3D target object positioning method and device, and electronic device

Abstract

The application discloses a 3D target object positioning method and device and electronic equipment, and belongs to the field of intelligent transportation. The method comprises the following steps: generating a first clustering center point set according to current frame point cloud data of a preset road region and generating a second clustering center point set according to last frame point cloud data of the preset road region; determining N clustering centers corresponding to N 3D target objects in the second clustering center point set according to the first clustering center point set and the second clustering center point set; establishing N sector regions corresponding to the N 3D target objects in the current frame according to the N clustering centers, current moving distances of the N 3D target objects and current turning radii; and determining N target clustering centers in the N sector regions as position information of the N 3D target objects at the current moment according to the current frame point cloud data. The application solves the technical problem that the accuracy of repositioning of the disappeared 3D target object is not high.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent transportation, and in particular to a 3D target object positioning method and device and electronic equipment. BACKGROUND

[0002] 3D target object tracking often appears 3D target object flickering or disappearing due to reasons such as point cloud sparseness, jitter, splicing, etc., thereby leading to poor tracking effect.

[0003] Once the 3D target object disappears in the process of tracking the 3D target object, the disappeared 3D target object needs to be repositioned. One is to use the 3D tracking trajectory to predict the position of the disappeared 3D target object, and continue tracking by taking the predicted position as the current position thereof. The other is to use object matching and mapping between 2D image and 3D image to obtain the position of the current 3D target object in the 3D image to continue tracking. However, the first way is effective for smooth moving objects, but has poor accuracy for non-stationary objects such as vehicles changing lanes. The second way has a large amount of calculation and requires high mapping accuracy between 2D target objects and 3D target objects, and is not suitable for projects with low computing power and low latency requirements. Therefore, the prior art has low accuracy in repositioning the disappeared 3D target object or has high computing power requirement. SUMMARY

[0004] The embodiments of the present application provide a 3D target object positioning method, device and electronic equipment, which at least solve the technical problem of low accuracy in repositioning the disappeared 3D target object in the prior art.

[0005] In a first aspect, the embodiments of the present application provide a 3D target object positioning method, comprising: generating a first cluster center point set according to current frame point cloud data of a preset road area, and generating a second cluster center point set according to last frame point cloud data of the preset road area, the current frame point cloud data having N 3D target objects disappearing relative to the last frame point cloud data, N being a positive integer; determining N cluster centers corresponding to the N 3D target objects in the second cluster center point set according to the first cluster center point set and the second cluster center point set; establishing N sector areas corresponding to the N 3D target objects in the current frame according to the N cluster centers, current moving distances of the N 3D target objects, and current turning radii.

[0006] Determining N target cluster centers in the preset N sector areas as position information of the N 3D target objects at the current time according to the current frame point cloud data.

[0007] Optionally, the generating a first clustering center point set according to current frame point cloud data of a preset road region and generating a second clustering center point set according to last frame point cloud data of the road region comprises:

[0008] filtering background point cloud information from the current frame point cloud data and the last frame point cloud data to obtain first residual point cloud data of the current frame point cloud data and second residual point cloud data of the last frame point cloud data;

[0009] performing clustering on the first residual point cloud data to obtain the first clustering center point set;

[0010] performing clustering on the second residual point cloud data to obtain the second clustering center point set.

[0011] Optionally, the filtering background point cloud information from the current frame point cloud data and the last frame point cloud data comprises: performing difference calculation on the background mask of the preset road region and the current frame point cloud data to filter background point cloud information in the current frame point cloud data; and performing difference calculation on the background mask of the preset road region and the last frame point cloud data to filter background point cloud information in the last frame point cloud data.

[0012] Optionally, the determining N clustering centers corresponding to the N 3D target objects in the second clustering center point set according to the first clustering center point set and the second clustering center point set comprises: removing clustering centers in which target class 3D objects are located and removing clustering centers in which noise points are located in the second clustering center point set to obtain N clustering centers corresponding to the N 3D target objects in the second clustering center point set; wherein the target class 3D object is a 3D target object that does not disappear at the current moment.

[0013] Optionally, if there is at least one pedestrian object in the N 3D target objects, after the determining N clustering centers corresponding to the N 3D target objects in the second clustering center point set according to the first clustering center point set and the second clustering center point set, the method further comprises: determining a circular region with a clustering center of the pedestrian object in the second clustering center point set as a center; wherein a size parameter of the circular region is determined in advance based on a walking speed of the pedestrian object.

[0014] Optionally, before the N sector regions are established according to the N cluster centers, the current moving distances of the N 3D target objects, and the current turning radii, the method further comprises: determining the driving parameters and the bounding box lengths of the N 3D target objects at a previous time through tracking of the trajectories of the N 3D target objects; determining the moving distances of the N 3D target objects within a time period for collecting a frame of point cloud data as the current moving distances according to the driving parameters of the N 3D target objects at the previous time; and determining the turning radii of the N 3D target objects as the current turning radii according to the bounding box lengths of the N 3D target objects at the previous time.

[0015] Optionally, the determining the N target cluster centers in the N sector regions from the current frame of point cloud data as the position information of the N 3D target objects at the current time comprises: removing the cluster center where the target 3D object is located from the first cluster center set to obtain M candidate cluster centers in the first cluster center set, M being an integer greater than 1; selecting the cluster center within the sector region of the current 3D target object from the M candidate cluster centers with the N 3D target objects as the current 3D target object respectively; if there are K candidate cluster centers in the sector region of the current 3D target object, determining the target cluster center corresponding to the current 3D target object from the K candidate cluster centers as the position information of the current 3D target object at the current time, K being a positive integer; and if there is no candidate cluster center in the sector region of the current 3D target object, re-clustering the point cloud data in the sector region in the current frame of point cloud data based on the adjusted clustering parameters, and determining the position information of the current 3D target object at the current time according to at least one new cluster center generated by the re-clustering.

[0016] Optionally, the determining the target cluster center corresponding to the current 3D target object from the K candidate cluster centers comprises: if there are multiple candidate cluster centers in the sector region of the current 3D target object, determining the target cluster center corresponding to the current 3D target object from the multiple candidate cluster centers through optimal matching; and if there is a single candidate cluster center in the sector region of the current 3D target object, determining the single candidate cluster center as the target cluster center corresponding to the current 3D target object.

[0017] Optionally, the method further comprises: in a state where there is no 3D target object in the preset road region, acquiring multiple frames of historical point cloud data of a position where the road region is located; fusing the multiple frames of historical point cloud data to obtain fused point cloud data; and processing the fused point cloud data to form a voxelized mask of the road region and saving the voxelized mask.

[0018] In a second aspect, an embodiment of the present application provides a positioning device for a 3D target object, comprising: a clustering unit configured to generate a first set of clustering center points according to current frame point cloud data of a preset road region, and generate a second set of clustering center points according to last frame point cloud data of the preset road region, the current frame point cloud data having N 3D target objects disappeared relative to the last frame point cloud data, N being a positive integer; a center matching unit configured to determine N clustering centers corresponding to the N 3D target objects in the second set of clustering center points according to the first set of clustering center points and the second set of clustering center points; a first region prediction unit configured to establish N sector regions corresponding to the N 3D target objects in a current frame according to the N clustering centers, current moving distances of the N 3D target objects, and a current turning radius; and a position determination unit configured to determine N target clustering centers in the N sector regions as position information of the N 3D target objects at a current time according to the current frame point cloud data.

[0019] In a third aspect, an embodiment of the present application provides an electronic device, comprising a memory, a processor, and code stored in the memory and executable on the processor, and the processor executes the code to implement the method in any of the embodiments of the first aspect.

[0020] The one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0021] If the current frame point cloud data of the preset road region has 3D target objects disappeared relative to the last frame point cloud data, the first set of clustering center points of the current frame point cloud data and the second set of clustering center points of the last frame point cloud data are generated according to the current frame point cloud data and the last frame point cloud data of the preset road region, the N clustering centers corresponding to the N 3D target objects disappeared in the second set of clustering center points are determined according to the first set of clustering center points and the second set of clustering center points, and the N sector regions corresponding to the N 3D target objects in the current frame are established according to the N clustering centers, current moving distances of the N 3D target objects, and a current turning radius, and the N target clustering centers in the N sector regions are determined as the position information of the N 3D target objects at a current time according to the current frame point cloud data. The above technical solution finds the clustering center corresponding to the 3D target objects disappeared in the current frame as the position information of the 3D target objects disappeared at a current time through clustering, thereby improving the accuracy of repositioning the 3D target objects disappeared.

[0022] Moreover, since the mapping between the 2D target object and the 3D target object is not needed, the technical solution provided in the embodiments of the present application requires less computing power, can reduce the delay, and can be applied to a scene with low computing power. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort based on these drawings.

[0024] Figure 1 A flow chart of the 3D target object positioning method in the embodiments of the present application;

[0025] Figure 2 A schematic diagram of the sector area corresponding to the 3D target object in the embodiments of the present application;

[0026] Figure 3 A structural schematic diagram of the positioning device of the 3D target object in the embodiments of the present application;

[0027] Figure 4 A structural schematic diagram of the electronic device in the embodiments of the present application. DETAILED DESCRIPTION

[0028] In order to better understand the above technical solutions, the above technical solutions will be described in detail in combination with the drawings of the specification and specific embodiments.

[0029] Firstly, it should be noted that the term "and / or" appearing in this paper only describes the association relationship of the associated objects, which means that there can be three kinds of relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents an "or" relationship between the front and rear associated objects.

[0030] The embodiments of the present application provide a 3D target object positioning method, which can be applied to an RSU (Road Side Unit). The RSU is a device installed on the roadside in an ETC system, which uses DSRC (Dedicated Short Range Communication) technology to communicate with an OBU (On Board Unit) to realize the functions of vehicle identity recognition, electronic deduction, etc.

[0031] Reference Figure 1 As shown in the drawings, the 3D target object positioning method provided by the embodiments of the present application includes the following steps:

[0032] S101, generating a first cluster center point set according to current frame point cloud data of a preset road area, and generating a second cluster center point set according to last frame point cloud data of the preset road area.

[0033] It can be understood that the 3D target tracking model based on deep learning detects and tracks the 3D target object in the preset road area, and assigns the same ID number to the same target object. In the process of tracking the 3D target object corresponding to each ID number by the 3D target tracking model based on deep learning, if the current position information of the 3D target object corresponding to one or more ID numbers in the preset road area cannot be obtained, it is represented that the 3D target object disappears at the current time, and the first cluster center point set is generated according to the current frame point cloud data of the preset road area, and the second cluster center point set is generated according to the last frame point cloud data of the road area.

[0034] It should be noted that the current frame point cloud data is a frame of point cloud data relative to the last frame point cloud data at the moment when at least one tracked 3D target object disappears in the preset road area. And it should be understood that each 3D target object tracked in the preset road area will not be newly added and disappeared in the actual scene, but because the 3D target tracking model based on deep learning detects the 3D target object in the current frame point cloud data, the sparse current frame point cloud data or the occlusion between different 3D target objects will cause the phenomenon of missing detection and tracking loss of 3D target objects, and then the 3D target tracking model based on deep learning cannot detect the 3D target object from the current frame point cloud data, so it cannot obtain the recognition box of the 3D target object in the current frame point cloud data, resulting in the disappearance of the tracked 3D target object at the current time.

[0035] It can be understood that the 3D target tracking model based on deep learning tracks the 3D target object in the preset road area, specifically, each 3D target object detected by the 3D target tracking model based on deep learning is numbered and tracked, and the running track of the 3D target object corresponding to each ID code is recorded.

[0036] The preset road area can be the central region of the road or the central region of the radar detection range determined from the point cloud data collected by the 3D sensing device according to the road traffic condition. The preset road area can exclude buildings, trees, poles, etc. outside the road area.

[0037] In the process of tracking the 3D target object in the preset road area by the 3D target tracking model based on deep learning, if at least one 3D target object disappears at the current time, further comprising: calculating the speed and heading angle of the 3D target object at the last time (i.e. the last time before disappearing) by the running track of the 3D target object, for subsequent use when establishing the corresponding sector area of the 3D target object in the current frame.

[0038] In step S101, in order to avoid the interference of background point cloud information on the clustering process, it is necessary to filter out the background point cloud information of the current frame point cloud data and the background point cloud information of the last frame point cloud data. Among them, the background point cloud information of the current frame point cloud data and the background point cloud information of the last frame point cloud data refer to the point cloud information other than the 3D target objects such as vehicles and pedestrians which are tracked. Specifically, the background point cloud information includes the point cloud occupied by the road, the buildings on both sides of the road, the trees, the billboards, the road signs and the like.

[0039] Specifically, by filtering out the background point cloud information in the current frame point cloud data, the first remaining point cloud data of the current frame point cloud data is obtained; by filtering out the background point cloud information in the last frame point cloud data, the second remaining point cloud data of the last frame point cloud data is obtained. It should be noted that the first remaining point cloud data is the point cloud information occupied by the vehicles and pedestrians in the current frame point cloud data. The second remaining point cloud data is the point cloud information occupied by the vehicles and pedestrians in the last frame point cloud data.

[0040] After obtaining the first remaining point cloud data, the first remaining point cloud data is clustered to obtain a first set of clustering center points; after obtaining the second remaining point cloud data, the second remaining point cloud data is clustered to obtain a second set of clustering center points. Among them, the first set of clustering center points includes the clustering center corresponding to each 3D target object, and the second set of clustering center points includes the clustering center corresponding to each 3D target object. It should be understood that whether a 3D target object disappears or does not disappear at the current moment, clustering the current frame point cloud data and the last frame point cloud data respectively will obtain the corresponding clustering center, because: even if a 3D target object A occupies sparse point cloud in the current frame point cloud data, resulting in that the 3D target tracking model based on deep learning cannot detect the bounding box of the 3D target object A in the current frame point cloud data, but by clustering the current frame point cloud data, the clustering center of the 3D target object A can still be obtained.

[0041] In order to filter out the background point cloud information in the current frame point cloud data and the background point cloud information in the last frame point cloud data, the background mask of the preset road region can be calculated by difference with the current frame point cloud data to filter out the background point cloud information in the current frame point cloud data, thereby obtaining the first remaining point cloud data; the background mask of the preset road region is calculated by difference with the last frame point cloud data to filter out the background point cloud information in the last frame point cloud data, thereby obtaining the second remaining point cloud data.

[0042] It should be noted that the background mask of the preset road region is pre-created and saved locally, and the background mask can be generated by using multiple frames of historical point cloud data collected in the preset road region. When it is necessary to filter out the background point cloud information, the background mask of the preset road region can be called.

[0043] Specifically, the background mask of the preset road region can be created in the following manner: in a state where there is no 3D target object in the preset road region (for example, in a state where there is no pedestrian and vehicle), a plurality of frames of historical point cloud data of a location where the road region is located is acquired; the plurality of frames of historical point cloud data is fused to obtain dense fused point cloud data; the fused point cloud data is processed to form a voxelized mask of the preset road region and is saved.

[0044] The processing of the fused point cloud data includes: by performing a closing operation on the fused point cloud data, each discrete point in the fused point cloud data can be connected to form a 3D ground solid shape of the road region; and the 3D ground solid shape is converted into a voxelized mask of the road region. In order to reduce the influence of the shaking of poles and trees caused by wind, the 3D ground solid shape can be dilated and then converted into a voxel representation closest to the object as the voxelized mask of the road region.

[0045] S102: According to the first cluster center point set and the second cluster center point set, N cluster centers corresponding to the N 3D target objects in the second cluster center point set are determined.

[0046] In step S102, the N cluster centers corresponding to the N 3D target objects that disappear are determined from the second cluster center point set. Specifically, the cluster center where the target class 3D object is located and the cluster center where the noise point in the second cluster center point set is located can be removed from the second cluster center point set, and the remaining cluster centers in the second cluster center point set are the N cluster centers corresponding to the N 3D target objects that disappear, wherein the target class 3D object refers to the 3D target object that does not disappear at the current time.

[0047] Suppose a 3D target object existing in an actual scene is a vehicle, the 3D target tracking model based on deep learning can detect the recognition box corresponding to the vehicle from the last frame of point cloud data, or can detect the recognition box corresponding to the vehicle from the current frame of point cloud data, which represents that the vehicle does not disappear at the current time, so it is not necessary to reposition the vehicle at the current time, and therefore the cluster center where the vehicle is located needs to be removed from the second cluster center point set.

[0048] S103: According to the N cluster centers, the current moving distance of the N 3D target objects, and the current turning radius, N sector regions corresponding to the N 3D target objects in the current frame are established.

[0049] Specifically, after obtaining the N cluster centers corresponding to the N 3D target objects that disappear from the second cluster center point set, the N cluster centers are taken as current cluster centers respectively, the multi-dimensional parameters of the current 3D target objects corresponding to the current cluster centers are obtained, and the fan-shaped regions corresponding to the current 3D target objects in the current frame are determined according to the multi-dimensional parameters of the current 3D target objects, and the process is sequentially performed, so that the N fan-shaped regions corresponding to the N 3D target objects in the current frame are obtained.

[0050] It should be noted that the 3D target object not disappearing at the current moment means that the current frame point cloud data does not disappear relative to the last frame point cloud data, that is, the 3D target tracking model based on deep learning can detect the recognition box corresponding to the 3D target object in both the current frame point cloud data and the last frame point cloud data.

[0051] It should be noted that the establishment of the N fan-shaped regions in step S103 can be only for the N 3D target objects being vehicle objects.

[0052] If there is at least one pedestrian object in the N 3D target objects, a circular region is established for the pedestrian object.

[0053] Specifically, a circular region with the cluster center of the pedestrian object in the second cluster center point set as the center is determined, and the size parameter of the circular region is determined in advance based on the walking speed of the pedestrian object. The walking speed here can be a general speed of pedestrians of various ages, or a speed of the pedestrian object determined by the 3D target tracking model based on deep learning before the disappearance. Since the walking speed of the pedestrian is slow, usually less than 3.6 m / s, and the turning radius is almost zero, the size parameter of the circular region can be the radius of the circular region, which is determined based on the point cloud frame rate and the walking speed. Assuming that the point cloud frame rate is 10 Hz, the radius of the circular region can be a value slightly larger than 3.6 m, such as 0.4 m.

[0054] If the 3D target object is a vehicle object, such as a motor vehicle or a non-motor vehicle, the driving parameters and the vehicle length of the vehicle object at the last moment before the disappearance (i.e., at the last moment before the disappearance) need to be considered.

[0055] Specifically, the 3D target tracking model based on deep learning tracks the trajectory of each 3D target object corresponding to an ID number to determine the driving parameters and the recognition box length of the N 3D target objects at the last moment; the moving distance of the N 3D target objects within the time consumed for collecting one frame of point cloud data is determined as the current moving distance according to the driving parameters of the N 3D target objects at the last moment; and the turning radius of the N 3D target objects is determined as the current turning radius according to the recognition box length of the N 3D target objects in the last frame.

[0056] In the case of the disappearing 3D target object being a vehicle object, the driving parameters include the speed of the 3D target object at the last moment before disappearing and the vehicle acceleration, etc., and the current moving distance of the vehicle object can be determined according to the speed of the vehicle object at the last moment, the vehicle acceleration, and the time consumption of a frame of point cloud data. In terms of the length of the vehicle, the length of the recognition box of the 3D target object in the last frame is used to represent the length of the vehicle in the embodiment of the present application. Generally, the acceleration per 100 kilometers is usually higher than 2 seconds, and the calculation shows that the moving distance of the motor vehicle between two frames is about 1.3 meters more than the uniform speed, and the moving distance of the non-motor vehicle between two frames is about 0.5 meters more than the uniform speed. Therefore, according to the specific category of the vehicle object, the moving distance of the motor vehicle is (0.1v+1.3) meters, and the moving distance of the non-motor vehicle is (0.1v+0.5) meters, where v is the speed of the vehicle at the last moment.

[0057] In the case of any one of the N 3D target objects being a current 3D target object, a sector area corresponding to the current 3D target object in the current frame is constructed based on the current turning radius R and the current moving distance L of the current 3D target object, which is represented as follows: the following three expression functions in the world coordinate system, as shown in FIG. 8, the sector area surrounded by the following three expression functions is the sector area corresponding to the current 3D target object in the current frame: Figure 2

[0058] Expression of the first radius: (x+R) 2 +y 2 =R 2 ;

[0059] Expression of the second radius: (x-R) 2 +y 2 =R 2 ;

[0060] Expression of the arc:

[0061] where x and y are variables for expressing the sector area, R is the turning radius of the current 3D target object, and L is the current moving distance of the current 3D target object.

[0062] If the road traffic data is sufficient, a series of driving parameters such as the type of the vehicle, the speed of the vehicle at the last moment, the steering angle, and the distance to the front vehicle can be used as support to estimate the sector area with the cluster center in the second cluster center point set as the origin through a LightBGM model, so as to further reduce the sector area corresponding to the vehicle and improve the positioning accuracy.

[0063] ​The three expression functions of the current 3D target object in the corresponding fan-shaped region in the current frame are converted in the coordinate system based on the heading angle of the current 3D target object at the previous moment, to obtain the expression functions of the fan-shaped region of the current 3D target object in the user coordinate system.

[0064] In step S104, the cluster centers of the target class 3D objects are removed from the first cluster center point set to obtain M candidate cluster centers in the first cluster center point set, wherein the target class 3D objects are the 3D target objects that do not disappear at the current moment; the N 3D target objects are respectively taken as the current 3D target object, and the cluster centers in the fan-shaped region belonging to the current 3D target object are selected from the M candidate cluster centers; if there are K candidate cluster centers in the fan-shaped region of the current 3D target object, a target cluster center corresponding to the current 3D target object is determined from the K candidate cluster centers as the position information of the current 3D target object at the current moment, and K is a positive integer.

[0065] It should be understood that if there is only a single candidate cluster center in the fan-shaped region of the current 3D target object, i.e., K = 1, the single candidate cluster center existing in the fan-shaped region of the 3D target object is directly taken as the target cluster center. If there are multiple candidate cluster centers in the fan-shaped region of the current 3D target object, i.e., K is an integer greater than 1, an optimal matching is performed to determine a cluster center from the multiple candidate cluster centers existing in the fan-shaped region of the current 3D target object as the target cluster center.

[0066] Specifically, the process of optimal matching can be: first, according to the Gaussian distance as the reference, the candidate cluster centers in the fan-shaped region belonging to the current 3D target object are added to the elimination matrix, and the elimination matrix is processed by the Hungarian algorithm to optimally match the candidate cluster centers in the fan-shaped region of the current 3D target object with the cluster centers of the current 3D target object in the second cluster center point set, so as to match a target cluster center from the candidate cluster centers in the fan-shaped region as the position information of the current 3D target object at the current moment, thereby completing the repositioning of the disappeared current 3D target object, so as to continue tracking the current 3D target object based on the position information at the current moment.

[0067] Specifically, the process of optimal matching can be: first, according to the Gaussian distance as the reference, the candidate cluster centers in the fan-shaped region belonging to the current 3D target object are added to the elimination matrix, and the elimination matrix is processed by the Hungarian algorithm to optimally match the candidate cluster centers in the fan-shaped region of the current 3D target object with the cluster centers of the current 3D target object in the second cluster center point set, so as to match a target cluster center from the candidate cluster centers in the fan-shaped region as the position information of the current 3D target object at the current moment, thereby completing the repositioning of the disappeared current 3D target object, so as to continue tracking the current 3D target object based on the position information at the current moment.

[0068] It should be understood that in actual implementation, there can also be no candidate cluster center in the sector region corresponding to the current 3D target object due to too few point clouds. Then, if there is no candidate cluster center in the sector region of the current 3D target object, the clustering parameters used when generating the first cluster center point set need to be adjusted, such as reducing the distance threshold during clustering, and the point cloud data in the sector region in the current frame point cloud data is re-clustered based on the adjusted clustering parameters, and the position information of the current 3D target object at the current time is determined according to at least one new cluster center generated by re-clustering.

[0069] Specifically, if there is only one new cluster center in the sector region corresponding to the current 3D target object after re-clustering, the new cluster center is taken as the position information of the current 3D target object at the current time; if there are multiple new cluster centers in the sector region corresponding to the current 3D target object after re-clustering, a new cluster center is matched from each new candidate cluster center as a target cluster center through optimal matching, and the specific optimal matching process is described above and will not be repeated here.

[0070] The above technical solution finds the cluster center corresponding to the disappeared 3D target object in the current frame through clustering as the position information of the disappeared 3D target object at the current time, thereby improving the accuracy of repositioning the disappeared 3D target object.

[0071] Based on the same inventive concept, the embodiments of the present application provide a 3D target object positioning device, as shown in Figure 3 The device includes: a clustering unit 301 configured to generate a first cluster center point set according to current frame point cloud data of a preset road region, and generate a second cluster center point set according to last frame point cloud data of the preset road region, the current frame point cloud data has N 3D target objects disappearing relative to the last frame point cloud data, N is a positive integer; a center matching unit 302 configured to determine N cluster centers corresponding to the N 3D target objects in the second cluster center point set according to the first cluster center point set and the second cluster center point set; a first region prediction unit 303 configured to establish N sector regions corresponding to the N 3D target objects in the current frame according to the N cluster centers, current moving distances of the N 3D target objects, and current turning radii; and a position determination unit 304 configured to determine N target cluster centers in the N sector regions as position information of the N 3D target objects at the current time according to the current frame point cloud data.

[0072] In some embodiments, the clustering unit 301 comprises: a background filtering subunit configured to filter background point cloud information from the current frame point cloud data and the previous frame point cloud data to obtain first remaining point cloud data of the current frame point cloud data and second remaining point cloud data of the previous frame point cloud data; a first clustering subunit configured to cluster the first remaining point cloud data to obtain the first set of clustering center points; and a second clustering subunit configured to cluster the second remaining point cloud data to obtain the second set of clustering center points.

[0073] In some embodiments, the background filtering subunit is specifically configured to: perform difference calculation on the background mask of the preset road region and the current frame point cloud data to filter background point cloud information in the current frame point cloud data; and perform difference calculation on the background mask of the preset road region and the previous frame point cloud data to filter background point cloud information in the previous frame point cloud data.

[0074] In some embodiments, the center matching unit 302 is specifically configured to: remove a clustering center in which a target class 3D object is located from the second set of clustering center points, and remove a clustering center in which a noise point is located from the second set of clustering center points, to obtain N clustering centers corresponding to the N 3D target objects in the second set of clustering center points; wherein the target class 3D object is a 3D target object that does not disappear at the current time.

[0075] In some embodiments, the method further comprises a second region prediction unit configured to: determine a circular region with a clustering center of the pedestrian object in the second set of clustering center points as a center; wherein a size parameter of the circular region is determined in advance based on a walking speed of the pedestrian object.

[0076] In some embodiments, the method further comprises: a parameter determination unit configured to determine driving parameters and bounding box lengths of the N 3D target objects at a previous time by tracking trajectories of the N 3D target objects; determine a movement distance generated by the N 3D target objects within a time consumption of collecting a frame of point cloud data as the current movement distance according to the driving parameters of the N 3D target objects at the previous time; and determine a turning radius of the N 3D target objects as the current turning radius according to the bounding box lengths of the N 3D target objects at the previous time.

[0077] In some embodiments, the position determining unit comprises: a center removing subunit configured to remove a cluster center in which the target class 3D object is located from the first cluster center point set to obtain M candidate cluster centers in the first cluster center point set, M being an integer greater than 1; a center screening subunit configured to screen, as a current 3D target object, a cluster center belonging to a sector region of the current 3D target object from the M candidate cluster centers; a first position determining subunit configured to determine, as position information of the current 3D target object at a current time, a target cluster center corresponding to the current 3D target object from K candidate cluster centers in the sector region of the current 3D target object, K being a positive integer; and a second position determining subunit configured to, if there is no candidate cluster center in the sector region of the current 3D target object, re-cluster point cloud data in the sector region in the current frame point cloud data based on the adjusted clustering parameters, and determine position information of the current 3D target object at the current time according to at least one new cluster center generated by the re-clustering.

[0078] In some embodiments, the first position determining subunit is specifically configured to: if there are multiple candidate cluster centers in the sector region of the current 3D target object, determine, as the target cluster center corresponding to the current 3D target object, a target cluster center from the multiple candidate cluster centers by optimal matching; and if there is a single candidate cluster center in the sector region of the current 3D target object, determine, as the target cluster center corresponding to the current 3D target object, the single candidate cluster center.

[0079] In some embodiments, the method further comprises a mask forming unit configured to: in a state in which the road region is free of 3D target objects, acquire a plurality of frames of historical point cloud data of a position at which the road region is located; fuse the plurality of frames of historical point cloud data to obtain fused point cloud data; and process the fused point cloud data to form a voxelized mask of the road region and save the voxelized mask.

[0080] As to the above apparatus, the specific functions of each module have been described in detail in the 3D target object positioning method embodiments provided in the present specification, and thus will not be described in detail here.

[0081] Based on the same inventive concept, the embodiments of the present application provide an electronic device, as shown in Figure 4 The electronic device comprises one or more memories 404, one or more processors 402, and at least one computer program (program code) stored in the memory 404 and executable on the processor 402, and the processor 402 implements the 3D target object positioning method as described above when executing the computer program.

[0082] wherein, in Figure 4 The bus architecture, represented by bus 400, can include any number of interconnected buses and bridges, the bus 400 linking together various circuits including the processor 402 represented by one or more processors and the memory 404 represented by memory. The bus 400 can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art, and therefore, will not be further described herein. The bus interface 405 provides an interface between the bus 400 and the receiver 401 and the transmitter 403. The receiver 401 and the transmitter 403 can be the same element, i.e., a transceiver, providing a means for communicating with various other apparatus over a transmission medium. The processor 402 is responsible for managing the bus 400 and general processing, while the memory 404 can be used for storing data used by the processor 402 in executing operations.

[0083] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to encompass within their scope all such variations and modifications as are included within the scope of the invention.

[0084] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A 3D target object positioning method, characterized by, The method comprises the following steps: generating a first clustering center point set according to current frame point cloud data of a preset road region, and generating a second clustering center point set according to last frame point cloud data of the preset road region, wherein the current frame point cloud data has N 3D target objects missing relative to the last frame point cloud data, and N is a positive integer; determining N clustering centers corresponding to the N 3D target objects in the second clustering center point set according to the first clustering center point set and the second clustering center point set; determining driving parameters and identification box lengths of the N 3D target objects at a last time through trajectory tracking of the N 3D target objects; determining a moving distance of the N 3D target objects within a time consumption of collecting a frame of point cloud data as a current moving distance according to the driving parameters of the N 3D target objects at the last time, and determining a turning radius of the N 3D target objects as a current turning radius according to the identification box lengths of the N 3D target objects at the last time; establishing N sector regions corresponding to the N 3D target objects in the current frame according to the N clustering centers, the current moving distance and the current turning radius of the N 3D target objects; determining N target clustering centers in the N sector regions as position information of the N 3D target objects at a current time according to the current frame point cloud data, comprising: removing a clustering center in which a target class 3D object is located from the first clustering center point set to obtain M candidate clustering centers in the first clustering center point set, wherein M is an integer greater than 1; taking the N 3D target objects as current 3D target objects respectively, and screening a clustering center in a sector region belonging to the current 3D target object from the M candidate clustering centers; if there are K candidate clustering centers in the sector region of the current 3D target object, determining a target clustering center corresponding to the current 3D target object from the K candidate clustering centers as position information of the current 3D target object at the current time, wherein K is a positive integer; if there is no candidate clustering center in the sector region of the current 3D target object, re-clustering point cloud data in the sector region in the current frame point cloud data based on adjusted clustering parameters, and determining position information of the current 3D target object at the current time according to at least one new clustering center generated by re-clustering, wherein the target class 3D object is a 3D target object that does not disappear at the current time.

2. The method of claim 1, wherein, The method comprises the following steps: filtering background point cloud information from the current frame point cloud data and the last frame point cloud data to obtain first residual point cloud data of the current frame point cloud data and second residual point cloud data of the last frame point cloud data; clustering the first residual point cloud data to obtain the first clustering center point set; clustering the second residual point cloud data to obtain the second clustering center point set.

3. The method of claim 2, wherein, The filtering background point cloud information from the current frame point cloud data and the last frame point cloud data comprises: differential calculation of the background mask of the preset road region and the current frame point cloud data to filter background point cloud information in the current frame point cloud data; differential calculation of the background mask of the preset road region and the last frame point cloud data to filter background point cloud information in the last frame point cloud data.

4. The method of claim 1, wherein, The determination of the N cluster centers corresponding to the N 3D target objects in the second cluster center point set according to the first cluster center point set and the second cluster center point set comprises: removing the cluster center where the target class 3D object is located from the second cluster center point set, and removing the cluster center where the noise point is located in the second cluster center point set, to obtain the N cluster centers corresponding to the N 3D target objects in the second cluster center point set.

5. The method of claim 1, wherein, If there is at least one pedestrian object in the N 3D target objects, after the determination of the N cluster centers corresponding to the N 3D target objects in the second cluster center point set according to the first cluster center point set and the second cluster center point set, it further comprises: determining a circular region with the cluster center of the pedestrian object in the second cluster center point set as the center; wherein the size parameter of the circular region is determined in advance based on the walking speed of the pedestrian object.

6. The method of claim 1, wherein, Determining the target cluster center corresponding to the current 3D target object from the K candidate cluster centers comprises: if there are multiple candidate cluster centers in the sector region of the current 3D target object, determining the target cluster center corresponding to the current 3D target object from the multiple candidate cluster centers by optimal matching; if there is a single candidate cluster center in the sector region of the current 3D target object, determining the single candidate cluster center as the target cluster center corresponding to the current 3D target object.

7. A positioning apparatus of a 3D target object, characterized by, It comprises: a clustering unit configured to generate a first cluster center point set according to current frame point cloud data of a preset road region, and generate a second cluster center point set according to last frame point cloud data of the preset road region, the current frame point cloud data having N 3D target objects disappeared relative to the last frame point cloud data, N being a positive integer; a center matching unit configured to determine N cluster centers corresponding to the N 3D target objects in the second cluster center point set according to the first cluster center point set and the second cluster center point set; a parameter determination unit configured to determine driving parameters and recognition box length of the N 3D target objects at a last time through trajectory tracking of the N 3D target objects; determine the moving distance of the N 3D target objects within the time consumption of collecting a frame of point cloud data as the current moving distance according to the driving parameters of the N 3D target objects at the last time, and determine the turning radius of the N 3D target objects as the current turning radius according to the recognition box length of the N 3D target objects at the last time; The first region prediction unit is configured to establish N fan-shaped regions corresponding to the N 3D target objects in the current frame according to the N cluster centers, current moving distances of the N 3D target objects, and current turning radii. The position determination unit is configured to determine N target cluster centers in the N fan-shaped regions as position information of the N 3D target objects at the current time according to the current frame point cloud data, including: removing a cluster center in which a target class 3D object is located from a first cluster center point set to obtain M candidate cluster centers in the first cluster center point set, M being an integer greater than 1; taking the N 3D target objects as current 3D target objects respectively, and screening cluster centers belonging to the fan-shaped region of the current 3D target object from the M candidate cluster centers; if there are K candidate cluster centers in the fan-shaped region of the current 3D target object, determining a target cluster center corresponding to the current 3D target object from the K candidate cluster centers as position information of the current 3D target object at the current time, K being a positive integer; if there is no candidate cluster center in the fan-shaped region of the current 3D target object, re-clustering point cloud data in the fan-shaped region in the current frame point cloud data based on adjusted clustering parameters, and determining position information of the current 3D target object at the current time according to at least one new cluster center generated by re-clustering, wherein the target class 3D object is a 3D target object that does not disappear at the current time.

8. An electronic device, comprising a memory, a processor, and code stored on the memory and executable on the processor, the code comprising instructions for: The processor implements the method in any one of claims 1-6 when executing the code. The processor implements the method in any one of claims 1-6 when executing the code.

Images