Target trajectory prediction method and system based on radar detection, and computing device
By dynamically adjusting the downsampling and filtering order and combining it with artificial intelligence models, the problems of excessive data resource screening and loss of details in existing technologies are solved, and efficient trajectory prediction and risk assessment are achieved in different scenarios.
Patent Information
- Application Number
- CN202510825167.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-19
- Publication Date
- 2025-10-14
AI Technical Summary
After acquiring lidar data, existing technologies directly use the operation of filtering first and then downsampling, which leads to excessive screening of data resources, blurring of small structures, and accidental deletion of edge points, resulting in excessive loss of important details in the target vehicle's trajectory and prediction bias.
The real-time operating scenario of the target vehicle is determined based on the LiDAR data. A downsampling-then-filtering or filtering-then-downsampling strategy is adopted. Downsampling is performed first and then filtering in dense scenarios, while filtering is performed first and then downsampling in sparse scenarios. A trajectory prediction model is constructed in conjunction with an artificial intelligence model, and the processing order is dynamically adjusted to retain more effective details.
Ensure real-time performance in highly complex scenarios, improve trajectory prediction accuracy, provide risk assessment basis through multi-trajectory parameter output, and improve the safety and prediction accuracy of autonomous driving systems.
Smart Images

Figure CN120779416A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of radar detection and relates to radar signal processing technology, specifically a target trajectory prediction method, system and computing equipment based on radar detection. Background Art
[0002] When military radar acquires point cloud data through lidar, outliers are caused by multipath effects due to sensor noise and environmental reflections. Point clouds are sparse in long-range areas and dense in dynamic target areas. The recognition of some small edge structures depends on the local distribution of point clouds.
[0003] After existing military radars obtain point cloud data through lidar, the traditional process is to use a filtering followed by downsampling operation. This does not take into account the applicable scenarios of this filtering followed by downsampling operation, resulting in excessive screening of data resources, blurring of small structures, and mistaken deletion of edge points, which leads to excessive loss of important details in the predicted trajectory of the enemy's target vehicle, resulting in technical problems such as prediction bias.
[0004] The present invention provides a target trajectory prediction method, system and computing device based on radar detection to solve the above technical problems. Summary of the Invention
[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a target trajectory prediction method and system based on radar detection, which is used to solve the technical problems that the prior art directly adopts the operation of filtering first and then downsampling after obtaining lidar data, without considering the applicable scenarios of the operation of filtering first and then downsampling, resulting in excessive screening of data resources, blurring of small structures, and erroneous deletion of edge points, thereby leading to excessive loss of important details.
[0006] To achieve the above objectives, a first aspect of the present invention provides a target trajectory prediction method based on radar detection, comprising:
[0007] Collecting lidar data and environmental data of the target vehicle using a lidar sensor; converting the lidar data into initial point cloud data, wherein the environmental data includes: real-time environmental visibility;
[0008] Determine whether the real-time operating scenario of the target vehicle is a preset dense scenario; if yes, downsample and then filter the initial point cloud data; if no, filter and then downsample the initial point cloud data to obtain preprocessed point cloud data;
[0009] The preprocessed point cloud data is used as input to a trajectory prediction model to obtain trajectory parameters corresponding to several trajectories of the target vehicle within a set period; the target trajectory is obtained based on the trajectory parameters; wherein the trajectory prediction model is constructed based on an artificial intelligence model.
[0010] Preferably, converting the laser radar data into initial point cloud data includes:
[0011] Retrieve the LiDAR data collected from the target vehicle and convert it into Cartesian coordinates. Using the Open3D library, combine the Cartesian coordinates with the reflection intensity and timestamp to form the initial point cloud data. The LiDAR data includes the distance r, speed, angle information, timestamp, and reflection intensity of the laser emission point returned by the laser sensor. The angle information includes the horizontal angle φ and vertical scanning angle θ of the laser beam.
[0012] The calculation method for converting the laser radar data in the polar coordinate system into the Cartesian coordinate form is: through the equation system The X-axis distance, Y-axis distance, and Z-axis distance of the LiDAR data in Cartesian coordinates are calculated respectively.
[0013] The present invention reflects the surface material characteristics of an object through reflection intensity, which is beneficial to improving classification accuracy when combined with Cartesian coordinates.
[0014] Preferably, the determining whether the real-time operating scenario of the target vehicle is a preset dense scenario includes:
[0015] The first characteristic of setting dense scenes is: marking scenes with point cloud density per unit volume greater than a preset volume density threshold and point cloud density per unit area greater than a preset area density threshold as scenes with high point cloud density;
[0016] The second characteristic of the dense scene is set as follows: the ratio of outlier points in the point cloud data to the total number of points is greater than the preset outlier ratio threshold, the reflection intensity is lower than the preset intensity threshold, and the ratio of dynamic points in the point cloud data to the total number of points is greater than the dynamic ratio threshold;
[0017] The third characteristic of the dense scene is set as follows: the maximum delay time during point cloud data acquisition is lower than the preset delay threshold and the minimum frame rate is greater than 10Hz;
[0018] If the initial point cloud data meets the dense scene characteristics one, two, and three, the real-time operation scene of the target car is marked as meeting the preset dense scene.
[0019] Preferably, the downsampling operation steps include:
[0020] S110: Retrieve the three-dimensional space in the Cartesian coordinate system and preset the initial voxel size in the three-dimensional space according to the point cloud density:
[0021] S111: Mapping the point cloud coordinates to the voxel grid index to divide the voxel into several small voxels;
[0022] S112: Obtain the centroid coordinates of each voxel, and use the point corresponding to the centroid coordinates as the representative point of each voxel; generate initial downsampled point cloud data from the representative points of several voxels;
[0023] S113: Determine whether the real-time operating scenario of the target vehicle remains unchanged;
[0024] If yes, the preset initial voxel size remains unchanged;
[0025] If not, determine whether the real-time operating scene of the target vehicle has changed from a dense scene to a non-dense scene; if the dense scene is changed to a non-dense scene, replace the initial voxel size according to the preset scene change parameter table; if the non-dense scene is changed to a dense scene, replace the initial voxel size according to the preset scene change parameter table, and perform an interpolation operation during the replacement process;
[0026] S114: Repeat steps S112-S113 to obtain target downsampled point cloud data, and mark it as completed downsampling.
[0027] It should be noted that the centroid coordinates of each voxel are the average coordinates of each voxel.
[0028] The present invention reduces the amount of data while retaining more valid point cloud data by changing the voxel size in different scenarios. Specifically, large voxels are used to merge outliers in dense areas, which helps to reduce noise interference; voxels are reduced in sparse areas to avoid noise points from being mistakenly deleted, thereby retaining valid signals in the point cloud data.
[0029] Preferably, the filtering operation steps include:
[0030] Set the filtered point cloud data and preset the neighborhood radius; obtain the centroid coordinates of the filtered point cloud data, and mark the point corresponding to the centroid coordinates of the filtered point cloud data as the target point;
[0031] Calculate the average distance μ and standard deviation δ of each point in the filtered point cloud data to the target point within the neighborhood radius;
[0032] Eliminate points whose distance exceeds μ+kδ; where k represents the preset standard deviation multiple.
[0033] It should be noted that if the operation of downsampling first and then filtering is performed, the filtered point cloud data recorded in step S120 is the target downsampling point cloud data; if the operation of filtering first and then downsampling is performed, the filtered point cloud data recorded in step S120 is the initial point cloud data; the standard deviation multiple is set according to the scene change parameter table.
[0034] The present invention uses statistical filtering to analyze the neighborhood statistical characteristics of each point (such as average distance and standard deviation) to identify and eliminate outliers that do not conform to the distribution law.
[0035] Preferably, the step of using the pre-processed point cloud data as input to the trajectory prediction model to obtain trajectory parameters corresponding to a plurality of trajectories of the target vehicle within a set period includes:
[0036] Extract the target car's t-second historical trajectory and obtain the spatial features of the pre-processed point cloud data; the spatial features include: point cloud density ρ and the speed change rate a of each point t , angular change rate ω t and curvature K(t); the real-time environmental visibility is denoted as j;
[0037] Integrate spatial features to form a multimodal input vector X t =[a t ,ω t , K(t), K(t)', j]; where K(t)' is the first-order derivative of curvature;
[0038] Retrieve the trajectory prediction model and input the multimodal vector X t Input into the trajectory prediction model to obtain the trajectory parameters of H trajectories of the target vehicle in the future set period; wherein the trajectory parameters include: the expected value μ of the X axis x , the expected value μ of the Y axis y , the variance of the X axis σ x and the variance σ of the Y axis y .
[0039] The present invention enhances the multimodal input vector and retains more details such as curvature and real-time environmental visibility, so that the point cloud data input to the trajectory prediction model has less noise and more complete details, which is beneficial to improving the accuracy of the trajectory prediction model.
[0040] Preferably, obtaining the target trajectory based on the trajectory parameters includes:
[0041] Retrieve the trajectory parameters, and if the trajectory parameters conform to the Gaussian distribution, generate specific trajectory points through reparameterized sampling technology: x c ~N(μ x , σ x 2 ), y c~N(μ y ,σ y 2 ); where x c represents the x coordinate of the cth trajectory point; y c Represents the y coordinate of the cth trajectory point; the value range of H is [0, c], where c is a positive integer;
[0042] By formula Calculate the driving probability P of the cth trajectory c ;
[0043] P c The trajectory corresponding to the maximum value is taken as the target trajectory.
[0044] By calculating the travel probability of each trajectory, the present invention can quantitatively assess the feasibility of each trajectory. The trajectory with the highest travel probability is selected as the target trajectory, thereby selecting the trajectory that best meets expectations and is most likely to be actually executed, which helps to improve the accuracy and reliability of the predicted target trajectory.
[0045] Preferably, the trajectory prediction model is constructed according to an artificial intelligence model, including:
[0046] Obtaining standard training data; wherein the standard training data includes standard input data consistent with the content attributes of the time prediction sequence and standard output data consistent with the content attributes of the time when the filter bag needs to be replaced;
[0047] The artificial intelligence model is trained using standard training data, and the trained artificial intelligence model is marked as a trajectory prediction model; wherein the artificial intelligence model includes a convolutional neural network model or a long short-term memory neural network model.
[0048] To achieve the above-mentioned object, a second aspect of the present invention provides a target trajectory prediction system based on radar detection, comprising: a data acquisition module, a data preprocessing module and a trajectory prediction module;
[0049] Data acquisition module: collects the target vehicle's lidar data and environmental data based on the lidar sensor; converts the lidar data into initial point cloud data, where the environmental data includes real-time environmental visibility;
[0050] Data preprocessing module: determines whether the real-time operation scene of the target vehicle is a preset dense scene; if yes, the initial point cloud data is first downsampled and then filtered; if not, the initial point cloud data is first filtered and then downsampled to obtain the preprocessed point cloud data;
[0051] The trajectory prediction module: taking the preprocessed point cloud data as the input of a trajectory prediction model to obtain trajectory parameters corresponding to a plurality of trajectories of the target vehicle in a set period; and obtaining a target trajectory based on the trajectory parameters; wherein the trajectory prediction model is constructed according to an artificial intelligence model.
[0052] To achieve the above-mentioned purpose, the third aspect of the present application provides a target trajectory prediction calculation device based on radar detection, comprising a memory and a processor, the memory storing executable instructions of the processor; wherein the processor is configured to execute the executable instructions to implement the target trajectory prediction method based on radar detection provided in the first aspect.
[0053] Compared with the prior art, the present application has the following advantages:
[0054] 1. The present application can quickly remove redundant point clouds by adopting a dense scene preprocessing strategy (i.e. the operation of first downsampling and then filtering), reduce the subsequent filtering calculation amount, and is suitable for high complexity scenes such as urban intersections, to ensure the real-time requirement; by adopting a non-dense scene preprocessing strategy (first filtering and then downsampling), noise points are preferentially removed, and more effective details are retained, which is suitable for structured scenes such as highways, and improves the trajectory prediction accuracy; by dynamically adjusting the downsampling and filtering sequence, a balance between calculation efficiency and data fidelity is achieved, and resource waste of the traditional fixed process is avoided.
[0055] 2. The present application generates a probability distribution prediction result by a multi-trajectory parameter output and target trajectory screening mechanism, provides a risk assessment basis for an automatic driving decision system, and improves the system safety; after dynamically adjusting the processing sequence, and enhancing the multi-modal features, more details such as curvature, real-time environmental visibility, etc. are retained, so that the point cloud data noise input into the trajectory prediction model is less, the details are more complete, and the accuracy of the trajectory prediction model is improved. BRIEF DESCRIPTION OF DRAWINGS
[0056] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.
[0057] Figure 1 The flowchart of the target trajectory prediction of the present application;
[0058] Figure 2 The schematic diagram of the specific steps of data preprocessing of the present application;
[0059] Figure 3 The schematic diagram of the specific steps of trajectory prediction of the present application;
[0060] Figure 4 The schematic diagram of the module relationship included in the present application. DETAILED DESCRIPTION
[0061] The technical solutions of the present application will be described clearly and completely in combination with the embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.
[0062] Referring to Figure 1 , the first aspect embodiment of the present application provides a target trajectory prediction method based on radar detection, comprising:
[0063] collecting laser radar data and environment data of the target vehicle by the laser radar sensor; and converting the laser radar data into initial point cloud data, wherein the environment data comprises real-time environment visibility;
[0064] judging whether the real-time running scene of the target vehicle is a preset dense scene; if yes, performing the operation of first downsampling and then filtering on the initial point cloud data; if no, performing the operation of first filtering and then downsampling on the initial point cloud data; and obtaining preprocessed point cloud data;
[0065] taking the preprocessed point cloud data as the input of a trajectory prediction model to obtain trajectory parameters corresponding to a plurality of trajectories of the target vehicle within a set period; and obtaining a target trajectory based on the trajectory parameters, wherein the trajectory prediction model is constructed according to an artificial intelligence model.
[0066] Referring to Figure 2 , the specific steps of data preprocessing are as follows: judging whether the real-time running scene of the target vehicle is a preset dense scene; wherein, the first characteristic of the dense scene is that a scene in which the point cloud density in a unit volume is greater than a preset volume density threshold value and the point cloud density in a unit area is greater than a preset area density threshold value is marked as a scene with high point cloud density;
[0067] the second characteristic of the dense scene is that the proportion of outliers in the point cloud data to the total number of points is greater than a preset outlier ratio threshold value, the reflection intensity is lower than a preset intensity threshold value, and the proportion of dynamic points in the point cloud data to the total number of points is greater than a dynamic ratio threshold value;
[0068] the third characteristic of the dense scene is that the maximum delay time in the process of obtaining the point cloud data is lower than a preset delay threshold value and the minimum frame rate is greater than 10 Hz;
[0069] if the initial point cloud data all satisfy the first, second and third characteristics of the dense scene, the real-time running scene of the target vehicle is marked as a scene that meets the preset dense scene.
[0070] If the real-time operation scene of the target car is a preset dense scene, the initial point cloud data is first downsampled and then filtered; otherwise, the initial point cloud data is first filtered and then downsampled;
[0071] The downsampling operation steps are as follows: S110: Retrieve the three-dimensional space in the Cartesian coordinate system, and preset the initial voxel size in the three-dimensional space according to the point cloud density:
[0072] S111: Mapping the point cloud coordinates to the voxel grid index to divide the voxel into several small voxels;
[0073] S112: Obtain the centroid coordinates of each voxel, and use the point corresponding to the centroid coordinates as the representative point of each voxel; generate initial downsampled point cloud data from the representative points of several voxels;
[0074] S113: Determine whether the real-time operating scenario of the target vehicle remains unchanged;
[0075] If yes, the preset initial voxel size remains unchanged;
[0076] If not, determine whether the real-time operating scene of the target vehicle has changed from a dense scene to a non-dense scene; if the dense scene is changed to a non-dense scene, replace the initial voxel size according to the preset scene change parameter table; if the non-dense scene is changed to a dense scene, replace the initial voxel size according to the preset scene change parameter table, and perform an interpolation operation during the replacement process;
[0077] S114: Repeat steps S112-S113 to obtain the target downsampled point cloud data, and mark it as completed downsampling;
[0078] The filtering operation steps are as follows: setting the filtered point cloud data and presetting the neighborhood radius; obtaining the centroid coordinates of the filtered point cloud data, and marking the point corresponding to the centroid coordinates of the filtered point cloud data as the target point;
[0079] Calculate the average distance μ and standard deviation δ of each point in the filtered point cloud data to the target point within the neighborhood radius;
[0080] Eliminate points whose distance exceeds μ+kδ; where k represents the preset standard deviation multiple.
[0081] For example, setting the scene feature 1 as follows: marking scenes with point cloud density exceeding 50,000 points / cubic meter per unit volume and point cloud density exceeding 5,000 points / square meter per unit area as scenes with high point cloud density;
[0082] The second scenario feature is set as follows: the proportion of outliers in the point cloud data to the total number of points is greater than 10%, the reflection intensity is lower than the preset intensity threshold, and the proportion of dynamic points in the point cloud data to the total number of points is greater than 15%;
[0083] The third scenario feature is set as follows: the maximum delay time of the point cloud data acquisition process is less than 100ms and the minimum frame rate is greater than 10Hz;
[0084] A car Q is currently operating in a dense scene that meets the set scenario characteristics one, two, and three. The car Q's built-in lidar sensor collects lidar data and environmental data. The lidar data is converted into initial point cloud data, which is then downsampled and then filtered. The point cloud density of the initial point cloud data is 5720 points per square meter.
[0085] S110: Retrieve the three-dimensional space in the Cartesian coordinate system, and preset the initial voxel size in the three-dimensional space to 0.17m according to the point cloud density;
[0086] S111: Mapping the point cloud coordinates to the voxel grid index to divide the voxel into several small voxels;
[0087] S112: Obtain the centroid coordinates of each voxel, and use the point corresponding to the centroid coordinates as the representative point of each voxel; generate initial downsampled point cloud data from the representative points of several voxels;
[0088] S113: At this time, the real-time operation scene of the car is changed from a dense scene to a non-dense scene (the point cloud density range per unit area is 3680 points / square meter), and the initial voxel size is replaced with 0.28m according to the preset scene change parameter table (such as Table 1);
[0089] S114: Repeat steps S112-S113 to obtain the target downsampled point cloud data, and mark it as completed downsampling;
[0090] Set the filtered point cloud data, where the filtered point cloud data has a preset neighborhood radius for the target downsampled point cloud data; obtain the centroid coordinates of the filtered point cloud data, and mark the point corresponding to the centroid coordinates of the filtered point cloud data as the target point;
[0091] Calculate the average distance μ and standard deviation δ of each point in the filtered point cloud data to the target point within the neighborhood radius;
[0092] At this time, the point cloud density range within the unit area is 3680 points / square meter, and the preset standard deviation multiple k is 3;
[0093] Then, points with a distance greater than μ+3δ are removed from the filtered point cloud data.
[0094] Table 1 Scene change parameters
[0095]
[0096] See also Figure 3 The specific steps of trajectory prediction are: extract the target car's t-second historical running trajectory and obtain the spatial features of the pre-processed point cloud data; the spatial features include: point cloud density ρ and the speed change rate of each point a t , direction angle change rate ω t and curvature K(t); the real-time environmental visibility is denoted as j;
[0097] Integrate spatial features to form a multimodal input vector X t =[a t ,ω t , K(t), K(t)', j]; where K(t)' is the first-order derivative of curvature;
[0098] Retrieve the trajectory prediction model and input the multimodal vector X t Input into the trajectory prediction model to obtain the trajectory parameters of H trajectories of the target vehicle in the future set period; wherein the trajectory parameters include: the expected value μ of the X axis x , the expected value of the Y axis μ y , the variance of the X axis σ x and the variance σ of the Y axis y ;
[0099] Retrieve the trajectory parameters, and if the trajectory parameters conform to the Gaussian distribution, generate specific trajectory points through reparameterized sampling technology: x c ~N(μ x ,σ x 2 ), y c ~N(μ y ,σ y 2 ); where x c represents the x coordinate of the cth trajectory point; y c Represents the y coordinate of the cth trajectory point; the value range of H is [0, c], where c is a positive integer;
[0100] By formula Calculate the driving probability P of the cth trajectory c ;
[0101] P c The trajectory corresponding to the maximum value is taken as the target trajectory.
[0102] For example, in foggy weather, the target vehicle Q's trajectory in the next 3 seconds is predicted, with a point every 0.1 seconds, for a total of 30 points; the real-time environmental visibility is j = 50 meters;
[0103] Now the vehicle is moving along the curve, and the first three time stamps of the target trajectory to be predicted are retrieved, and some lidar data of the trajectory points within the first three time stamps are obtained (as shown in Table 2); among them, the direction angle
[0104] Table 2 Obtaining partial lidar data of trajectory points within the first three time stamps
[0105] Timestamp t x-coordinate y coordinate Speed (m / s) Direction angle γ 0 0 0 10 0° 0.1 1.0 0.5 10 30° 0.2 2.8 2.0 9 45°
[0106] Retrieve pre-processed point cloud data, point cloud density ρ = 50 points / m 2 ; Speed change rate Direction angle change rate Assume that at t = 0.2, the curvature K(t) = 0.05, K(t)′ = 0.2;
[0107] Integrate spatial features to form a multimodal input vector X t =[-1, 150, 0.05, 0.2, 50];
[0108] The multimodal input vector X t Input into the trajectory prediction model to obtain the trajectory parameters of the three trajectories in the future set period; among them, the trajectory parameter of the first trajectory (μ x , μ y , σ x , σ y ) is (10.2, 5.1, 0.3, 0.4); the trajectory parameters of the second trajectory are (20.5, 10.3, 0.5, 0.6); the trajectory parameters of the third trajectory are (30.8, 15.6, 0.7, 0.8);
[0109] Since the trajectory parameters conform to the Gaussian distribution, the first specific trajectory point generated by the reparameterized sampling technique is: x = 10.23, y = 5.08; the second specific trajectory point is: x = 6.19, y = 4.56; the third specific trajectory point is: x = 9.26, y = 12.55;
[0110] At this time, the probability of the first specific trajectory point p1 = 0.56; the probability of the second specific trajectory point p2 = 0.78; the probability of the third specific trajectory point p3 = 0.86; then the trajectory corresponding to the third specific trajectory point is selected as the target trajectory.
[0111] See also Figure 4 , a second aspect of the present invention provides a target trajectory prediction system based on radar detection, comprising: a data acquisition module, a data preprocessing module and a trajectory prediction module;
[0112] Data acquisition module: collects the target vehicle's lidar data and environmental data based on the lidar sensor; converts the lidar data into initial point cloud data, where the environmental data includes real-time environmental visibility;
[0113] Data preprocessing module: determines whether the real-time operation scene of the target vehicle is a preset dense scene; if yes, the initial point cloud data is first downsampled and then filtered; if not, the initial point cloud data is first filtered and then downsampled to obtain the preprocessed point cloud data;
[0114] Trajectory prediction module: The preprocessed point cloud data is used as the input of the trajectory prediction model to obtain the trajectory parameters corresponding to several trajectories of the target vehicle within a set period; the target trajectory is obtained based on the trajectory parameters; wherein, the trajectory prediction model is constructed based on the artificial intelligence model.
[0115] The third aspect of the present invention provides a target trajectory prediction and calculation device based on radar detection, comprising: a memory and a processor, wherein the memory stores executable instructions of the processor; wherein the processor is configured to execute a target trajectory prediction method based on radar detection provided by the first aspect by executing the executable instructions.
[0116] Some of the data in the above formula are calculated by removing the dimensions and taking their numerical values. The formula is a formula that is closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.
[0117] The working principle of the present invention is as follows: the present invention collects laser radar data and environmental data of a target vehicle through a laser radar sensor; converts the laser radar data into initial point cloud data; determines whether the real-time operation scene of the target vehicle is a preset dense scene; if so, downsamples and then filters the initial point cloud data; if not, filters and then downsamples the initial point cloud data; obtains preprocessed point cloud data; uses the preprocessed point cloud data as input to a trajectory prediction model to obtain trajectory parameters corresponding to several trajectories of the target vehicle within a set period; and obtains the target trajectory based on the trajectory parameters.
[0118] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.
Claims
1. A target trajectory prediction method based on radar detection, characterized in that: include: Collect lidar data of the target vehicle and environmental data using the lidar sensor; Convert the LiDAR data into initial point cloud data, where the environmental data includes: real-time environmental visibility; Determine whether the real-time operating scenario of the target vehicle is a preset dense scenario; if yes, downsample and then filter the initial point cloud data; if no, filter and then downsample the initial point cloud data to obtain preprocessed point cloud data; The preprocessed point cloud data is used as input to a trajectory prediction model to obtain trajectory parameters corresponding to several trajectories of the target vehicle within a set period; the target trajectory is obtained based on the trajectory parameters; wherein the trajectory prediction model is constructed based on an artificial intelligence model.
2. The target trajectory prediction method based on radar detection according to claim 1, characterized in that: The step of converting the laser radar data into initial point cloud data includes: Retrieve the LiDAR data collected from the target vehicle, convert it into Cartesian coordinate data, and use the 3D data processing library to combine the Cartesian coordinates with the reflection intensity and timestamp to form the initial point cloud data: the LiDAR data includes: the distance, speed, angle information, timestamp, and reflection intensity of the laser emission point returned by the laser radar to the surface of the LiDAR sensor; the angle information includes: the horizontal angle and vertical scanning angle of the laser beam.
3. The target trajectory prediction method based on radar detection according to claim 1, characterized in that: The determining whether the real-time operating scenario of the target vehicle is a preset dense scenario includes: The first characteristic of setting dense scenes is: marking scenes with point cloud density per unit volume greater than a preset volume density threshold and point cloud density per unit area greater than a preset area density threshold as scenes with high point cloud density; The second characteristic of the dense scene is set as follows: the ratio of outlier points in the point cloud data to the total number of points is greater than the preset outlier ratio threshold, the reflection intensity is lower than the preset intensity threshold, and the ratio of dynamic points in the point cloud data to the total number of points is greater than the dynamic ratio threshold; The third characteristic of the dense scene is set as follows: the maximum delay time during point cloud data acquisition is lower than the preset delay threshold and the minimum frame rate is greater than 10Hz; If the initial point cloud data meets the dense scene characteristics one, two, and three, the real-time operation scene of the target car is marked as meeting the preset dense scene.
4. The target trajectory prediction method based on radar detection according to claim 1, characterized in that: The downsampling operation steps include: S110: Retrieve the three-dimensional space in the Cartesian coordinate system and preset the initial voxel size in the three-dimensional space according to the point cloud density: S111: Mapping the point cloud coordinates to the voxel grid index to divide the voxel into several small voxels; S112: Obtain the centroid coordinates of each voxel, and use the point corresponding to the centroid coordinates as the representative point of each voxel; generate initial downsampled point cloud data from the representative points of several voxels; S113: Determine whether the real-time operating scenario of the target vehicle remains unchanged; If yes, the preset initial voxel size remains unchanged; If not, determine whether the real-time operating scene of the target vehicle has changed from a dense scene to a non-dense scene; if the dense scene is changed to a non-dense scene, replace the initial voxel size according to the preset scene change parameter table; if the non-dense scene is changed to a dense scene, replace the initial voxel size according to the preset scene change parameter table, and perform an interpolation operation during the replacement process; S114: Repeat steps S112-S113 to obtain target downsampled point cloud data, and mark it as completed downsampling.
5. The target trajectory prediction method based on radar detection according to claim 1, characterized in that: The filtering operation steps include: Set the filtered point cloud data and preset the neighborhood radius; obtain the centroid coordinates of the filtered point cloud data, and mark the point corresponding to the centroid coordinates of the filtered point cloud data as the target point; Calculate the average distance μ and standard deviation δ of each point in the filtered point cloud data to the target point within the neighborhood radius; Eliminate points whose distance exceeds μ+kδ; where k represents the preset standard deviation multiple.
6. The target trajectory prediction method based on radar detection according to claim 1, characterized in that: The pre-processed point cloud data is used as the input of the trajectory prediction model to obtain trajectory parameters corresponding to several trajectories of the target vehicle within a set period, including: Extract the target car's t-second historical trajectory and obtain the spatial features of the pre-processed point cloud data; the spatial features include: point cloud density ρ and the speed change rate a of each point t , direction angle change rate ω t and curvature K(t); the real-time environmental visibility is denoted as j; Integrate spatial features to form a multimodal input vector X t =[a t ,ω t , K(t), K(t), j]; where K(t) is the first-order derivative of curvature; Retrieve the trajectory prediction model and input the multimodal vector X t Input into the trajectory prediction model to obtain the trajectory parameters of H trajectories of the target vehicle in the future set period; wherein the trajectory parameters include: the expected value μ of the X axis x , the expected value μ of the Y axis y , the variance of the X axis σ x and the variance σ of the Y axis y .
7. The target trajectory prediction method based on radar detection according to claim 1, characterized in that: The obtaining of the target trajectory based on the trajectory parameters includes: Retrieve the trajectory parameters, and if the trajectory parameters conform to the Gaussian distribution, generate specific trajectory points through the reparameterized sampling technology: Among them, x c represents the x coordinate of the cth trajectory point; y c Represents the y coordinate of the cth trajectory point; the value range of H is [0, c], where c is a positive integer; By formula Calculate the driving probability P of the cth trajectory c ; P c The trajectory corresponding to the maximum value is taken as the target trajectory.
8. The target trajectory prediction method based on radar detection according to claim 1, characterized in that: The trajectory prediction model is constructed based on an artificial intelligence model and includes: Obtaining standard training data; wherein the standard training data includes standard input data consistent with the content attributes of the time prediction sequence and standard output data consistent with the content attributes of the time when the filter bag needs to be replaced; The artificial intelligence model is trained using standard training data, and the trained artificial intelligence model is marked as a trajectory prediction model; wherein the artificial intelligence model includes a convolutional neural network model or a long short-term memory neural network model.
9. A target trajectory prediction system based on radar detection, adapted to the target trajectory prediction method based on radar detection according to any one of claims 1 to 8, characterized in that: include: Data acquisition module, data preprocessing module and trajectory prediction module; Data acquisition module: collects the target vehicle's lidar data and environmental data based on the lidar sensor; Convert the LiDAR data into initial point cloud data, where the environmental data includes: real-time environmental visibility; Data preprocessing module: determines whether the real-time operation scene of the target vehicle is a preset dense scene; if yes, the initial point cloud data is first downsampled and then filtered; if not, the initial point cloud data is first filtered and then downsampled to obtain the preprocessed point cloud data; Trajectory prediction module: The preprocessed point cloud data is used as the input of the trajectory prediction model to obtain the trajectory parameters corresponding to several trajectories of the target vehicle within a set period; the target trajectory is obtained based on the trajectory parameters; wherein, the trajectory prediction model is constructed based on the artificial intelligence model.
10. A target trajectory prediction and calculation device based on radar detection, characterized in that: include: A memory and a processor, wherein the memory stores executable instructions of the processor; wherein the processor is configured to execute a target trajectory prediction method based on radar detection according to any one of claims 1 to 8 by executing the executable instructions.
Citation Information
Cited By
Multi-source information high-precision positioning method and system for unmanned forklift
CN121932995A
A high-precision positioning method and system for unmanned forklifts using multi-source information
CN121932995B
