Point cloud data labeling method and related device

Abstract

The invention provides a point cloud data labeling method and a related device, and relates to the technical field of automatic labeling. And calculating the search radius of the current point cloud data frame by using the initial motion information of the target point in the previous point cloud data frame and the inter-frame time interval, and determining the first motion information of the target point in the search radius of the current point cloud data frame. Besides, second motion information of the target point in the current point cloud data frame can be predicted according to the state transition model, that is, when point cloud data labeling is carried out, two motion information determination modes are adopted; and subsequently, under the condition that the relative coordinate distance between the position coordinate in the first motion information and the position coordinate in the second motion information is smaller than a target distance threshold value, taking the first motion information of the target point in the current point cloud data frame as a labeling result of the target point in the current point cloud data frame. According to the method, the labeling result is obtained through mutual verification of the two motion information determination modes, and the accuracy of point cloud data labeling is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic labeling, in particular to a point cloud data labeling method and related device. BACKGROUND

[0002] With the rapid development of intelligent transportation and automatic driving technology, millimeter wave radar plays an increasingly important role in multi-sensor fusion perception systems due to its all-weather, anti-interference, and ability to obtain target speed information. However, compared with image and laser radar sensors, millimeter wave radar point clouds have characteristics such as sparsity, low resolution, and strong dispersion, making the training of target detection models based on them more dependent on accurate, continuous, and structured labeling data of millimeter wave radar point clouds.

[0003] Currently, the labeling method of millimeter wave radar point cloud data usually adopts a manual frame-by-frame labeling method. This method is prone to inconsistencies, omissions, and errors in labeling, which affect accuracy. SUMMARY

[0004] In view of the above problems, the present application provides a point cloud data labeling method and related device to improve the accuracy of point cloud data labeling. The specific scheme is as follows:

[0005] The first aspect of the present application provides a point cloud data labeling method, comprising:

[0006] Obtaining initial motion information of a target point in a previous point cloud data frame;

[0007] Using the initial motion information of the target point in the previous point cloud data frame, the inter-frame time interval between the current point cloud data frame and the previous point cloud data frame, calculating the search radius of the current point cloud data frame;

[0008] Determining the first motion information of the target point within the search radius of the current point cloud data frame;

[0009] Obtaining a state transition model for performing a motion trajectory prediction operation;

[0010] According to the state transition model, predicting the second motion information of the target point in the current point cloud data frame;

[0011] Calculating the coordinate relative distance between the position coordinates in the first motion information and the position coordinates in the second motion information;

[0012] In the case where the coordinate relative distance is less than a target distance threshold, the first motion information of the target point in the current point cloud data frame is taken as the labeling result of the target point in the current point cloud data frame.

[0013] In a possible implementation, the search radius of the current point cloud data frame is calculated based on initial motion information of the target point in the previous point cloud data frame, the current point cloud data frame, and an inter-frame time interval of the previous point cloud data frame, and the search radius of the current point cloud data frame is calculated by:

[0014] The inter-frame time interval is calculated based on a collection timestamp of the current point cloud data frame and a collection timestamp of the previous point cloud data frame.

[0015] The position offset of the target point is calculated based on the initial motion information of the target point in the previous point cloud data frame and the inter-frame time interval.

[0016] The first predicted position coordinate of the target point in the current point cloud data frame is predicted based on the position offset of the target point and a position coordinate in the initial motion information of the target point in the previous point cloud data frame.

[0017] The search radius of the current point cloud data frame corresponding to the first predicted position coordinate is determined.

[0018] In a possible implementation, the search radius of the current point cloud data frame corresponding to the first predicted position coordinate is determined by:

[0019] The absolute value of the difference between the first predicted position coordinate and the position coordinate in the initial motion information of the target point in the previous point cloud data frame is calculated.

[0020] The search radius of the current point cloud data frame is determined based on the sum of the absolute value and a preset value.

[0021] In a possible implementation, the first motion information of the target point in the search radius of the current point cloud data frame is determined by:

[0022] The initial value of the translation information is obtained by performing initial transformation estimation based on the position offset.

[0023] The rigid transformation matrix of the previous point cloud data frame and the current point cloud data frame is calculated based on the initial value of the translation information.

[0024] The second predicted position coordinate of the target point in the current point cloud data frame is obtained by performing transformation operation on the position coordinate in the initial motion information based on the rigid transformation matrix.

[0025] In a case where the relative distance between the first predicted position coordinate and the second predicted position coordinate is less than a preset distance threshold, the discrete point closest to the second predicted position coordinate is searched in the search radius with the second predicted position coordinate as the origin in the current point cloud data frame.

[0026] The motion information of the discrete point is determined as first motion information of the target point in the current point cloud data frame.

[0027] In a possible implementation, the position coordinates in the initial motion information are transformed using the rigid transformation matrix to obtain second predicted position coordinates of the target point in the current point cloud data frame, including:

[0028] The product of the rigid transformation matrix and the position coordinates in the initial motion information is taken as the second predicted position coordinates of the target point in the current point cloud data frame.

[0029] In a possible implementation, the second motion information of the target point in the current point cloud data frame is predicted according to the state transition model, including:

[0030] The state transition model is used to perform trajectory prediction on the initial motion information of the target point in the previous point cloud data frame to obtain the second motion information of the target point in the current point cloud data frame.

[0031] In a possible implementation, the coordinate relative distance between the position coordinates in the first motion information and the position coordinates in the second motion information is calculated, including:

[0032] The Euclidean distance between the position coordinates in the first motion information and the position coordinates in the second motion information is calculated.

[0033] The Euclidean distance is taken as the coordinate relative distance.

[0034] In a possible implementation, after the first motion information of the target point in the current point cloud data frame is taken as the labeled result of the target point in the current point cloud data frame, the method further includes:

[0035] The labeled result is converted into structured data, and the structured data is saved.

[0036] A data set including the structured data is constructed, and model training is performed using the data set.

[0037] The second aspect of the present application provides a point cloud data labeling device, including:

[0038] An information acquisition module is configured to acquire initial motion information of a target point in a previous point cloud data frame.

[0039] A radius calculation module is configured to calculate a search radius of a current point cloud data frame using the initial motion information of the target point in the previous point cloud data frame, the frame interval between the current point cloud data frame and the previous point cloud data frame.

[0040] an information determining module configured to determine first motion information of the target point within a search radius of the current point cloud data frame;

[0041] a model obtaining module configured to obtain a state transition model used for performing a motion trajectory prediction operation;

[0042] a prediction module configured to predict second motion information of the target point in the current point cloud data frame according to the state transition model;

[0043] a distance calculating module configured to calculate a coordinate relative distance between a position coordinate in the first motion information and a position coordinate in the second motion information;

[0044] a result determining module configured to, in a case where the coordinate relative distance is less than a target distance threshold, determine the first motion information of the target point in the current point cloud data frame as a labeling result of the target point in the current point cloud data frame.

[0045] A third aspect of the present application provides an electronic device, comprising at least one processor and a memory connected with the processor, wherein:

[0046] the memory is configured to store a computer program;

[0047] the processor is configured to execute the computer program, so that the electronic device can implement the point cloud data labeling method described above.

[0048] By means of the technical scheme, the application provides a point cloud data labeling method and a related device. In the application, when labeling point cloud data, the initial motion information of a target point in a previous point cloud data frame, the frame interval between a current point cloud data frame and the previous point cloud data frame, and the search radius of the current point cloud data frame are calculated to determine the first motion information of the target point in the search radius of the current point cloud data frame. In addition, a state transition model for performing a motion trajectory prediction operation is obtained, and the second motion information of the target point in the current point cloud data frame is predicted according to the state transition model. That is, when labeling point cloud data, two motion information determination methods are adopted, and the coordinate relative distance between the position coordinates in the first motion information and the position coordinates in the second motion information is calculated. If the coordinate relative distance is less than a target distance threshold, the first motion information of the target point in the current point cloud data frame is taken as the labeling result of the target point in the current point cloud data frame. That is, the labeling result is obtained through mutual verification of the two motion information determination methods, which can improve the accuracy of point cloud data labeling, and the automatic labeling process does not require human intervention, thereby avoiding problems such as inconsistent labeling of previous and subsequent frames, missing labeling, and incorrect labeling. BRIEF DESCRIPTION OF DRAWINGS

[0049] The above and other features, advantages, and aspects of the present disclosure will become more apparent by describing in detail the following specific embodiments thereof with reference made to the accompanying drawings. Throughout the drawings, the same or similar reference numerals are used to refer to the same or similar elements. It should be understood that the drawings are schematic and elements and features are not necessarily to scale.

[0050] Figure 1 A flowchart of a point cloud data labeling method provided by the application;

[0051] Figure 2 A flowchart of calculating a search radius provided by the application;

[0052] Figure 3 A flowchart of determining first motion information provided by the application;

[0053] Figure 4 A structural schematic diagram of a point cloud data labeling device provided by the application;

[0054] Figure 5 A structural schematic diagram of an electronic device provided by the application. DETAILED DESCRIPTION

[0055] The embodiments of the present application are described below with reference to the accompanying drawings. The terms used in the embodiment part of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0056] The embodiments of the present application are described below with reference to the accompanying drawings. Those skilled in the art can know that, with the development of technology and the emergence of new scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0057] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged under appropriate circumstances, and this is only a way of distinguishing the objects with the same attributes in the description of the embodiments of the present application. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, so that the processes, methods, systems, products or equipment containing a series of units do not have to be limited to those units, but can include other units not clearly listed or inherent to these processes, methods, products or equipment.

[0058] With the rapid development of intelligent transportation and automatic driving technology, security monitoring and other fields, millimeter wave radar plays an increasingly important role in multi-sensor fusion perception systems due to its all-weather, anti-interference, and ability to obtain target speed information. However, compared with image and laser radar sensors, millimeter wave radar point clouds have the characteristics of sparsity, low resolution, and strong dispersion, making the training of target detection models based on them more dependent on accurate, continuous, and structured labeling data of millimeter wave radar point clouds.

[0059] Currently, the millimeter wave radar point cloud data labeling method usually adopts manual frame-by-frame labeling method. This method is prone to inconsistencies between frames, omissions, and errors that affect accuracy. In addition, this method is difficult to adapt to the actual needs of generating massive training data. At the same time, this method is difficult to retain the motion information of the target in the time sequence, limiting the learning ability of the model to the dynamic behavior of the target. In addition, due to the sparsity of radar point cloud data, manual labeling is more difficult, and labeling personnel need to frequently compare frames and identify targets, resulting in a large workload, extremely low efficiency, high subjective error, and high cost.

[0060] To solve the problem of manual frame-by-frame labeling, the embodiment of the application introduces target tracking logic into the millimeter wave radar point cloud data labeling process. Specifically, the target is manually labeled in the initial labeling frame, and the GICP (Generalized Iterative Closest Point) algorithm (also known as GICP registration algorithm) and Kalman filtering are combined to realize automatic tracking of the target in consecutive frames. When the registration result of the GICP algorithm is highly consistent with the predicted trajectory of the Kalman filter, the system automatically generates labeling information for the target, including position, speed, and identifier ID, etc., realizing continuous labeling of the target across frames. This method not only greatly reduces the number of manual operations, but also makes the target labeling time-consistent, improves the labeling accuracy, and reduces the cost. In addition, the motion information of the target in the time sequence is preserved, which meets the actual needs of continuous target perception in the automatic driving system.

[0061] In addition, the embodiment of the application also integrates a visualization and manual correction interface to support user intervention and adjustment of the automatic labeling result. In the case of new targets or tracking failures, the user can manually add or modify target labeling in any frame to trigger a new tracking process, achieving efficient human-machine collaboration. This hybrid labeling mode combining artificial intelligence and manual correction takes into account both efficiency and accuracy, making the entire labeling process efficient, automatic, flexible, and reliable.

[0062] Based on the above, an embodiment of the application provides a point cloud data labeling method, which refers to Figure 1 may include:

[0063] S11, obtaining initial motion information of a target point in a previous point cloud data frame.

[0064] In this embodiment, when tracking the target, the labeling operation of the next point cloud data frame is mainly based on the target point in the previous point cloud data frame. If the next point cloud data frame is the current point cloud data frame, the previous point cloud data frame is one point cloud data frame before the current point cloud data frame.

[0065] In actual scenarios, the millimeter wave radar can periodically (i.e., at fixed time intervals) or at non-fixed time intervals collect point cloud data. In the visualization interface of the labeling system, the user can select any data frame in the point cloud data as the initial labeling frame. In an example, the user loads the millimeter wave radar point cloud data frame by frame through the system interface and selects a representative data frame as the initial labeling frame, where the representative can mean that the target first appears or includes clearly identifiable target points.

[0066] The user can select one or more target points of interest in the initial labeling frame by clicking, the system assigns a unique target ID to each labeled target point, and records the initial spatial position (i.e. position coordinates) and velocity information (such as velocity vector) of the target point, and sets the labeling state to "initialized", providing initialization conditions for subsequent tracking algorithms.

[0067] In a specific implementation, the system provides a graphical user interface (GUI) to display the point cloud data collected by the millimeter wave radar in time sequence frame by frame. Each frame of millimeter wave radar point cloud data represents the spatial distribution information at a time, containing multiple points, and each point has attributes such as position coordinates (such as spatial coordinates (x, y)), radar cross section (RCS), velocity information (including radial velocity and lateral velocity), and other optional attributes (such as target category label)

[0068] The user can use the "s" key and "d" key on the keyboard to play the previous and next data frame sequence to observe the changes in the point cloud data. When a target of interest (such as a pedestrian, vehicle, or obstacle) is first clearly visible or has typical characteristics, the frame is selected as the initial labeling frame.

[0069] In the initial labeling frame, the system displays the millimeter wave radar point cloud in three dimensions, and the user selects one or more target points in the initial labeling frame by left-clicking with the mouse. For each selected target point, the system automatically extracts and records the motion information of the target point, which specifically refers to the x coordinate, y coordinate, x-axis direction velocity, y-axis direction velocity, and timestamp of the target point.

[0070] The system assigns a globally unique ID to each selected target point, such as Target_001, Target_002, etc. If multiple target points belong to the same target, the system classifies these target points under the same ID.

[0071] The state of each newly labeled target point is set to "initialized", and subsequent tracking and multi-frame labeling will be based on this state for tracking updates or confirmation of changes.

[0072] In addition, the system uses color to indicate the selected target points to assist the user in confirming whether the operation is correct. If the user mistakenly selects a non-target point, the system supports the undo / reselect function.

[0073] If the initial labeling frame is the first point cloud data frame when labeling begins, the initial motion information of the target points in the initial labeling frame is the x coordinate, y coordinate, x-axis direction velocity, y-axis direction velocity, and timestamp described above.

[0074] If partial annotation has been performed, the previous point cloud data frame is the point cloud data frame before and immediately adjacent to the current point cloud data frame, and the target point in the previous point cloud data frame also has corresponding initial motion information.

[0075] In an implementation manner, the system maintains a "target point list", and traverses the annotated target point list. The motion information of each target point includes the following information:

[0076] a unique ID of the target point (such as Target_001);

[0077] a position coordinate P_prev=(x,y) in the previous point cloud data frame;

[0078] velocity information in the previous point cloud data frame, such as a velocity vector V_prev=(vx,vy);

[0079] a timestamp T_prev of the previous point cloud data frame.

[0080] S12, calculating a search radius of the current point cloud data frame by using the initial motion information of the target point in the previous point cloud data frame, the frame interval between the current point cloud data frame and the previous point cloud data frame.

[0081] In the embodiment of the application, after the system annotates the target point in the initial annotation frame, the position information and the velocity feature of the target point are combined to realize point cloud registration in the continuous frame by using the GICP algorithm. When the GICP algorithm is used, it is necessary to determine the search radius of the current point cloud data frame. The search radius refers to a coordinate point in the current point cloud data frame corresponding to the position coordinate in the initial motion information of the target point in the previous point cloud data frame as the center, and a search range determined by the search radius.

[0082] In specific implementation, key feature information for subsequent tracking is extracted from the annotated target point in the previous point cloud data frame (which can be the initial annotation frame or not), mainly including position coordinates and velocity information, to provide basic data for motion estimation and matching.

[0083] Then, according to the velocity information, position coordinates and frame interval of the target point, a spatial range in which the target point can appear in the current point cloud data frame is calculated, and a neighborhood radius threshold is set. The neighborhood radius threshold is the search radius, which is used to limit the matching range, reduce the search space, improve the efficiency and reduce the false matching.

[0084] S13, determining first motion information of the target point in the search radius of the current point cloud data frame.

[0085] In the embodiment, the GICP algorithm is performed on the point cloud data of the current point cloud data frame and the point cloud data of the previous point cloud data frame, accurate point cloud registration is realized, the point correspondence relationship between adjacent frames is established, the position of the target in the current point cloud data frame is predicted through the nearest neighbor algorithm, and then the first motion information of the target point in the search radius of the current point cloud data frame is found. The first motion information is the same as the initial motion information described above, including position coordinates and speed information.

[0086] S14, obtain a state transition model used for motion trajectory prediction operation.

[0087] In the embodiment, a Kalman filter is introduced, a state equation is established according to the position coordinates and speed information of the target point in the previous point cloud data frame, and the position of the target point in the current point cloud data frame is predicted through trajectory prediction operation.

[0088] When the Kalman filter is introduced, a Kalman filter model needs to be constructed, and the core of the Kalman filter model is a state transition model. When the state transition model is specifically constructed, system state variables x, y, v_x and v_y are defined, wherein (x, y) is the position coordinates and (v_x, v_y) is the speed. Considering the two-dimensional position and speed of the target point, a target state vector x = [x, y, v_x, v_y] is constructed.

[0089] Based on the state (position, speed) of the target point in multiple point cloud data frames before the current point cloud data frame, a uniform straight line motion model is used to establish a state transition model, wherein the interframe time interval of adjacent point cloud data frames is delta_t.

[0090] The state transition model includes a state transition matrix and an observation model. In the state transition model, a process noise covariance matrix and an observation noise covariance matrix R need to be set. When the state transition model is used for the first time, the process noise covariance matrix and the observation noise covariance matrix R can be set to the initial values configured in advance.

[0091] The state transition model is used to predict the motion trajectory, position and speed of the target point in the current point cloud data frame, and the prediction result is used for matching verification with the GICP result.

[0092] S15, predict second motion information of the target point in the current point cloud data frame according to the state transition model.

[0093] In the embodiment, the Kalman filter can be used for trajectory prediction, and the second motion information of the target point in the current point cloud data frame is output.

[0094] In specific implementation, according to the initial motion information of the target point in the previous point cloud data frame, the state transition model is used for trajectory prediction operation to obtain the second motion information of the target point in the current point cloud data frame.

[0095] In detail, the initial motion information of the target point in the previous point cloud data frame is brought into the state transition model to obtain the second motion information of the target point in the current point cloud data frame.

[0096] The second motion information includes a predicted position (P_x, P_y) and a predicted velocity (v_x, v_y), and the position of the target point in the current point cloud data frame is denoted as P_kal.

[0097] S16, calculate the coordinate relative distance between the position coordinates in the first motion information and the position coordinates in the second motion information.

[0098] In this embodiment, the GICP algorithm can find the specific position of the target point in the current point cloud data frame, and the Kalman filter predicts the predicted position of the target point in the current point cloud data frame through trajectory prediction. If the positions determined by the two algorithms are similar, it means that this tracking is successful, and the position of the target point in the current point cloud data frame is found. Since the GICP algorithm determines the real position of the target point, compared with the Kalman filter, the determined position is real, so the target point obtained by the GICP algorithm can be used as the labeling result of the current point cloud data frame and used to update the target state. If the positions determined by the two algorithms are far apart, it means that this tracking fails, and the automatic labeling process of the target is terminated.

[0099] In detail, the initial motion information of the target point in the previous point cloud data frame is brought into the state transition model to obtain the second motion information of the target point in the current point cloud data frame.

[0100] In one implementation manner, the Euclidean distance between the position coordinates in the first motion information and the position coordinates in the second motion information can be calculated, and the Euclidean distance is used as the coordinate relative distance.

[0101] In detail, the Euclidean distance between P_gicp and P_kal is calculated as follows:

[0102] In detail, the Euclidean distance between P_gicp and P_kal is calculated as follows:

[0103] D_gk =sqrt{(P_gx -P_kx)^2 + (P_gy - P_ky)^2}。

[0104] S17, in the case of the coordinate relative distance being less than the target distance threshold, taking the first motion information of the target point in the current point cloud data frame as the labeling result of the target point in the current point cloud data frame.

[0105] Specifically, the target distance threshold is set to 0.1 (unit: meter) according to actual configuration in an implementation manner.

[0106] If the Euclidean distance between the target point matched by the GICP algorithm and the target point position predicted by the Kalman filter is less than the target distance threshold, i.e., D_gk≤ε, it is considered that the registration result is reliable and the tracking is successful, and the first motion information of the target point in the current point cloud data frame obtained by the GICP algorithm is taken as the labeling result of the target point in the current point cloud data frame. Specifically, for each successfully tracked target, the system automatically extracts and generates labeling data, including target ID, position, speed and other information, and updates the target point state in the Kalman filter and sets the target point labeling state to "tracking success" to ensure the stability and reliability of the labeling result.

[0107] If D_gk>ε, it is determined that the matching fails, the system suspends the automatic labeling of the target point, and sets the labeling state to "lost" or "needs manual confirmation".

[0108] On the basis of any of the above embodiments, the system in the embodiments of the present application is further configured with a visualization and user correction function. Specifically, to meet the needs of manual correction and complex scene processing, the system also provides a matching visualization tool interface, which displays all target points and the current point cloud data frame in the labeling process in real time. When displaying, each point cloud data frame is displayed in the main view, and the point cloud data frame is initially displayed as a white point, indicating a non-target point. After obtaining the target point, the position of each target point in the current frame (x, y) plane is extracted, and the target point is identified by a red and larger volume point to distinguish from non-target points. The information panel in the upper left corner of the main view displays the current frame number, target number, and play state. In an implementation manner, if an abnormality occurs, such as target jumping, abnormal drift, or sudden increase in the number of the current frame, a "suggestion for user review" mark appears in the upper left corner of the main view.

[0109] In addition, the system also allows adding a target in any frame and starting a new round of automatic tracking, realizing an efficient human-computer collaborative labeling process. Specifically, a user can intuitively view the system labeling situation in the interface to assist in judging whether there is mislabeling or missing labeling. In addition, the system supports the user in quickly editing and supplementing the mislabeling and missing labeling. When the system detects the appearance of a new target or the interruption of tracking of an original target, the user can label again, at this time, the user can manually label, edit and supplement the target, the user input will be used as a new labeling result, the system automatically identifies the new target and starts a new tracking process, or continues to update the original target trajectory and automatically starts a new tracking process.

[0110] In an example, if there is an obviously unlabeled point in the current point cloud data frame, the point can be a new target, the user can click the mouse to select a point for labeling; if it is a new target, the system creates a new target data structure, initializes the trajectory, state, Kalman filter and the like of the target data structure. For all new targets, the system automatically starts the GICP registration and Kalman prediction process.

[0111] When a tracked target is lost in consecutive frames, the user can click the mouse to perform re-labeling. When a point is mislabeled as a target point, the user can select the point to perform a deletion operation.

[0112] In the embodiment, when labeling the point cloud data, the initial motion information of the target point in the previous point cloud data frame, the inter-frame time interval between the current point cloud data frame and the previous point cloud data frame are used to calculate a search radius of the current point cloud data frame, and first motion information of the target point in the search radius of the current point cloud data frame is determined. In addition, a state transition model used for a motion trajectory prediction operation is obtained, and second motion information of the target point in the current point cloud data frame is predicted according to the state transition model. That is, when labeling the point cloud data, two motion information determination manners are adopted, a coordinate relative distance between a position coordinate in the first motion information and a position coordinate in the second motion information is calculated, and in a case where the coordinate relative distance is less than a target distance threshold, the first motion information of the target point in the current point cloud data frame is taken as a labeling result of the target point in the current point cloud data frame. That is, the labeling result is obtained through mutual verification of the two motion information determination manners, which can improve the accuracy of the point cloud data labeling, and the automatic labeling process does not need human participation, and can avoid problems such as inconsistency between front and back frames, missing labeling and mislabeling in manual labeling.

[0113] In addition, the GICP point cloud registration and Kalman filter prediction are organically fused in the embodiment of the present application. The GICP establishes a point-to-point matching relationship locally in space, and the Kalman filter estimates the target state from the time dimension, and the two complement each other to enhance. Through a matching verification mechanism, false matches are effectively filtered, and the accuracy and robustness of tracking are improved.

[0114] In addition, the application can automatically start the target tracking process after the initial labeled frame is manually calibrated. The entire labeling process does not require frequent human intervention. The system can continuously predict, match and update based on historical labeling and target trajectory, and truly realize automated batch labeling. For large-scale millimeter wave radar point cloud dataset construction scenarios, such as autonomous driving systems, intelligent rail transit and industrial monitoring applications, it has high engineering practicability and significantly reduces labor costs and time overhead.

[0115] On the basis of any of the above embodiments, with reference to Figure 2 , the search radius of the current point cloud data frame can be calculated using the initial motion information of the target point in the previous point cloud data frame, the inter-frame time interval of the current point cloud data frame and the previous point cloud data frame, which can include:

[0116] S21, calculate the inter-frame time interval according to the acquisition time stamp of the current point cloud data frame and the acquisition time stamp of the previous point cloud data frame.

[0117] In this embodiment, when calculating the inter-frame time interval delat_t, the acquisition time stamp T_curr of the current point cloud data frame and the acquisition time stamp T_prev of the previous point cloud data frame can be obtained, and then:

[0118] delta_t = T_curr - T_prev.

[0119] In actual scenarios, if the millimeter wave radar collects point cloud data periodically, the delat_t of adjacent data frames is the same, and then delat_t only needs to be calculated once, without the need for repeated calculation. If the millimeter wave radar does not collect point cloud data periodically, the delat_t of adjacent data frames can be different, and then the delat_t of adjacent data frames needs to be calculated in real time.

[0120] S22, calculate the position offset of the target point using the initial motion information of the target point in the previous point cloud data frame and the inter-frame time interval.

[0121] In this embodiment, the initial motion information of the target point in the previous point cloud data frame includes:

[0122] The unique ID of the target point (such as Target_001)

[0123] The position coordinates P_prev = (x, y) in the previous point cloud data frame

[0124] The velocity information in the previous point cloud data frame, such as the velocity vector V_prev = (vx, vy)

[0125] The time stamp T_prev of the previous point cloud data frame.

[0126] Firstly, the position offset delta_position is calculated by using V_prev and delta_t, and the specific calculation formula is:

[0127] delta_position = V_prev x delta_t.

[0128] S23, according to the position offset of the target point and the position coordinate in the initial motion information of the target point in the previous point cloud data frame, the first predicted position coordinate of the target point in the current point cloud data frame is predicted.

[0129] Specifically, after the position offset delta_position is calculated, the position coordinate of the target point in the previous point cloud data frame is known, and the position offset of the target point between the previous point cloud data frame and the current point cloud data frame is known, the sum of the two is the first predicted position coordinate P_predicted of the target point in the current point cloud data frame, and the specific calculation formula is

[0130] P_predicted = P_prev + delta_position.

[0131] S24, determining the search radius of the current point cloud data frame corresponding to the first predicted position coordinate.

[0132] After the position coordinate of the target point in the previous point cloud data frame and the first predicted position coordinate of the target point in the current point cloud data frame are known, the approximate moving range of the target point can be known, so as to determine the search radius.

[0133] In an implementation manner, when the search radius is determined, the absolute value of the difference between the first predicted position coordinate and the position coordinate in the initial motion information of the target point in the previous point cloud data frame can be calculated, and the search radius r of the current point cloud data frame is determined based on the sum of the absolute value and a preset value, and the specific calculation formula is:

[0134] r = abs(P_predicted - P_prev) + 0.2.

[0135] Wherein, abs refers to the absolute value, and 0.2 refers to the preset value, which can be configured according to actual conditions.

[0136] After the search radius r is known, the coordinate point in the current point cloud data frame corresponding to the position coordinate of the target point in the previous point cloud data frame, i.e. the second predicted position coordinate, can be taken as the origin, and the search radius is taken as the search radius, so as to find the final position of the target point.

[0137] On the basis of the embodiment, refer to Figure 3determine first motion information of the target point in a current point cloud data frame within a search radius, including:

[0138] S31, performing initial transformation estimation based on the position offset to obtain an initial value of the translation information.

[0139] In a specific implementation, when the point cloud registration is performed based on the GICP algorithm, the previous point cloud data frame is taken as a reference frame of the GICP algorithm, denoted as P_ref, and the current point cloud data frame is taken as a target frame of the GICP algorithm, denoted as P_src.

[0140] The position offset delat_position predicted through the above steps is used as the initial transformation estimation to calculate the initial value of the translation information, so as to improve the convergence speed of the GICP.

[0141] In addition, the GICP algorithm parameters can be configured as follows: the maximum number of iterations is 50 times, the convergence threshold is 1e -6 The covariance matrix of the points is used to enhance the matching accuracy.

[0142] S32, calculating the rigid transformation matrix of the previous point cloud data frame and the current point cloud data frame by using the initial value of the translation information.

[0143] In this embodiment, the GICP algorithm is performed to obtain the rigid transformation matrix T between the current point cloud data frame and the previous point cloud data frame. The rigid transformation matrix T includes rotation and translation information. When the rigid transformation matrix T is calculated, the translation information used is the initial value of the translation information, that is, the new rotation and translation information are calculated by using the initial value of the translation information.

[0144] S33, performing transformation operation on the position coordinates in the initial motion information by using the rigid transformation matrix, to obtain the second predicted position coordinates of the target point in the current point cloud data frame.

[0145] In this embodiment, the position of the target point in the previous point cloud data frame is transformed by using the rigid transformation matrix T to predict the accurate position of the target point in the current point cloud data frame.

[0146] In an implementation manner, the product of the rigid transformation matrix and the position coordinates in the initial motion information can be taken as the second predicted position coordinates P_curr_estimated of the target point in the current point cloud data frame. The specific calculation formula is as follows:

[0147] P_curr_estimated = T * P_prev.

[0148] P_prev is the position coordinates of the target point in the previous point cloud data frame.

[0149] S34, judge whether the relative distance of the first predicted position coordinate and the second predicted position coordinate is less than a preset distance threshold; if yes, execute step S35.

[0150] Specifically, the first predicted position coordinate is calculated based on the position offset of the target point, and the second predicted position coordinate is predicted based on the GICP algorithm. If the relative distance of the first predicted position coordinate and the second predicted position coordinate is less than the preset distance threshold, it means that the prediction results of the two are similar, and it also means that the prediction result of the GICP algorithm is accurate, and the subsequent steps can be executed.

[0151] In an implementation manner, in a case where the relative distance of the first predicted position coordinate and the second predicted position coordinate is greater than or equal to the preset distance threshold, it is indicated that the mutual verification of the two predicted positions fails, and at this time, prompt information representing target point tracking failure is output.

[0152] S35, in the current point cloud data frame, a discrete point closest to the second predicted position coordinate is searched within a search radius with the second predicted position coordinate as the origin.

[0153] In the embodiment, a region can be obtained with the second predicted position coordinate as the origin and the search radius as the radius.

[0154] Then, P_curr_estimated and P_src are taken as inputs of the nearest neighbor algorithm, and the discrete point P_gicp closest to the second predicted position coordinate P_curr_estimated is searched in the region by using the nearest neighbor algorithm.

[0155] In the embodiment, the discrete point closest to the second predicted position coordinate needs to be searched by using the nearest neighbor algorithm, because the second predicted position coordinate predicted by the GICP algorithm may not be an actual discrete point in the current point cloud data frame, and therefore an actually existing discrete point closest to the second predicted position coordinate needs to be found.

[0156] S36, motion information of the discrete point is determined as first motion information of the target point in the current point cloud data frame.

[0157] In the embodiment, after the discrete point is found, the current point cloud data frame carries motion information, such as position coordinates and velocity, of the discrete point, and the motion information of the discrete point is the first motion information of the target point in the current point cloud data frame.

[0158] In the embodiment, the target point can be automatically tracked and labeled in continuous frames by using the GICP algorithm.

[0159] On the basis of any of the above embodiments, after taking the first motion information of the target point in the current point cloud data frame as the labeling result of the target point in the current point cloud data frame, the labeling result can be converted into structured data, the structured data is saved, a data set including the structured data is constructed, and a model training operation is performed using the data set.

[0160] Specifically, when performing model training, the data in the data set used is structured data, and therefore, in the embodiments of the present application, the labeling result needs to be converted into structured data and archived in chronological order, and the system finally forms a multi-target, multi-frame continuous labeling data set.

[0161] When converting into structured data, the data structure of the target in the current point cloud data frame is traversed, and targets with a state of "tracking success" are filtered out. For each successfully tracked target, the following key labeling elements are extracted:

[0162] ID: unique identifier of the target;

[0163] Position: (x, y) coordinates in the current frame;

[0164] Velocity: two-dimensional velocity vector (v_x, v_y);

[0165] Frame number: current frame index (e.g., frame_id = 128).

[0166] The labeling information of each target extracted is packaged into a unified formatted structure, the labeling file is saved in chronological order according to the frame, each frame is saved as an independent file in turn, and the naming format is, for example, / annotations / frame_000128.csv.

[0167] Subsequently, an index file such as index.csv can be created to record frame number, timestamp, file path, etc. information for quick retrieval.

[0168] In addition, the following standard data set folder structure can be constructed for model training, use and deployment:

[0169] / dataset_root /

[0170] ├── pointclouds / # original radar point cloud data frame

[0171] │ ├── 000128.bin

[0172] │ └──...

[0173] ├── annotations / # labeling information (by frame)

[0174] │ ├── frame_000128.csv

[0175] │ └──...

[0176] ├── index / # mapping files

[0177] │ └── index.csv

[0178] └── meta.yaml # dataset description (frame rate, field description, etc.)

[0179] After the above steps, the system finally forms a multi-target and multi-frame continuous annotation dataset for subsequent model training.

[0180] In summary, the present application combines target motion modeling and data-driven (i.e., GICP algorithm) point cloud automatic annotation method. By extracting the motion characteristics (such as speed) of the target in the millimeter wave radar point cloud, the system constructs a state transition model of the target state. Combined with Kalman filter prediction and GICP point cloud registration algorithm, it can automatically track and annotate the target points in consecutive frames. This method overcomes the low efficiency of traditional radar point cloud manual annotation and the difficulty of continuous tracking of targets, realizes continuous, multi-target, high-precision automatic annotation, significantly improves data production efficiency and annotation consistency, and can provide a dataset for subsequent model training.

[0181] In addition, the target ID and trajectory tracking logic are also included, providing a more time-sequential and consistent dataset for millimeter wave radar point cloud-based training tasks.

[0182] Compared with the traditional frame-by-frame static annotation method, the present application realizes dynamic annotation of point cloud data through target tracking, greatly improving the annotation efficiency and data consistency, while having good real-time performance and scalability. This technology is suitable for intelligent transportation, industrial monitoring, unmanned vehicle perception and other scenarios, can effectively reduce the data annotation cost and improve the model training quality, is an important support means for promoting the development of millimeter wave radar intelligent perception system, provides an innovative technical path for the construction of automatic annotation system and high-quality perception data platform, and provides strong support for the rapid iteration and engineering landing of radar perception models.

[0183] On the basis of the above embodiment of the point cloud data annotation method, another embodiment of the present application provides a point cloud data annotation device, which refers to Figure 4 may include:

[0184] An information acquisition module 11 is configured to acquire initial motion information of a target point in a previous point cloud data frame.

[0185] The radius calculation module 12 is configured to calculate a search radius of the current point cloud data frame by using initial motion information of the target point in the previous point cloud data frame, and a time interval between the current point cloud data frame and the previous point cloud data frame.

[0186] The information determination module 13 is configured to determine first motion information of the target point in the search radius of the current point cloud data frame.

[0187] The model acquisition module 14 is configured to acquire a state transition model used for performing the motion trajectory prediction operation.

[0188] The prediction module 15 is configured to predict second motion information of the target point in the current point cloud data frame according to the state transition model.

[0189] The distance calculation module 16 is configured to calculate a coordinate relative distance between the position coordinates in the first motion information and the position coordinates in the second motion information.

[0190] The result determination module 17 is configured to, in a case where the coordinate relative distance is less than a target distance threshold, determine the first motion information of the target point in the current point cloud data frame as a labeling result of the target point in the current point cloud data frame.

[0191] In an implementation manner, the radius calculation module 12 includes:

[0192] The interval calculation sub-module is configured to calculate the time interval between the current point cloud data frame and the previous point cloud data frame according to a collection time stamp of the current point cloud data frame and a collection time stamp of the previous point cloud data frame.

[0193] The offset calculation sub-module is configured to calculate a position offset of the target point by using the initial motion information of the target point in the previous point cloud data frame and the time interval.

[0194] The coordinate prediction sub-module is configured to predict a first predicted position coordinate of the target point in the current point cloud data frame according to the position offset of the target point and the position coordinate in the initial motion information of the target point in the previous point cloud data frame.

[0195] The radius determination sub-module is configured to determine the search radius of the current point cloud data frame corresponding to the first predicted position coordinate.

[0196] In an implementation manner, the radius determination sub-module is specifically configured to:

[0197] Calculate an absolute value of a difference between the first predicted position coordinate and the position coordinate in the initial motion information of the target point in the previous point cloud data frame, and determine the search radius of the current point cloud data frame based on a sum of the absolute value and a preset value.

[0198] In an implementation manner, the information determination module 13 includes:

[0199] An initial value calculation submodule is configured to perform initial transformation estimation based on the position offset to obtain initial translation information;

[0200] A matrix calculation submodule is configured to calculate a rigid transformation matrix between the previous point cloud data frame and the current point cloud data frame using the initial translation information.

[0201] A transformation submodule is configured to perform transformation operation on the position coordinates in the initial motion information using the rigid transformation matrix to obtain second predicted position coordinates of the target point in the current point cloud data frame.

[0202] A search submodule is configured to, in a case where the relative distance between the first predicted position coordinates and the second predicted position coordinates is less than a preset distance threshold, search for a discrete point closest to the second predicted position coordinates within a search radius with the second predicted position coordinates as the origin in the current point cloud data frame.

[0203] An information determination submodule is configured to determine the motion information of the discrete point as the first motion information of the target point in the current point cloud data frame.

[0204] In an implementation manner, the transformation submodule is specifically configured to:

[0205] The product of the rigid transformation matrix and the position coordinates in the initial motion information is taken as the second predicted position coordinates of the target point in the current point cloud data frame.

[0206] In an implementation manner, the prediction module 15 is specifically configured to:

[0207] According to the initial motion information of the target point in the previous point cloud data frame, trajectory prediction operation is performed using a state transition model to obtain second motion information of the target point in the current point cloud data frame.

[0208] In an implementation manner, the distance calculation module 16 is specifically configured to:

[0209] The Euclidean distance between the position coordinates in the first motion information and the position coordinates in the second motion information is calculated, and the Euclidean distance is taken as the coordinate relative distance.

[0210] In an implementation manner, the method further includes:

[0211] A data conversion module is configured to convert the labeling result into structured data and save the structured data.

[0212] A construction module is configured to construct a data set including the structured data, and perform model training operation using the data set.

[0213] In the embodiment, when the point cloud data is labeled, the initial motion information of the target point in the previous point cloud data frame, the frame interval between the current point cloud data frame and the previous point cloud data frame are used to calculate the search radius of the current point cloud data frame, and the first motion information of the target point in the search radius of the current point cloud data frame is determined. In addition, a state transition model used for the motion trajectory prediction operation is also obtained, and the second motion information of the target point in the current point cloud data frame is predicted according to the state transition model. That is, when the point cloud data is labeled, two motion information determination methods are adopted, and the coordinate relative distance between the position coordinates in the first motion information and the position coordinates in the second motion information is calculated subsequently. In the case where the coordinate relative distance is less than the target distance threshold, the first motion information of the target point in the current point cloud data frame is taken as the labeling result of the target point in the current point cloud data frame. That is, through mutual verification of the two motion information determination methods, the labeling result is obtained, which can improve the accuracy of the point cloud data labeling, and the automatic labeling process does not need human participation, and can avoid the problems of inconsistent labeling, missing labeling, wrong labeling and the like between the previous frame and the subsequent frame in the manual labeling.

[0214] It should be noted that the working processes of the various modules and sub-modules in the embodiment are described above, and will not be repeated here.

[0215] In the embodiment, an electronic device is also provided, which includes at least one processor and a memory connected with the processor, wherein:

[0216] The memory is used to store a computer program.

[0217] The processor is used to execute the computer program, so that the electronic device can implement the point cloud data labeling method described above.

[0218] Reference Figure 5 As shown in FIG. 1, a structure schematic diagram of an electronic device suitable for implementing the electronic device in the embodiment of the present application is shown. The electronic device in the embodiment of the present application can include but is not limited to a fixed terminal such as a mobile phone, a notebook computer, a PDA (Personal Digital Assistant), a PAD (Tablet Personal Computer), a desktop computer and the like. Figure 5 The electronic device shown is only an example, and should not bring any limitation to the functions and use range of the embodiment of the present application.

[0219] As Figure 5As shown, the electronic device can include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 601 that can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 602 or loaded into a random access memory (RAM) 603 from a storage device 608. In a state in which the electronic device is powered on, various programs and data required for operation of the electronic device are also stored in the RAM 603. The processing device 601, the ROM 602, and the RAM 603 are connected to each other through a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0220] Generally, the following devices can be connected to the I / O interface 605: input devices 606 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; output devices 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 608 including, for example, a memory card, a hard disk, etc.; and communication devices 609. The communication devices 609 can allow the electronic device to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 5 The electronic device is shown with various devices, but it should be understood that not all of the shown devices are required to be implemented or present. More or fewer devices can alternatively be implemented or present.

[0221] The embodiment of the present application also provides a computer program product comprising computer readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the point cloud data labeling methods provided by the embodiments of the present application.

[0222] The embodiment of the present application also provides a computer readable storage medium, which carries one or more computer programs, and when the one or more computer programs are executed by an electronic device, the electronic device can implement any of the point cloud data labeling methods provided by the embodiments of the present application.

[0223] In addition, it should be noted that the device embodiments described above are only schematic, wherein the units shown as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., they can be located in one place, or distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment. In addition, the connection relationship between the modules in the device embodiment provided by the present application indicates that there is a communication connection between them, which can be implemented as one or more communication buses or signal lines.

[0224] Those skilled in the art can clearly understand that the application can be implemented by means of software plus necessary universal hardware, of course, also can be implemented by special hardware including special integrated circuit, special CPU, special memory, special component, etc. Generally, the functions completed by computer program can be easily implemented by corresponding hardware, and the specific hardware structure for implementing the same function can be various, such as analog circuit, digital circuit or special circuit, etc. However, for the application, the software program implementation is the better embodiment. Based on such understanding, the technical solutions of the application can be embodied in the form of software product, which is stored in a readable storage medium, such as a floppy disk, a U disk, a mobile hard disk, a ROM, a RAM, a magnetic disk or an optical disk, etc., including a plurality of instructions for making a computer device (which can be a personal computer, a training device, or a network device, etc.) execute the methods described in various embodiments of the application.

[0225] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product.

[0226] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, training device or data center to another website, computer, training device or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be stored by a computer or a data storage device such as a training device, a data center, etc. integrated with one or more available media sets. The available media can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

Claims

1. A point cloud data labeling method, characterized in that, The method comprises the following steps: acquiring initial motion information of a target point in a previous point cloud data frame; calculating a search radius of a current point cloud data frame based on the initial motion information of the target point in the previous point cloud data frame, the current point cloud data frame and a frame interval between the previous point cloud data frame and the current point cloud data frame; determining first motion information of the target point in the search radius of the current point cloud data frame; acquiring a state transition model used for a motion trajectory prediction operation; predicting second motion information of the target point in the current point cloud data frame based on the state transition model; calculating a coordinate relative distance between a position coordinate in the first motion information and a position coordinate in the second motion information; in a case where the coordinate relative distance is less than a target distance threshold, taking the first motion information of the target point in the current point cloud data frame as a labeling result of the target point in the current point cloud data frame.

2. The point cloud data labeling method of claim 1, wherein, The method for calculating the search radius of the current point cloud data frame based on the initial motion information of the target point in the previous point cloud data frame, the current point cloud data frame and the frame interval between the previous point cloud data frame and the current point cloud data frame comprises the following steps: calculating a frame interval based on a collection time stamp of the current point cloud data frame and a collection time stamp of the previous point cloud data frame; calculating a position offset of the target point based on the initial motion information of the target point in the previous point cloud data frame and the frame interval; predicting a first predicted position coordinate of the target point in the current point cloud data frame based on the position offset of the target point and a position coordinate in the initial motion information of the target point in the previous point cloud data frame; determining a search radius of the current point cloud data frame corresponding to the first predicted position coordinate.

3. The point cloud data labeling method of claim 2, wherein, The method for determining the search radius of the current point cloud data frame corresponding to the first predicted position coordinate comprises the following steps: calculating an absolute value of a difference between the first predicted position coordinate and a position coordinate in the initial motion information of the target point in the previous point cloud data frame; determining the search radius of the current point cloud data frame based on a sum of the absolute value and a preset value.

4. The point cloud data labeling method of claim 2, wherein, The method for determining the first motion information of the target point in the search radius of the current point cloud data frame comprises the following steps: performing initial transformation estimation based on the position offset to obtain a translation information initial value; calculating a rigid transformation matrix of the previous point cloud data frame and the current point cloud data frame based on the translation information initial value; performing transformation operation on the position coordinate in the initial motion information by using the rigid transformation matrix to obtain a second predicted position coordinate of the target point in the current point cloud data frame; in a case where a relative distance between the first predicted position coordinate and the second predicted position coordinate is less than a preset distance threshold, searching for a discrete point closest to the second predicted position coordinate within the search radius in the current point cloud data frame with the second predicted position coordinate as an origin; determining motion information of the discrete point as the first motion information of the target point in the current point cloud data frame.

5. The point cloud data labeling method of claim 4, wherein, transforming the position coordinate in the initial motion information using the rigid transformation matrix to obtain a second predicted position coordinate of the target point in the current point cloud data frame, comprising: multiplying the rigid transformation matrix and the position coordinate in the initial motion information as the second predicted position coordinate of the target point in the current point cloud data frame.

6. The point cloud data labeling method of claim 1, wherein, predicting second motion information of the target point in the current point cloud data frame according to the state transition model, comprising: performing trajectory prediction operation on the target point in the current point cloud data frame according to the initial motion information of the target point in the previous point cloud data frame using the state transition model to obtain the second motion information of the target point in the current point cloud data frame.

7. The point cloud data labeling method of claim 1, wherein, calculating the coordinate relative distance of the position coordinate in the first motion information and the position coordinate in the second motion information, comprising: calculating the Euclidean distance of the position coordinate in the first motion information and the position coordinate in the second motion information; taking the Euclidean distance as the coordinate relative distance.

8. The point cloud data labeling method of claim 1, wherein, after taking the first motion information of the target point in the current point cloud data frame as the labeling result of the target point in the current point cloud data frame, further comprising: converting the labeling result into structured data and saving the structured data; constructing a data set including the structured data to perform model training operation using the data set. 9.A point cloud data labeling apparatus, characterized by comprising: comprising: an information acquisition module configured to acquire initial motion information of a target point in a previous point cloud data frame; a radius calculation module configured to calculate a search radius of a current point cloud data frame using the initial motion information of the target point in the previous point cloud data frame, the current point cloud data frame and a frame interval between the previous point cloud data frame and the current point cloud data frame; an information determination module configured to determine first motion information of the target point within the search radius of the current point cloud data frame; a model acquisition module configured to acquire a state transition model used for motion trajectory prediction operation; a prediction module configured to predict second motion information of the target point in the current point cloud data frame according to the state transition model; a distance calculation module configured to calculate a coordinate relative distance of a position coordinate in the first motion information and a position coordinate in the second motion information; a result determination module configured to take the first motion information of the target point in the current point cloud data frame as a labeling result of the target point in the current point cloud data frame in a case that the coordinate relative distance is less than a target distance threshold.

10. An electronic device, comprising: comprising at least one processor and a memory connected with the processor, wherein: the memory is configured to store a computer program; the processor is configured to execute the computer program to enable the electronic device to implement the point cloud data labeling method according to any one of claims 1 to 8.

Images