Dynamic target detection method and device, storage medium and terminal

Abstract

A dynamic target detection method and device, a storage medium and a terminal, the dynamic target detection method comprising: projecting three-dimensional point cloud data in an environment around a vehicle obtained to a polar coordinate grid to obtain two-dimensional projected point cloud, the three-dimensional point cloud data being collected by a laser radar, and the polar coordinate grid being constructed on a horizontal plane with a center of the laser radar as a coordinate origin; performing point cloud clustering on the two-dimensional projected point cloud according to positions of the two-dimensional projected point cloud on the polar coordinate grid to obtain a point cloud clustering result; and performing dynamic target detection according to the point cloud clustering result. The above scheme can improve the accuracy of a dynamic target detection result, so when the dynamic target is a dynamic vehicle, it is expected to improve the accuracy of dynamic vehicle detection.

Description

Technical Field

[0001] This invention relates to the field of target detection, and more particularly to a dynamic target detection method and apparatus, storage medium, and terminal. Background Technology

[0002] Environmental perception is one of the essential technologies for achieving autonomous driving. Within environmental perception, dynamic vehicle detection is crucial for autonomous vehicles to perceive the dynamic environment around them and make corresponding decisions and control strategies based on these dynamic changes.

[0003] Dynamic vehicle detection methods can generally be divided into two categories: learning-based dynamic vehicle detection methods and model-based dynamic vehicle detection methods. Learning-based dynamic vehicle detection methods often require a large amount of training data. Although detection accuracy generally increases with the amount of training data, these methods struggle to correctly detect targets not present in the training database. Furthermore, learning-based dynamic vehicle detection methods consume significant computational resources, especially given the recent surge in research interest in deep learning, which places high demands on hardware resources. Model-based dynamic vehicle detection methods, on the other hand, typically utilize model fitting. They do not rely on large amounts of training data and have lower computational resource requirements compared to learning-based methods. Since real-world driving environments are complex and variable, it is often difficult to collect complete data for training learning-based dynamic vehicle detection methods. Therefore, model-based dynamic vehicle detection methods play a crucial supporting role in enabling autonomous vehicles to make autonomous judgments in unknown environments.

[0004] Model-based dynamic vehicle detection methods typically involve LiDAR point cloud clustering, such as using Radially Bounded Nearest Neighbor (RBNN) clustering. However, this clustering method is prone to under-segmentation or over-segmentation of the point cloud, affecting the accuracy of the clustering results and consequently the accuracy of subsequent target detection. Summary of the Invention

[0005] The technical problem solved by the embodiments of the present invention is the low accuracy of dynamic target detection results.

[0006] To address the aforementioned technical problems, this invention provides a dynamic target detection method, comprising: projecting acquired three-dimensional point cloud data of the vehicle's surrounding environment onto a polar coordinate grid to obtain a two-dimensional projected point cloud, wherein the three-dimensional point cloud data is acquired by a lidar, and the polar coordinate grid is constructed on a horizontal plane with the center of the lidar as the origin; performing point cloud clustering on the two-dimensional projected point cloud based on its position on the polar coordinate grid to obtain a point cloud clustering result; and performing dynamic target detection based on the point cloud clustering result.

[0007] Optionally, the method for constructing the polar coordinate grid includes: taking the center of the lidar as the origin of the coordinate system, dividing the horizontal plane along the circumferential and radial directions respectively to obtain the polar coordinate grid.

[0008] Optionally, dividing the horizontal plane along the circumferential and radial directions respectively includes: dividing the horizontal plane into multiple sector segments along the circumferential direction with equal central angles; and dividing the horizontal plane into multiple annular segments along the radial direction with equal ring widths.

[0009] Optionally, the step of clustering the two-dimensional projected point cloud based on its position on the polar coordinate grid includes: marking the occupancy status of each grid in the polar coordinate grid to obtain a polar coordinate grid marked with grid status, wherein the grid status includes an occupied state and a free state, the occupied state indicating that the grid contains the two-dimensional projected point cloud, and the free state indicating that the grid does not contain the two-dimensional projected point cloud; unfolding the polar coordinate grid marked with grid status in the radial and circumferential directions to obtain an unfolded grid diagram; and clustering the two-dimensional projected point cloud based on the grid status of each grid in the unfolded grid diagram.

[0010] Optionally, the step of clustering the two-dimensional projection points based on the grid state of each grid in the unfolded grid diagram includes: binarizing the unfolded grid diagram to obtain a binary image; performing morphological dilation on the binary image; extracting connected regions from the morphologically dilated binary image to obtain multiple connected regions; and marking the two-dimensional projection point clouds corresponding to the grids in each connected region as the same cluster.

[0011] Optionally, the step of extracting connected regions from the binary image after morphological dilation includes: searching for connected regions within the binary image after morphological dilation using a neighborhood connectivity approach.

[0012] Optionally, the step of clustering the two-dimensional projected point cloud according to the grid state of each grid in the unfolded grid diagram includes: taking each grid in the occupied state as the center in turn, drawing a circle with a set threshold as the radius, marking all grids in the occupied state inside the circle as the same cluster, until all grids in the occupied state are traversed in a set order.

[0013] Optionally, the dynamic target detection method further includes: before projecting the acquired three-dimensional point cloud data of the surrounding environment of the vehicle onto the polar coordinate grid, filtering out ground point clouds from the three-dimensional point cloud data to obtain ground point clouds.

[0014] Optionally, the step of performing dynamic target detection based on the point cloud clustering results includes: performing candidate clustering based on the relationship between the point cloud clustering results and preset target attribute filtering conditions to obtain candidate clusters; performing dynamic target detection based on the positions of the two-dimensional projected point clouds corresponding to the candidate clusters in adjacent frame point clouds in the polar coordinate grid, and determining candidate dynamic targets based on the dynamic target detection results; and confirming the candidate dynamic targets to obtain the type of the dynamic targets.

[0015] Optionally, the step of filtering candidate clusters based on the relationship between the point cloud clustering results and preset target attribute filtering conditions to obtain candidate clusters includes: determining whether the point cloud clustering results meet the target attribute filtering conditions; and selecting point cloud clusters whose point cloud clustering results meet the target attribute filtering conditions as candidate clusters; wherein the target attribute filtering conditions include: the number of points in the cluster meets a set number and the size of the cluster meets a set size.

[0016] Optionally, the step of performing dynamic target detection based on the positions of the two-dimensional projected point clouds corresponding to the candidate clusters in the polar coordinate grid in adjacent frame point clouds, and determining candidate dynamic targets based on the dynamic target detection results, includes: unifying the coordinates of the point clouds in adjacent frames to the same coordinate system; determining whether the object corresponding to the candidate cluster is a dynamic target based on the position changes of the point clouds corresponding to the candidate clusters in the polar coordinate grid in adjacent frames; and determining the candidate cluster as a candidate dynamic target if it is a dynamic target.

[0017] Optionally, determining whether the object corresponding to the candidate cluster is a dynamic target based on the position changes of the point clouds corresponding to the candidate clusters in the polar coordinate grid in adjacent frames includes: performing a difference between the grid occupancy of the point clouds corresponding to the candidate clusters in one frame of adjacent frames and the grid occupancy of the point clouds corresponding to the candidate clusters in another frame of adjacent frames in the polar coordinate grid to obtain the number of grids whose grid occupancy has changed; when the number of grids whose grid occupancy has changed is greater than a set dynamic object detection threshold, the candidate cluster is determined to be the candidate dynamic target.

[0018] Optionally, the dynamic object detection threshold is related to the shortest side length of the dynamic target projection, the relative distance between the center of the dynamic target and the lidar, and the resolution of the polar coordinate grid.

[0019] Optionally, the step of confirming the candidate dynamic target to obtain the type of the dynamic target includes: making a consistency judgment on the motion state changes of the candidate dynamic target based on N consecutive frames of point cloud, where N is a positive integer greater than 1; and determining the type of the dynamic target based on the consistency judgment result.

[0020] Optionally, the step of performing a consistency judgment on the motion state changes of the candidate dynamic target based on N consecutive frames of point cloud includes: performing pose estimation on the candidate dynamic target to obtain the center position and orientation angle of the candidate dynamic target at time t, where t is a positive integer; performing data association on the frame point cloud acquired at time t+1 based on the orientation angle and motion speed of the candidate dynamic target at time t to obtain a first association cluster associated with the candidate dynamic target at time t+1; performing pose estimation on the first association cluster to obtain the center position and orientation angle of the first association cluster at time t+1; performing data association on the frame point cloud acquired at time t+2 based on the motion direction corresponding to the orientation angle of the first association cluster at time t+1 to obtain a second association cluster; and performing a consistency judgment on the motion state changes based on the motion speed and orientation angle of the candidate cluster, the first association cluster, and the second association cluster.

[0021] Optionally, the step of determining the consistency of motion state changes based on the motion speed and direction angle of the candidate cluster, the first associated cluster, and the second associated cluster includes: calculating a first speed difference between the candidate cluster and the first associated cluster, and a second speed difference between the first associated cluster and the second associated cluster; calculating a first direction angle difference between the candidate cluster and the first associated cluster, and a second direction angle difference between the first associated cluster and the second associated cluster; determining whether the first speed difference and the second speed difference are both less than a preset speed difference, and whether the first direction angle difference and the second direction angle difference are both less than a preset direction angle difference; if the first speed difference and the second speed difference are both less than the preset speed difference, and the first direction angle difference and the second direction angle difference are both less than the preset direction angle difference, then the motion state changes are determined to be consistent.

[0022] This invention also provides a dynamic target detection device, comprising: a projection unit for projecting acquired three-dimensional point cloud data of the vehicle's surrounding environment onto a polar coordinate grid to obtain a two-dimensional projected point cloud, wherein the three-dimensional point cloud data is acquired by a lidar, and the polar coordinate grid is constructed on a horizontal plane with the center of the lidar as the origin; a clustering unit for performing point cloud clustering on the two-dimensional projected point cloud according to its position on the polar coordinate grid to obtain a point cloud clustering result; and a dynamic target detection unit for performing dynamic target detection based on the point cloud clustering result.

[0023] This invention also provides a storage medium, which is a non-volatile or non-transient storage medium, on which a computer program is stored. When the computer program is run by a processor, it executes the steps of any of the above-described dynamic target detection methods.

[0024] This invention also provides a terminal, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the steps of any of the above-described dynamic target detection methods when running the computer program.

[0025] Compared with the prior art, the technical solution of the embodiments of the present invention has the following beneficial effects:

[0026] The acquired 3D point cloud data of the vehicle's surrounding environment is projected onto a polar coordinate grid to obtain a 2D projected point cloud. Based on the position of the 2D projected point cloud on the polar coordinate grid, point cloud clustering is performed to obtain the clustering results. Since the polar coordinate grid is constructed on a horizontal plane with the center of the LiDAR as the origin, it can better characterize the distribution characteristics of the LiDAR point cloud. This results in a high consistency between the distribution characteristics of the 2D projected point cloud on the polar coordinate grid and the distribution characteristics of the LiDAR point cloud, thus solving the undersegmentation and oversegmentation problems caused by the fixed distance threshold used in the RBNN algorithm. Therefore, clustering the 2D projected point cloud based on its position on the polar coordinate grid yields highly accurate clustering results. Consequently, when performing dynamic target detection based on the clustering results, the accuracy of dynamic target detection can be improved. Attached Figure Description

[0027] Figure 1 This is a flowchart of a dynamic target detection method according to an embodiment of the present invention;

[0028] Figure 2 This is a schematic diagram of a polar coordinate grid in an embodiment of the present invention;

[0029] Figure 3 yes Figure 1 One specific implementation of step S13;

[0030] Figure 4 This is a schematic diagram of a rule-based point cloud preprocessing method in an embodiment of the present invention;

[0031] Figure 5 This is a flowchart of a dynamic object detection process in an embodiment of the present invention;

[0032] Figure 6 This is a rule-based dynamic vehicle confirmation flowchart in an embodiment of the present invention;

[0033] Figure 7 This is a schematic diagram of the structure of a dynamic target detection device according to an embodiment of the present invention. Detailed Implementation

[0034] As mentioned above, model-based dynamic vehicle detection methods typically involve LiDAR point cloud clustering, such as using Radially Bounded Nearest Neighbor (RBNN) clustering. Research has shown that while RBNN uses a fixed radius for bounding sphere calculations, the point cloud density distribution acquired by the vehicle-mounted LiDAR is uneven. Using the same distance threshold for both near and far locations contradicts the characteristics of the LiDAR point cloud distribution. Furthermore, using a fixed radius for bounding sphere calculations can easily lead to under-segmentation or over-segmentation of the point cloud, resulting in low clustering accuracy and consequently affecting the accuracy of subsequent target detection.

[0035] To address the aforementioned issues, in this embodiment of the invention, the acquired 3D point cloud data of the vehicle's surrounding environment is projected onto a polar coordinate grid to obtain a 2D projected point cloud. Based on the position of the 2D projected point cloud on the polar coordinate grid, point cloud clustering is performed to obtain clustering results. Since the polar coordinate grid is constructed on a horizontal plane with the center of the LiDAR as the origin, it can better characterize the distribution characteristics of the LiDAR point cloud. This results in a high consistency between the distribution characteristics of the 2D projected point cloud on the polar coordinate grid and the distribution characteristics of the LiDAR point cloud, thus solving the under-segmentation and over-segmentation problems caused by the fixed distance threshold used in the RBNN algorithm. Therefore, clustering the 2D projected point cloud based on its position on the polar coordinate grid yields highly accurate clustering results. Consequently, when performing dynamic target detection based on the clustering results, the accuracy of dynamic target detection can be improved.

[0036] To make the above-mentioned objectives, features and beneficial effects of the embodiments of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0037] This invention provides a dynamic target detection method, referring to... Figure 1 The flowchart of a dynamic target detection method provided in this embodiment of the invention may specifically include the following steps:

[0038] Step S11: Project the acquired 3D point cloud data of the environment surrounding the vehicle onto a polar coordinate grid to obtain a 2D projected point cloud.

[0039] In practice, the 3D point cloud data is collected by lidar.

[0040] The polar coordinate grid is constructed on a horizontal plane with the center of the lidar as the origin.

[0041] In some embodiments, a polar coordinate grid can be constructed as follows: with the center of the lidar as the origin of the coordinate system, the horizontal plane is divided along the circumferential and radial directions respectively to obtain a polar coordinate grid.

[0042] Specifically, the horizontal plane is divided along the circumference with equal central angles to obtain multiple sector-shaped segments. The horizontal plane is also divided radially with equal ring widths to obtain multiple annular segments. After circumferential and radial division, the horizontal plane can be divided into multiple grids, resulting in a polar coordinate grid. The farther the grid is from the origin, the larger its size, which is consistent with the radial divergence characteristic of lidar point clouds.

[0043] Reference Figure 2 A schematic diagram of a polar coordinate grid in an embodiment of the present invention is provided.

[0044] To distinguish different grids, each grid can be identified using an index. Specifically, the index of a vertex of a grid can be used as the index of that grid.

[0045] In some embodiments, the index of the polar coordinate grid can be represented by the following formula:

[0046] n g = < ρ i / γ g >; (1)

[0047] n θ = <θ i / γ θ >; (2)

[0048] Among them, (n g ,n θ ) is the index of one of the grids, ρ i Let θ be the polar radius of a vertex of this grid. i γ is the polar angle of a vertex of this grid. g γ is the radial resolution of the polar coordinate grid. θ This represents the circumferential resolution (also known as angular resolution) of the polar coordinate grid.

[0049] Projecting 3D point cloud data onto a polar coordinate grid yields a 2D projected point cloud. The coordinates of the 2D projected point cloud are then converted to polar coordinates, which can be expressed using the polar radius ρ. i and polar angle θ i It is represented as follows. Where, the polar radius ρ i The polar angle θ can be calculated using formula (3). i It can be calculated using the following formula (4):

[0050]

[0051]

[0052] Where, (ρ i ,θ i p is a two-dimensional projected point cloud i polar coordinates, ρ i For point cloud p i The polar radius, θ i For point cloud p i The polar angle.

[0053] The set of 3D point clouds is P = {p i} i＝1,...,n ,p i =(x i y i , z i ), where p i =(x i y i , z i Let be the three-dimensional coordinates of the i-th point cloud. The polar coordinate set of the two-dimensional projected point cloud is P = {p i} i＝1,...,n ,p i =(ρ i θ i Where n is the total number of 3D point clouds, n is a positive integer, 1≤i≤n.

[0054] In practice, the size of the equal central angle and the ring width affect the resolution of the polar coordinate grid, that is, the density of the polar coordinate grid. The size of the equal central angle and the ring width can be set according to the application scenario. For example, the equal central angle in urban traffic scenarios is smaller than that in highway traffic scenarios, and the ring width in urban traffic scenarios is smaller than that in highway traffic scenarios.

[0055] In practice, the 3D point cloud data collected by lidar may include ground point clouds and surface point clouds. Ground point clouds refer to the point clouds corresponding to objects above ground. Surface point clouds refer to the point clouds corresponding to road surfaces, etc. Dynamic target detection typically focuses primarily on objects on the ground.

[0056] To improve point cloud processing speed, in this embodiment of the invention, the acquired 3D point cloud data of the vehicle's surrounding environment is preprocessed before being projected onto the polar coordinate grid. Preprocessing mainly includes filtering out ground point clouds to obtain ground point clouds. These ground point clouds are then used as 3D point cloud data for subsequent dynamic target detection. By filtering out ground point clouds, point cloud data that is not beneficial or has minimal benefit for dynamic target detection can be removed, reducing the number of point clouds that need to be processed in the subsequent dynamic target detection process, thereby improving point cloud processing speed and reducing computational resource consumption.

[0057] In practical implementation, various methods can be used to filter out ground point clouds. In some embodiments, the vehicle's position is obtained using a Global Positioning System (GPS). The vehicle's attitude is obtained using an Inertial Measurement Unit (IMU), and ground state tracking is performed using an extended Kalman filter, or a ground elevation model is established using Gaussian process regression. Ground point cloud segmentation is then performed on the 3D point cloud data acquired by the lidar.

[0058] Step S12: Based on the position of the two-dimensional projected point cloud on the polar coordinate grid, perform point cloud clustering on the two-dimensional projected point cloud to obtain the point cloud clustering result.

[0059] Specifically, based on the position of the two-dimensional projected point cloud on the polar coordinate grid, the occupancy status of each grid in the polar coordinate grid is marked, resulting in a polar coordinate grid marked with grid states. The grid states include occupied states and free states. An occupied state indicates that the grid contains two-dimensional projected point clouds, meaning that two-dimensional projected point clouds fall within the grid. A free state indicates that the grid does not contain the two-dimensional projected point cloud, meaning that no two-dimensional projected point cloud falls within the grid.

[0060] The polar coordinate grid marked with grid states is expanded radially and circumferentially to obtain an expanded grid map. Based on the grid states of each grid in the expanded grid map, the two-dimensional projected point cloud is clustered.

[0061] In some embodiments, occupied and free states can be distinguished by different identifiers, such as 1 representing occupied state and 0 representing free state. It is understood that other identifiers can also be used to identify occupied or free states, as long as they can distinguish between them; examples will not be provided here.

[0062] In practice, point cloud clustering can be performed in various ways.

[0063] In one embodiment, the unfolded mesh image is binarized to obtain a binary image; morphological dilation is performed on the binary image; connected component extraction is performed on the morphologically dilated binary image to obtain multiple connected regions; and the two-dimensional projected point clouds corresponding to the meshes within each connected region are labeled as belonging to the same cluster. That is, the three-dimensional point clouds corresponding to the two-dimensional projected point clouds of the meshes within each connected region belong to the same cluster.

[0064] For example, if the resulting connected regions are A, B, and C, then the two-dimensional projected point clouds within connected region A belong to the same cluster. The two-dimensional projected point clouds within connected region B belong to the same cluster. The two-dimensional projected point clouds within connected region C belong to the same cluster.

[0065] In some non-limiting embodiments, a neighborhood connectivity approach is used to search for connected regions within the morphologically dilated binary graph. For example, an 8-neighborhood connectivity approach is used to search for connected regions within the morphologically dilated binary graph.

[0066] In another embodiment, each grid in the occupied state is used as the center in turn, and a circle is drawn with a set threshold as the radius. All grids in the occupied state inside the circle are marked as the same cluster, until all grids in the occupied state are traversed in a set order to complete the point cloud clustering.

[0067] Furthermore, to improve point cloud clustering efficiency, a grid g with a state of 1 can be selected according to a set rule. i , with the selected grid g i Set a threshold t centered on the target. g Draw a circle with radius g, and enclose a grid whose state is occupied by the grid in state 1 and g. i They belong to the same cluster. Skip the ones starting with g. i Centered on, distance threshold t g Within a circle with a radius of 1, the grid is covered. Then, circles are drawn with the grid outside the circle that is in occupied state 1, using a set threshold as the center, until all grids in occupied state 1 have been traversed. During point cloud clustering, if a grid within the circle already has a cluster category, all grids enclosed by the circle that are in occupied state belong to that same cluster category. If none of the grids within the circle have a cluster category, then all grids in occupied state within the circle are marked as a new cluster category.

[0068] Step S13: Perform dynamic target detection based on the point cloud clustering results.

[0069] As shown above, the acquired 3D point cloud data of the vehicle's surrounding environment is projected onto a polar coordinate grid to obtain a 2D projected point cloud. Based on the position of the 2D projected point cloud on the polar coordinate grid, point cloud clustering is performed to obtain the clustering results. Since the polar coordinate grid is constructed on a horizontal plane with the center of the LiDAR as the origin, it can better characterize the distribution characteristics of the LiDAR point cloud. This results in a high consistency between the distribution characteristics of the 2D projected point cloud on the polar coordinate grid and the distribution characteristics of the LiDAR point cloud, thus solving the undersegmentation and oversegmentation problems caused by the fixed distance threshold used in the RBNN algorithm. Therefore, clustering the 2D projected point cloud based on its position on the polar coordinate grid yields highly accurate clustering results. Consequently, when performing dynamic target detection based on the clustering results, the accuracy of dynamic target detection can be improved.

[0070] Furthermore, reducing 3D point cloud data to 2D projected point cloud and processing it based on polar coordinate grids can reduce the computational cost of point cloud clustering.

[0071] In some embodiments, refer to Figure 3 , give Figure 1 One specific implementation of step S13 may include the following steps:

[0072] Step S131: Based on the relationship between the point cloud clustering results and the preset target attribute filtering conditions, candidate clusters are filtered to obtain candidate clusters.

[0073] In this embodiment of the invention, candidate clustering can be screened according to set rules, such as configuring target attribute screening conditions by setting corresponding rules based on the attributes of the target to be detected.

[0074] Specifically, it is determined whether the point cloud clustering results meet the target attribute filtering conditions; point cloud clusters whose point cloud clustering results meet the target attribute filtering conditions are selected as candidate clusters. The target attribute filtering conditions include: the number of points in the cluster meets a set number, and the size of the cluster meets a set size.

[0075] The requirement that the number of point clouds in a cluster meets the set number can include the number of point clouds in a cluster being less than the set number, or the number of point clouds in a cluster being more than the preset number. The relationship between the number of point clouds in a cluster and the set number can be configured according to the actual application scenario requirements.

[0076] The size of a cluster can include its length and width. Meeting the set size criteria can include: the cluster length being less than the set length, the cluster width being less than the set width, and the cluster height being less than the set height. Meeting the set size criteria can also include: the cluster length being greater than the set length, the cluster width being greater than the set width, and the cluster height being greater than the set height.

[0077] Understandably, to further improve the accuracy of candidate cluster selection, the relationship between the cluster size and the set size can have other variations. For example, if the set size is a size range, the cluster size must meet the set size range. The relationship between the cluster size and the set size can be configured according to the specific application scenario requirements.

[0078] Since different types of targets have different shapes and sizes, the target attribute filtering conditions can be set according to the type of target to be detected. For example, the target attribute filtering conditions for a vehicle are different from those for a cyclist.

[0079] LiDAR possesses precise ranging capabilities, and despite uneven point cloud density distribution, it can still accurately reflect the dimensions of the target's outer surface. By setting cluster size and the number of clustered point clouds in the target attribute filtering conditions for initial screening, large obstacles such as walls and trees can be filtered out, improving the efficiency of subsequent dynamic target detection and avoiding unnecessary computation.

[0080] Step S132: Based on the positions of the two-dimensional projected point clouds corresponding to the candidate clusters in the adjacent frame point clouds in the polar coordinate grid, perform dynamic target detection, and determine candidate dynamic targets based on the dynamic target detection results.

[0081] Vehicle-mounted sensors can acquire continuous sensing data. Therefore, research has found that the motion properties of an object in the environment can be determined by detecting whether the reflection point cloud of an object changes in the received three-dimensional point cloud data at adjacent time points.

[0082] For targets requiring dynamic detection, corresponding grid state change rules can be set to filter candidate dynamic targets. These candidate dynamic targets can include dynamic vehicles, dynamic pedestrians, or dynamic cyclists, etc. The configured grid state change rules differ depending on the dynamic target being detected.

[0083] In specific implementation, the coordinates of point clouds in adjacent frames are unified into the same coordinate system; based on the position changes of the point clouds corresponding to the candidate clusters in the polar coordinate grid in the adjacent frames, it is determined whether the object corresponding to the candidate cluster is a dynamic target; if it is a dynamic target, the candidate cluster is determined to be a candidate dynamic target.

[0084] The grid occupancy of the point cloud corresponding to the candidate cluster in one frame of the adjacent frame is compared with the grid occupancy of the point cloud corresponding to the candidate cluster in another frame of the adjacent frame in the polar coordinate grid to obtain the number of grids whose grid occupancy has changed. When the number of grids whose grid occupancy has changed is greater than the set dynamic object detection threshold, the candidate cluster is determined as the candidate dynamic target.

[0085] The dynamic object detection threshold is related to factors such as the shortest side length of the dynamic target projection, the relative distance between the center of the dynamic target and the lidar, and the resolution of the polar coordinate grid.

[0086] In this embodiment of the invention, the dynamic object detection threshold can be expressed by the following formula (5):

[0087] T D = <L O / (||O||·γ θ (5)

[0088] Among them, L O γ represents the shortest side length of the dynamic target projection; ||O|| represents the distance between the center of the dynamic target and the lidar, and γ θ This represents the angular resolution of the polar coordinate grid.

[0089] The shortest side length of the dynamic target projection is related to the category of the target to be detected. That is, the shortest side length of the dynamic target projection can be determined based on the category of the target to be detected.

[0090] To help those skilled in the art better understand the determination of candidate dynamic targets, a specific embodiment is described below.

[0091] Taking dynamic vehicle detection as an example, after obtaining candidate clusters, the vehicle's position information collected by GPS and its pose information acquired by IMU are first used. The pose information can include position information and orientation angle. The coordinate system of two consecutive frames of point cloud at time t and t-1 is placed in the coordinate system corresponding to time t. The coordinates of the two frames of point cloud are projected onto the horizontal plane and polar coordinate grid mapping is performed. The polar coordinate grid is centered on the lidar and the horizontal plane is divided into sectors with equal central angles φ. At the same time, the horizontal plane is divided into rings with equal ring widths along the radial direction, dividing the horizontal plane into multiple grids. The farther the grid is from the lidar, the larger its size, which is consistent with the radial divergence characteristic of the lidar point cloud.

[0092] At time t, search for candidate clusters C. i The occupied grid positions and number correspond to the grid occupancy at the same position at time t-1. The candidate clusters C at time t are then... iThe number of grid cells with different occupancy statuses (i.e., those whose grid states have changed) is identified by subtracting the occupant's grid position from the grid cell at the same position at time t-1. This is represented by the function DIFF(). i This is the candidate cluster C. i Mesh states that change at times t and t-1. (DIFF(C)) i ) and dynamic object detection threshold T D Comparison, if DIFF(C) i ) greater than T D Then the candidate cluster C i It was determined to be a candidate dynamic vehicle.

[0093] Step S133: Perform dynamic target confirmation on the candidate dynamic target to obtain the type of the dynamic target.

[0094] In practical implementation, the consistency of motion state changes of the candidate dynamic target can be judged based on N consecutive frames of point cloud data, where N is a positive integer; the type of the dynamic target is determined based on the consistency judgment result, where N is a positive integer greater than 1. The consistency judgment result indicates whether the motion state changes of the candidate dynamic target are consistent; if consistent, the candidate dynamic target is confirmed as a dynamic target, and its type is determined.

[0095] In some embodiments, to improve the accuracy of determining the type of dynamic target, N can be 3, or other values ​​such as 4 or 5.

[0096] In specific implementation, pose estimation is performed on the candidate dynamic target to obtain the center position and orientation angle of the candidate dynamic target at time t, where t is a positive integer; based on the orientation angle and motion speed of the candidate dynamic target at time t, data association is performed on the frame point cloud acquired at time t+1 to obtain the first association cluster associated with the candidate dynamic target at time t+1; pose estimation is performed on the first association cluster to obtain the center position and orientation angle of the first association cluster at time t+1; based on the motion direction corresponding to the orientation angle of the first association cluster at time t+1, data association is performed on the frame point cloud acquired at time t+2 to obtain the second association cluster; based on the motion speed and orientation angle of the candidate cluster, the first association cluster, and the second association cluster, a consistency judgment of motion state changes is performed.

[0097] Calculate the first velocity difference between the candidate cluster and the first associated cluster, and the second velocity difference between the first associated cluster and the second associated cluster; calculate the first orientation angle difference between the candidate cluster and the first associated cluster, and the second orientation angle difference between the first associated cluster and the second associated cluster; determine whether the first velocity difference and the second velocity difference are both less than a preset velocity difference, and whether the first orientation angle difference and the second orientation angle difference are both less than a preset orientation angle difference.

[0098] If both the first speed difference and the second speed difference are less than a preset speed difference, and both the first direction angle difference and the second direction angle difference are less than a preset direction angle difference, then the change in motion state is determined to be consistent.

[0099] In practice, the preset velocity difference and preset orientation angle difference differ depending on the type of dynamic target. Adjusting the preset velocity difference and preset orientation angle difference according to the type of dynamic target can adapt to various types of target detection.

[0100] When the changes in motion state are consistent, the type of dynamic target can be determined based on preset velocity difference, preset direction angle difference, or the aforementioned dynamic object detection threshold.

[0101] In practice, model fitting methods such as PE-CPD and PE-MSS can be used for pose estimation.

[0102] For candidate dynamic targets, model fitting methods such as PE-CPD and PE-MSS are used to estimate the target pose. The center position and orientation angle of the candidate dynamic target at time t are obtained. Subsequently, data association is performed with the point cloud at time t+1. Data association can be performed by combining the orientation angle of the candidate dynamic target at time t and the randomly sampled velocity to find associated targets, resulting in the first associated cluster A at time t+1. i For D i and A i When making rule judgments, it is necessary to consider that the target is moving dynamically and the state of the grid occupied by the target will switch between occupied and free states. Furthermore, the grid state changes differently as the candidate dynamic target moves closer to or further away from the vehicle.

[0103] For ease of description, the number of grid cells that are occupied at time t and free at time t+1 is denoted as OCCUPY(D). i The number of grid cells that are occupied at time t+1 and free at time t is denoted as FREE(D). i ).

[0104] When a candidate dynamic target approaches the vehicle, a portion of the grid in the occupied state at time t will become a free state at time t+1. The grid change satisfies the following first condition: FREE(D i ) greater than OCCUPY(D i ), and OCCUPY(D i () is greater than 0.

[0105] When a candidate dynamic target moves away from the vehicle, the free state mesh at time t will change to the occupied state at time t+1. The mesh change satisfies the following second condition: OCCUPY(D i ) greater than FREE(D i ), and FREE(D i () is greater than 0.

[0106] The relative motion relationship between the candidate dynamic target and the vehicle is determined by whether the changes in the grid state at adjacent time steps satisfy the first or second condition. Furthermore, the attitude information of the first association cluster can be determined based on the changes in the grid state at adjacent time steps.

[0107] Considering environmental occlusion and the inherent errors in lidar measurements, a weighted average can be applied to the number of occupied grid cells to improve the accuracy of determining grid state changes between adjacent time points. The weighted average of the number of grid cells that are occupied at time t and free at time t+1 yields WOCCUPY(Di) = w * OCCYPY(Di), where 0 <w<1。

[0108] In the event that a candidate dynamic target is close to the vehicle, FREE(D) i ) is greater than WOCCUPY(Di), and WOCCUPY(Di) is greater than 0.

[0109] When the candidate dynamic target is far away from the vehicle, WOCCUPY(Di) is greater than FREE(D). i ), and FREE(D i () is greater than 0.

[0110] For the first association cluster A obtained i Pose estimation is performed on it to obtain the first association cluster A at time t+1. i The location of the center and its orientation angle are determined, and the first association cluster A is used. i By associating the motion direction indicated by the orientation angle at time t+2, a second association cluster H is obtained. i Combined with D i With A iPosition and velocity information are compared between time t+1 and t+2 to determine the target type, i.e., the consistency of the three-dimensional motion velocity and direction of the associated target across three consecutive frames is used. DIFFS() represents the velocity difference, and DIFFA() represents the angular difference. If DIFFS() is less than a preset threshold T... S And DIFFA() is less than T A Then the second association cluster H at time t+2 can be... i If the cluster is determined to be of the preset dynamic target type, the type confirmation for this cluster is completed.

[0111] To facilitate a better understanding and implementation of the embodiments of the present invention by those skilled in the art, a dynamic vehicle is used as an example below, with reference to... Figures 4 to 6 The specific implementation process of the dynamic target detection method is explained. Among other things, Figure 4 A schematic diagram of a rule-based point cloud preprocessing method is provided in an embodiment of the present invention. Figure 5 A flowchart of a dynamic object detection process according to an embodiment of the present invention is provided. Figure 6 A rule-based dynamic vehicle confirmation flowchart is provided in an embodiment of the present invention.

[0112] Step S401: Obtain the original point cloud P.

[0113] The original point cloud P can be collected by the vehicle's onboard LiDAR. The original point cloud P is also the 3D point cloud data mentioned above.

[0114] Step S402: Point cloud ground segmentation.

[0115] In practical implementation, GPS combined with IMU can be used to obtain the vehicle's position and attitude, and extended Kalman filtering can be used for ground state tracking to perform point cloud ground segmentation. Alternatively, Gaussian process regression can be used to establish a ground elevation model for point cloud ground segmentation. The specific implementation method for ground point cloud segmentation can be found in the description of point cloud ground segmentation in the above embodiments, and will not be repeated here.

[0116] Step S403, obtain the ground point cloud p a .

[0117] Step S404, point cloud clustering.

[0118] In practical implementation, point cloud clustering can be performed based on polar coordinate grids. Let point cloud P = {p i} i＝1,...,n ,p i =(x i ,y i ,z iFirst, the ground point cloud is projected onto the horizontal plane to obtain a two-dimensional projected point cloud. Then, the coordinates of the two-dimensional projected point cloud are converted into polar coordinates using the above formulas (3) and (4). i ,θ i ).

[0119] Furthermore, with the center of the lidar as the origin of the coordinate system, the horizontal plane is divided into multiple sector segments with equal central angles, and simultaneously, the plane is divided into multiple annular segments with equal ring widths along the radial direction. These two segmentation methods form a polar coordinate grid on the horizontal plane. The polar coordinate grid is identified by an index, and the calculation method for the polar coordinate grid index can be found in the above formulas (1) and (2).

[0120] Based on the polar coordinates of the 2D projected point cloud and the index of the polar coordinate grid, it can be determined whether the grid contains the point cloud. If the grid contains the 2D projected point cloud, the grid state is occupied and marked as 1; if the grid does not contain the 2D projected point cloud, the grid state is free and marked as 0. The polar coordinate grid is unfolded radially and circumferentially. The unfolded grid diagram is obtained. Based on the grid state of each grid in the unfolded grid diagram, the 2D projected point cloud is clustered.

[0121] For details on the specific implementation of two-dimensional point cloud clustering, please refer to the description in step S12 of the above embodiments, which will not be repeated here.

[0122] Step S405, obtain cluster {C i}

[0123] Step S406: Using the rules set for the attributes of the target to be detected, candidate clusters are selected and filtered to obtain candidate cluster C. i .

[0124] In practice, candidate clustering can be filtered according to set rules, such as configuring target attribute filtering conditions by setting corresponding rules based on the attributes of the target to be detected.

[0125] For example, the target attribute filtering condition is LESS(NUMBER(C i ),T N ), LESS (LENGTH (C) i ), T L ), LESS (HEIGHT (C) i ),T H And LESS(WIDTH(C) i ),T D Among them, LESS(NUMBER(C) i ),T N ) represents the number of point clouds in the cluster. NUMBER(C) iLess than the set quantity T N LESS(LENGTH(C) i ),T L LENGTH(C) represents the length of the cluster. i Less than the set length T L LESS (HEIGHT (C) i ),T H WIDTH(C) represents the height of the cluster. i Less than the set height T D LESS(WIDTH(C) i ),T D HEIGHT(C) represents the width of the cluster. i (less than the set width T) H .

[0126] Clusters that meet the target attribute selection criteria are designated as candidate clusters. These candidate clusters then proceed to the subsequent dynamic object detection process. The following section combines... Figure 5 The process of detecting dynamic objects is explained.

[0127] Step S501, coordinate transformation.

[0128] Specifically, based on the candidate clusters {C} obtained in step S406 i First, using the vehicle's location information collected and the vehicle's pose information obtained by the IMU, the coordinate system of the point cloud for two consecutive frames at time t and time t-1 is determined to be in the coordinate system at time t.

[0129] Step S502: Obtain point clouds of adjacent frames with the same origin.

[0130] Step S503, polar coordinate grid mapping.

[0131] Then, the coordinates of the point clouds of adjacent frames with the same origin are projected onto the horizontal plane to perform polar coordinate grid mapping.

[0132] Step S504: Virtual scan mapping.

[0133] Candidate cluster C can be searched by using a virtual scan map. i The grid positions and number of cells occupied at time t and time t-1.

[0134] Step S505: According to the set rules, determine whether the mesh whose mesh state has changed exceeds the dynamic object detection threshold T. D .

[0135] At time t, search for candidate clusters C. iThe occupied grid positions and quantities are compared with the grid occupancy at the same position at time t-1. Difference is performed to find the number of grids with different occupancy states, i.e., the grids whose state has changed. This is then calculated using the function DIFF(C). i ) indicates that, compared with the dynamic object detection threshold T D For comparison, the calculation formula is as follows. If the number of changing grid cells exceeds T... D Then the candidate cluster is judged as candidate vehicle D. i .

[0136] Among them, the dynamic object detection threshold T D It can be obtained using the above formula (5), which will not be repeated here.

[0137] Step S506, obtain candidate vehicle D i .

[0138] For the detected candidate vehicle D i Typically, feature extraction is performed, and the model is fitted to the constructed vehicle measurement model. Candidate vehicle types are then determined based on a pre-defined objective function. The rule-based dynamic vehicle detection process is described in [link to documentation]. Figure 6 As shown below. (Combined with...) Figure 6 The process for confirming dynamic vehicles is explained.

[0139] Step S601: Perform pose estimation for the candidate vehicles.

[0140] In specific implementation, based on the candidate vehicle D obtained in step S506 i For candidate vehicle D i Pose estimation is performed for candidate vehicle D. i Target pose estimation can be performed using model fitting methods such as PE-CPD and PE-MSS.

[0141] Step S602, obtain candidate vehicle D at time t. i The center position and the direction angle

[0142] Step S603, data association.

[0143] Step S604: Obtain the association cluster A associated at time t+1. i

[0144] Based on the orientation angle and velocity of the candidate dynamic target at time t, the associated cluster A at time t+1 is obtained. i That is, we obtain the first association cluster A. i The speed of movement can be the speed of random sampling.

[0145] Step S605, for D i and Ai Perform rule-based judgment.

[0146] In specific implementation, for D i and A i The rules are used to make judgments. Considering that the target is moving dynamically, the grid it occupies will switch between occupied and free states. Furthermore, the grid state changes differently as the target moves closer to or further away from the vehicle. Therefore, the cases are divided.

[0147] When the target is close to the vehicle, the mesh state change satisfies the following first condition: MORE(FREE(D i ),OCCUPY(D i And MORE(OCCUPY(D) i ), 0).

[0148] Among them, MORE(FREE(D) i ),OCCUPY(D i )) means FREE(D i ) greater than OCCUPY(D i MORE(OCCUPY(D) i ), 0) represents OCCUPY(D i () is greater than 0.

[0149] When the target is far away from the vehicle, the mesh state change satisfies the following second condition: MORE(OCCUPY(D i ), FREE(D i And MORE(FREE(D) i ), 0).

[0150] MORE(OCCUPY(D i ), FREE(D i )) represents OCCUPY(D i ) Big FREE (D i MORE(FREE(D)) i ), 0) means FREE(D i () is greater than 0.

[0151] Among them, OCCUPY(D i FREE(D) represents the number of grid cells that are occupied at time t and free at time t+1. i ) represents the number of grid cells that are occupied at time t+1 and free at time t.

[0152] Considering environmental occlusion and the inherent errors in lidar measurements, the number of occupied state grids is weighted before being compared with the number of free states. The weighted number of grids that are occupied at time t and free at time t+1 is then used to obtain WOCCUPY(D). i )=w*OCCYPY(D i ), 0 <w<1。

[0153] Therefore, for the case where the target is close to the vehicle, when weighting the number of grids that are occupied at time t and free at time t+1, the condition FREE(D) should be satisfied. i ) is greater than WOCCUPY(Di), and WOCCUPY(Di) is greater than 0.

[0154] When the target is far from the vehicle, and considering the number of grid cells that are occupied at time t and free at time t+1, the weighted average should satisfy WOCCUPY(Di) greater than FREE(D). i ), and FREE(D i () is greater than 0.

[0155] Step S606, for the first association cluster A i Perform pose estimation.

[0156] Step S607, obtain the first association cluster A at time t+1. i The center position and its orientation angle.

[0157] Step S608: Perform data association.

[0158] Step S609, obtain the second association cluster H at time t+2. i .

[0159] In practical implementation, the first association cluster A at time t+1 can be used as a basis. i The location of the center and its orientation angle are determined, and the first association cluster A is used. i By associating the motion direction indicated by the orientation angle at time t+2, a second association cluster H is obtained. i .

[0160] Step S610, combined with D i With A i Using location and speed information, rule-based judgments are performed to determine the dynamic vehicle H. i .

[0161] In practical implementation, it can be combined with D i With A iPosition and velocity information are compared between time t+1 and t+2 to determine the target type, i.e., the consistency of the three-dimensional motion velocity and direction of the associated target across three consecutive frames is used. DIFFS() represents the velocity difference, and DIFFA() represents the angular difference. If DIFFS() is less than a preset threshold T... S And DIFFA() is less than T A Then the second association cluster H at time t+2 can be... i Determined to be a dynamic vehicle H i This completes the type confirmation for this cluster.

[0162] Understandably, the detection process for other types of dynamic targets, such as pedestrians, cyclists, and animals, can refer to the above description, and will not be listed here.

[0163] This invention also provides a dynamic target detection device, referring to... Figure 7 The present invention provides a schematic diagram of the structure of a dynamic target detection device according to an embodiment of the present invention. The dynamic target detection device 70 may include:

[0164] Projection unit 71 is used to project the acquired three-dimensional point cloud data of the environment around the vehicle onto a polar coordinate grid to obtain a two-dimensional projected point cloud. The polar coordinate grid is constructed on a horizontal plane with the center of the lidar as the origin.

[0165] Clustering unit 72 is used to perform point cloud clustering on the two-dimensional projected point cloud according to the position of the two-dimensional projected point cloud on the polar coordinate grid, and obtain point cloud clustering results;

[0166] The dynamic target detection unit 73 is used to perform dynamic target detection based on the point cloud clustering results.

[0167] In specific implementation, the working principle and workflow of the dynamic target detection device 70 can be referred to the description in the dynamic target detection method provided in any of the above embodiments of the present invention, and will not be repeated here.

[0168] This invention also provides a storage medium, which is a non-volatile or non-transient storage medium, on which a computer program is stored. When the computer program is run by a processor, it executes the steps of any of the above-described dynamic target detection methods.

[0169] This invention also provides a terminal, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the steps of any of the above-described dynamic target detection methods when running the computer program.

[0170] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in any computer-readable storage medium, which may include ROM, RAM, disk, or optical disk, etc.

[0171] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A dynamic target detection method, characterized in that, include: The three-dimensional point cloud data of the surrounding environment of the vehicle is projected onto a polar coordinate grid to obtain a two-dimensional projected point cloud. The three-dimensional point cloud data is collected by a lidar, and the polar coordinate grid is constructed on a horizontal plane with the center of the lidar as the origin. Based on the position of the two-dimensional projected point cloud on the polar coordinate grid, point cloud clustering is performed on the two-dimensional projected point cloud to obtain the point cloud clustering result; Based on the point cloud clustering results, dynamic target detection is performed; The step of clustering the two-dimensional projected point cloud based on its position on the polar coordinate grid includes: Based on the position of the two-dimensional projected point cloud on the polar coordinate grid, the occupancy status of each grid in the polar coordinate grid is marked to obtain a polar coordinate grid marked with grid status. The grid status includes an occupied state and a free state. The occupied state means that there is a two-dimensional projected point cloud in the grid, and the free state means that there is no two-dimensional projected point cloud in the grid. The polar coordinate grid marked with the grid state is unfolded radially and circumferentially to obtain the unfolded grid diagram. Based on the grid state of each grid in the unfolded grid diagram, point cloud clustering is performed on the two-dimensional projected point cloud; The step of clustering the two-dimensional projected point cloud based on the grid state of each grid in the unfolded grid diagram includes: Using each grid cell in the occupied state as the center, draw a circle with a set threshold as the radius. Mark all occupied grid cells inside the circle as belonging to the same cluster, until all grid cells in the occupied state are traversed in a set order. During point cloud clustering, if a grid cell in the circle has already obtained a cluster category, then all grid cells in the occupied state enclosed by the circle belong to the same cluster category. If none of the grid cells in the circle have a cluster category, then all grid cells in the occupied state in the circle are marked as a new cluster category.

2. The dynamic target detection method as described in claim 1, characterized in that, The method for constructing the polar coordinate grid includes: Using the center of the lidar as the origin of the coordinate system, the horizontal plane is divided along the circumferential and radial directions to obtain the polar coordinate grid.

3. The dynamic target detection method as described in claim 2, characterized in that, The division of the horizontal plane along the circumferential and radial directions respectively includes: The horizontal plane is divided into multiple sector-shaped sections along the circumference with equal central angles; The horizontal plane is divided into multiple annular segments with equal ring widths along the radial direction.

4. The dynamic target detection method as described in claim 1, characterized in that, The step of performing point cloud clustering on the two-dimensional projection points based on the grid state of each grid in the unfolded grid diagram includes: The unfolded mesh image is binarized to obtain a binary image; Morphological dilation is performed on the binary image; Image connectivity is extracted from the binary image after morphological dilation to obtain multiple connected regions; The two-dimensional projected point clouds corresponding to the grids in each connected region are marked as the same cluster.

5. The dynamic target detection method as described in claim 4, characterized in that, The step of extracting connected components from the morphologically dilated binary image includes: The neighborhood connectivity method is used to search for connected regions within the binary graph after morphological dilation.

6. The dynamic target detection method as described in claim 1, characterized in that, Also includes: Before projecting the acquired 3D point cloud data of the vehicle's surrounding environment onto the polar coordinate grid, the 3D point cloud data is filtered to remove ground point cloud data, resulting in ground point cloud.

7. The dynamic target detection method according to any one of claims 1 to 6, characterized in that, The dynamic target detection based on the point cloud clustering results includes: Based on the relationship between the point cloud clustering results and the preset target attribute filtering conditions, candidate clusters are filtered to obtain candidate clusters; Based on the positions of the two-dimensional projected point clouds corresponding to the candidate clusters in the adjacent frame point clouds in the polar coordinate grid, dynamic target detection is performed, and candidate dynamic targets are determined based on the dynamic target detection results. The candidate dynamic targets are identified through dynamic target confirmation to determine the type of the dynamic target.

8. The dynamic target detection method as described in claim 7, characterized in that, The step of filtering candidate clusters based on the relationship between the point cloud clustering results and preset target attribute filtering conditions to obtain candidate clusters includes: Determine whether the point cloud clustering result meets the target attribute filtering conditions; Point cloud clusters whose clustering results satisfy the target attribute filtering conditions are selected as candidate clusters. The target attribute filtering conditions include: the number of point clouds in the cluster meets a set quantity and the size of the cluster meets a set size.

9. The dynamic target detection method as described in claim 7, characterized in that, The step of performing dynamic target detection based on the positions of the two-dimensional projected point clouds corresponding to the candidate clusters in the polar coordinate grid, and determining candidate dynamic targets based on the dynamic target detection results, includes: Unify the coordinates of point clouds in adjacent frames to the same coordinate system; Based on the position changes of the point clouds corresponding to the candidate clusters in the polar coordinate grid in the adjacent frames, it is determined whether the object corresponding to the candidate cluster is a dynamic target; If the target is dynamic, the candidate cluster is determined as a candidate dynamic target.

10. The dynamic target detection method as described in claim 9, characterized in that, The step of determining whether the object corresponding to the candidate cluster is a dynamic target based on the position changes of the point clouds corresponding to the candidate clusters in the polar coordinate grid in adjacent frames includes: The grid occupancy of the point cloud corresponding to the candidate cluster in one frame of the adjacent frame in the polar coordinate grid is compared with the grid occupancy of the point cloud corresponding to the candidate cluster in another frame of the adjacent frame in the polar coordinate grid to obtain the number of grids whose grid occupancy has changed. When the number of grids whose grid occupancy changes exceeds the set dynamic object detection threshold, the candidate cluster is determined as the candidate dynamic target.

11. The dynamic target detection method as described in claim 10, characterized in that, The dynamic object detection threshold is related to the shortest side length of the dynamic target projection, the relative distance between the center of the dynamic target and the lidar, and the resolution of the polar coordinate grid.

12. The dynamic target detection method as described in claim 7, characterized in that, The process of confirming the dynamic target of the candidate dynamic target to obtain the type of the dynamic target includes: Based on N consecutive frames of point cloud, a consistency judgment is made on the motion state changes of the candidate dynamic target, where N is a positive integer greater than 1; The type of the dynamic target is determined based on the consistency judgment result.

13. The dynamic target detection method as described in claim 12, characterized in that, The step of making a consistency judgment on the motion state changes of the candidate dynamic target based on N consecutive frames of point cloud includes: The candidate dynamic target is pose estimated to obtain the center position and orientation angle of the candidate dynamic target at time t, where t is a positive integer; Based on the orientation angle and speed of the candidate dynamic target at time t, the frame point cloud acquired at time t+1 is correlated to obtain the first correlation cluster of the candidate dynamic target at time t+1. Pose estimation is performed on the first associated cluster to obtain the center position and orientation angle of the first associated cluster at time t+1; Based on the motion direction corresponding to the direction angle of the first association cluster at time t+1, the frame point cloud acquired at time t+2 is associated with data to obtain the second association cluster; Based on the motion speed and direction angle of the candidate clusters, the first associated clusters, and the second associated clusters, a consistency judgment is made on the changes in motion state.

14. The dynamic target detection method as described in claim 13, characterized in that, The step of determining the consistency of motion state changes based on the motion speed and direction angle of the candidate clusters, the first associated clusters, and the second associated clusters includes: Calculate the first velocity difference between the candidate cluster and the first associated cluster, and the second velocity difference between the first associated cluster and the second associated cluster, respectively; Calculate the first directional angle difference between the candidate cluster and the first associated cluster, and the second directional angle difference between the first associated cluster and the second associated cluster, respectively; Determine whether both the first speed difference and the second speed difference are less than a preset speed difference, and whether both the first direction angle difference and the second direction angle difference are less than a preset direction angle difference; If both the first speed difference and the second speed difference are less than a preset speed difference, and both the first direction angle difference and the second direction angle difference are less than a preset direction angle difference, then the change in motion state is determined to be consistent.

15. A dynamic target detection device, characterized in that, include: The projection unit is used to project the acquired three-dimensional point cloud data of the environment around the vehicle onto a polar coordinate grid to obtain a two-dimensional projected point cloud. The three-dimensional point cloud data is collected by the lidar, and the polar coordinate grid is constructed on a horizontal plane with the center of the lidar as the coordinate origin. A clustering unit is used to perform point cloud clustering on the two-dimensional projected point cloud based on the position of the two-dimensional projected point cloud on the polar coordinate grid, so as to obtain the point cloud clustering result; A dynamic target detection unit is used to perform dynamic target detection based on the point cloud clustering results; The clustering unit is used to mark the occupancy status of each grid in the polar coordinate grid according to the position of the two-dimensional projected point cloud on the polar coordinate grid, so as to obtain a polar coordinate grid marked with grid status. The grid status includes an occupied state and a free state. The occupied state means that there is a two-dimensional projected point cloud in the grid, and the free state means that there is no two-dimensional projected point cloud in the grid. The polar coordinate grid marked with grid states is unfolded radially and circumferentially to obtain an unfolded grid map. Point cloud clustering is performed on the two-dimensional projected point cloud based on the grid states of each grid in the unfolded grid map. This point cloud clustering includes: using each grid in the occupied state as the center, drawing a circle with a set threshold as the radius, and marking all occupied grids within the circle as belonging to the same cluster, until all grids in the occupied state are traversed in a set order. During point cloud clustering, if a grid within the circle has already obtained a cluster category, then all grids in the occupied state enclosed by the circle belong to that cluster category; if none of the grids within the circle have a cluster category, then all grids in the occupied state within the circle are marked as a new cluster category.

16. A storage medium, said storage medium being a non-volatile storage medium or a non-transient storage medium, wherein a computer program is stored thereon, characterized in that, The computer program, when run by a processor, performs the steps of the dynamic target detection method according to any one of claims 1 to 14.

17. A terminal comprising a memory and a processor, wherein the memory stores a computer program capable of running on the processor, characterized in that, When the processor runs the computer program, it performs the steps of the dynamic target detection method according to any one of claims 1 to 14.

Images