be aggregated into targets with practical significance, adapting to scenarios with irregular target distribution and uncertain number, and improving the accuracy and robustness of target positioning.

[0062] Furthermore, for the set of high-reflection points, the local gradient of each point is calculated, and points with gradient magnitudes exceeding a threshold are extracted as edge points. Using edge segmentation, the set of high-reflection points is segmented into multiple independent regions. For each segmented region, its geometric center, contour centroid, or target coordinates corresponding to that region are determined by matching it with a preset target template.

[0063] Figure 2 This is a schematic diagram of a laser target recognition process provided in an embodiment of the present application. Figure 2As shown in the figure, during the laser target recognition process, the radar data, imu data and wheel speed data are input into the local point cloud estimator. The local point cloud estimator and the sliding window voxel are used to construct a local dense point cloud. The target coordinates are obtained by clustering or edge segmenting the local dense point cloud.

[0064] Figure 3 This is a schematic diagram of the structure of a laser target recognition device provided in an embodiment of the present application. Figure 3 As shown, the laser target recognition device 200 includes: at least one processor 201; and a memory 202 communicatively connected to the at least one processor 201; wherein the memory 202 stores instructions that can be executed by the at least one processor 201, and the instructions are executed by the at least one processor 201 so that the at least one processor 201 can: obtain laser data transmitted by a non-repetitive scanning radar, and obtain a local prediction initial value corresponding to the laser data based on the imu data or the wheel speed data; determine a local area point based on the local prediction initial value, and determine a local gradient value between the local prediction initial value and the local area point, so as to extract point cloud feature points according to the local gradient value; determine the neighboring feature points corresponding to each point cloud feature point, and construct a point pair structure based on the point cloud feature points and the neighboring feature points; perform least squares optimization on the point pair structure to obtain a transformed pose of the current point cloud relative to the historical point cloud; construct a local dense point cloud based on a preset time window and the transformed pose, so as to obtain the target coordinates based on the local dense point cloud.

[0065] A non-volatile computer storage medium provided in an embodiment of the present application stores computer-executable instructions, and the computer-executable instructions are configured to: obtain laser data transmitted by a non-repetitive scanning radar, and obtain a local prediction initial value corresponding to the laser data based on imu data or wheel speed data; determine a local area point based on the local prediction initial value, and determine a local gradient value between the local prediction initial value and the local area point, so as to extract point cloud feature points based on the local gradient value; determine the neighboring feature points corresponding to each point cloud feature point, and construct a point pair structure based on the point cloud feature points and the neighboring feature points; perform least squares optimization on the point pair structure to obtain a transformed pose of the current point cloud relative to the historical point cloud; and construct a local dense point cloud based on a preset time window and the transformed pose, so as to obtain target coordinates based on the local dense point cloud.

[0066] The various embodiments in this application are described in a progressive manner. Similar portions between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from the other embodiments. In particular, the device, apparatus, and non-volatile computer storage medium embodiments are generally similar to the method embodiments, so their descriptions are relatively simple. For relevant portions, refer to the descriptions of the method embodiments.

[0067] The above merely describes the embodiments of the present application, and is not intended to limit the present application. The embodiments of the present application can be variously changed and modified by those skilled in the art. Any modification or replacement does not cause the corresponding technical solution to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A laser target recognition method, characterized in that: The method comprises: Acquire laser data transmitted by a non-repeated scanning radar, and obtain a local prediction initial value corresponding to the laser data based on imu data or wheel speed data; Determining a local area point based on the local prediction initial value, and determining a local gradient value between the local prediction initial value and the local area point, so as to extract point cloud feature points according to the local gradient value; Determining neighboring feature points corresponding to each of the point cloud feature points, and constructing a point pair structure based on the point cloud feature points and the neighboring feature points; Performing least square optimization on the point pair structure to obtain a transformed pose of the current point cloud relative to the historical point cloud; Based on the preset time window and the transformed posture, a local dense point cloud is constructed to obtain the target coordinates based on the local dense point cloud.

2. A laser target recognition method according to claim 1, characterized in that: The obtaining of the local prediction initial value corresponding to the laser data based on the imu data or the wheel speed data specifically includes: Align the point cloud data corresponding to the laser data with the imu data and the wheel speed data by time stamp; performing rotational dedistortion processing on the point cloud data based on the imu data, and / or performing translational dedistortion processing on the point cloud data based on the wheel speed data; The pose transformation matrix is ​​obtained by pre-integrating the imu data within the time interval of two frames of point cloud, and / or performing track calculation based on the wheel speed data to construct a pose transformation matrix, so as to obtain the local prediction initial value according to the pose transformation matrix.

3. The laser target recognition method according to claim 1, characterized in that: Determining a local gradient value between the local prediction initial value and the local area point to extract point cloud feature points according to the local gradient value specifically includes: Constructing a local neighborhood for each of the local prediction initial values, and determining a covariance matrix corresponding to the local neighborhood; Performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues; Determining geometric feature description information corresponding to the local prediction value based on the feature value; The geometric feature description information is compared with a preset threshold condition, and the feature points are classified based on the comparison result to obtain the point cloud feature points; wherein the point cloud feature points include plane points, line points and corner points.

4. A laser target recognition method according to claim 1, characterized in that: The determining of the neighboring feature points corresponding to each of the point cloud feature points, and constructing a point pair structure based on the point cloud feature points and the neighboring feature points, specifically includes: Dividing the point cloud feature points into different feature point sets according to their types, and constructing ivox point cloud data structures based on each of the feature point sets; Based on the feature points corresponding to the current frame, the corresponding ivox point cloud data structure is selected for query to determine the neighboring feature points; The point pair structure is constructed based on the point cloud feature points and the neighboring feature points.

5. The laser target recognition method according to claim 1, characterized in that: The least squares optimization of the point pair structure is performed to obtain the transformed pose of the current point cloud relative to the historical point cloud, specifically including: Based on the point pair structure, constructing a weighted square error sum objective function; The transformation matrix is ​​parameterized to perform linear approximation processing on the objective function, and the increment is solved by an iterative optimization algorithm to update the estimated value; The initial transformation matrix is ​​obtained through the random sampling consistency algorithm, and the interior points are screened; Applying constraints to the transformation matrix; wherein the constraints include at least a rotation matrix orthogonality constraint, a translation range constraint, and a time continuity constraint; When the preset threshold condition is reached, the iterative optimization is terminated and the transformed pose of the current point cloud relative to the historical point cloud is obtained.

6. A laser target recognition method according to claim 1, characterized in that: The constructing of a local dense point cloud based on the preset time window and the transformed pose specifically includes: Based on the preset time window, synchronously collect the original point cloud data and the local transformed pose at the corresponding moment; Determine the transformation matrix of the postures at other times relative to the reference posture according to the reference posture corresponding to the last moment; For each point cloud, the corresponding transformation matrix is ​​used to transform all point clouds into the reference coordinate system; The converted point clouds corresponding to the preset time windows are superimposed to obtain the local dense point cloud.

7. The laser target recognition method according to claim 1, characterized in that: Before constructing the local dense point cloud based on the preset time window and the transformed pose, the method further includes: Set time window threshold; When the window data volume is greater than the time window threshold, the old data batch deletion mechanism is triggered to keep the time window length constant.

8. The laser target recognition method according to claim 1, characterized in that: Obtaining target coordinates based on the local dense point cloud specifically includes: Extracting a high-reflection point set from the local dense point cloud based on a preset high-reflection point screening condition; Based on the high reflection point set, obtaining the target point coordinates by clustering; Alternatively, edge segmentation is performed based on the high reflection point set to obtain the target point coordinates.

9. A laser target recognition device, characterized in that: The device comprises a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to execute the method according to any one of claims 1 to 8.

10. A non-volatile computer storage medium storing computer executable instructions, characterized in that: The computer executable instructions can execute the method according to any one of claims 1 to 8.

Images