Laser positioning method and system for automatic driving

The lightweight laser positioning method for autonomous driving preprocesses laser point clouds and semantic maps to enhance efficiency and accuracy, addressing the time-consuming issues of existing SLAM technologies.

CN119087446APending Publication Date: 2024-12-06SHANGHAI CANTRON ROBOTICS TECH LTD CO
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411188427.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

Existing laser SLAM technologies for autonomous driving are time-consuming due to extensive point cloud registration calculations, especially on computing platforms with limited resources, affecting real-time performance.

Method used

A lightweight laser positioning method involving preprocessing of multi-frame laser point clouds and semantic maps to obtain BEV and semantic point clouds, using multi-angle hypotheses and multi-resolution registration to determine the vehicle's initial and real-time posture.

Benefits of technology

Reduces calculation time by about 30% and memory usage by about 60%, improving algorithm performance and enabling real-time positioning on limited computing platforms.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119087446A_ABST
    Figure CN119087446A_ABST
Patent Text Reader

Abstract

The invention provides a laser positioning method and system for automatic driving. The laser positioning method comprises the steps of obtaining a prior position of a vehicle; obtaining a multi-frame laser point cloud and a semantic map of the driving environment of the vehicle; preprocessing the multiple frames of laser point clouds and the semantic map to respectively obtain multiple frames of BEV point clouds and semantic point clouds; performing initial positioning on the vehicle based on the prior position, an initial frame BEV point cloud in the multiple frames of BEV point clouds and the semantic point cloud to determine an initial pose of the vehicle; obtaining the pose increment of the vehicle in the driving process of the vehicle; and determining the real-time pose of the vehicle in the driving process based on the initial pose and the pose increment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application generally relates to the field of autonomous driving, and more particularly to a laser positioning method and system for autonomous driving. Background Art

[0002] In the field of autonomous driving and robotics, in order for a vehicle or robot to automatically perform tasks, it is first necessary to let the vehicle or robot know its own posture information in the environment for use by downstream modules (for example, path planning modules, etc.). SLAM (Simultaneous Localization and Mapping) is a key technology for achieving vehicle or robot posture estimation. In the field of SLAM, the more popular research directions include multi-sensor fusion laser SLAM and visual SLAM. Benefiting from the excellent active ranging and scanning capabilities of lidar for three-dimensional environments, laser SLAM is currently more widely used in L4 autonomous driving or robotics applications, while visual SLAM is more used for L2 assisted driving.

[0003] The existing laser SLAM technology obtains the initial pose through GPS (Global Positioning System) / RTK (Real Time Kinematic) during positioning, and then performs 3D NDT (Normal Distribution Transform) registration on the laser point cloud of the current frame and the point cloud in the high-precision map to obtain an accurate initial pose. In the subsequent positioning process, the pose increment is applied to the pose of the previous moment, and the result is used as the pose prediction of the next moment, and then the point cloud registration is continued to optimize the pose. However, when the map data is large, the 3D registration of the point cloud is quite time-consuming. This is mainly because the registration algorithm uses too many points in the source point cloud (current frame point cloud) and the target point cloud (local point cloud in the high-precision map). In the NDT registration algorithm, the target point cloud needs to be divided into grid areas first, and then the predicted pose is used to transform the point cloud into a coordinate system to obtain a predicted point cloud (predicted_cloud). The grid area coordinates are calculated for each point in the cloud, and then the residual is constructed for nonlinear optimization. The residual is the distance between the point in the predicted point cloud and the centroid of the corresponding grid area. The result of the optimization is used as the final pose. On a computing platform with limited resources, when the scale of the points is large, the above series of calculations is very time-consuming, which will affect the real-time performance of the algorithm and even make it unusable.

[0004] In view of this, the present application proposes a lightweight laser positioning method and system for autonomous driving to solve the above-mentioned defects of the prior art. Summary of the invention

[0005] A brief summary of one or more aspects is given below to provide a basic understanding of these aspects. This summary is not an exhaustive overview of all conceived aspects, and is neither intended to identify the key or critical elements of all aspects nor to define the scope of any or all aspects. Its only purpose is to give some concepts of one or more aspects in a simplified form as a prelude to a more detailed description that will be given later.

[0006] The present application provides a laser positioning method for autonomous driving, comprising: obtaining a priori position of a vehicle; obtaining a multi-frame laser point cloud and a semantic map of the driving environment of the vehicle; preprocessing the multi-frame laser point cloud and the semantic map to obtain a multi-frame BEV point cloud and a semantic point cloud, respectively; initializing the positioning of the vehicle based on the priori position, an initial frame BEV point cloud in the multi-frame BEV point cloud, and the semantic point cloud to determine an initial posture of the vehicle; obtaining a posture increment of the vehicle during driving of the vehicle; and determining a real-time posture of the vehicle during driving based on the initial posture and the posture increment.

[0007] In some embodiments, the a priori position is obtained by at least one of a GPS module and an RTK module of the vehicle.

[0008] In some embodiments, the multi-frame laser point cloud is obtained by the vehicle's laser radar scanning the vehicle's driving environment, and the semantic map is stored in the vehicle or downloaded from a server.

[0009] In some embodiments, preprocessing the multi-frame laser point cloud to obtain a multi-frame BEV point cloud further includes: filtering out points in the multi-frame laser point cloud whose heights are not within a preset interval; projecting the remaining points to a plane of z=0; downsampling the projected points using a voxel downsampling filter to obtain a downsampled point cloud; and performing boundary extraction on the downsampled point cloud to obtain the multi-frame BEV point cloud.

[0010] In some embodiments, the semantic map includes a plurality of vector elements, and preprocessing the semantic map to obtain a semantic point cloud further includes: densely upsampling each vector element in the semantic map to obtain the semantic point cloud.

[0011] In some embodiments, initializing the positioning of the vehicle based on the prior position, the initial frame BEV point cloud and the semantic point cloud to determine the initial posture of the vehicle further includes: determining a set of candidate postures based on the prior position; applying each candidate posture in the candidate posture set to the initial frame BEV point cloud to obtain multiple initial predicted point clouds; using a point cloud registration algorithm to register each initial predicted point cloud and the semantic point cloud to obtain multiple registered postures and corresponding registration scores; and selecting the registered posture with the highest registration score as the initial posture of the vehicle.

[0012] In some embodiments, determining the candidate pose set based on the prior position further includes: obtaining a plurality of candidate pose angles with the prior position as the center; determining the candidate pose set T based on the prior position and the plurality of candidate pose angles all ={T0,T1,…,T N-1}, N represents the number of candidate pose angles, where T = [RP], P is the prior position, R is the rotation matrix and is obtained by Rodriguez formula:

[0013] R=I+(1-cosθ)*n ^ *n ^ +sinθ*n ^

[0014] where θ = ‖[α β γ] T ‖, α represents the pitch angle, β represents the roll angle, γ represents the yaw angle, n is the normalized rotation vector, n = [α β γ] T / θ=[n1 n2 n3] T , n^ represents the antisymmetric matrix of n.

[0015] In some embodiments, the pose increment is obtained by:

[0016] ΔT DR =T DR_k0 -1 *T DR_k1 ,

[0017] Where T DR_k0 represents the initial position determined by the vehicle’s dead reckoning (DR) module, T DR_k1 represents the pose determined by the DR module at the current moment, ΔT DR Indicates the position increment of the current moment relative to the initial moment.

[0018] In some embodiments, determining the real-time posture of the vehicle during driving based on the initial posture and the posture increment further includes: determining the real-time predicted posture based on the initial posture and the posture increment; applying the real-time predicted posture to the BEV point cloud of the current frame to obtain the real-time predicted point cloud; and using a point cloud registration algorithm to align the real-time predicted point cloud and the semantic point cloud to obtain the registered posture as the real-time posture of the vehicle.

[0019] The present application also provides a laser positioning system for autonomous driving, including: a data acquisition module, which is configured to: obtain a priori position of a vehicle; obtain a multi-frame laser point cloud and a semantic map of the vehicle's driving environment; a preprocessing module, which preprocesses the multi-frame laser point cloud and the semantic map to obtain a multi-frame BEV point cloud and a semantic point cloud, respectively; an initialization positioning module, which initializes the positioning of the vehicle based on the priori position, the initial frame BEV point cloud in the multi-frame BEV point cloud, and the semantic point cloud to determine the initial posture of the vehicle; a posture increment determination module, which obtains the posture increment of the vehicle during the driving process of the vehicle; and a real-time positioning module, which determines the real-time posture of the vehicle during driving based on the initial posture and the posture increment.

[0020] In some embodiments, the a priori position is obtained by at least one of a GPS module and an RTK module of the vehicle.

[0021] In some embodiments, the multi-frame laser point cloud is obtained by a laser radar of the vehicle scanning the driving environment of the vehicle, and wherein the semantic map is stored in the vehicle or downloaded from a server.

[0022] In some embodiments, the preprocessing module is further configured to preprocess the multi-frame laser point cloud by the following operations: filtering out points in the multi-frame laser point cloud whose heights are not within a preset interval; projecting the remaining points to the plane of z=0; downsampling the projected points using voxel downsampling filtering to obtain a downsampled point cloud; and performing boundary extraction on the downsampled point cloud to obtain the multi-frame BEV point cloud.

[0023] In some embodiments, the semantic map includes a plurality of vector elements, and the pre-processing module is further configured to pre-process the semantic map by densely upsampling each vector element in the semantic map to obtain a semantic point cloud.

[0024] In some embodiments, the initialization positioning module is further configured to: determine a set of candidate poses based on the prior position; apply each candidate pose in the candidate pose set to the initial frame BEV point cloud to obtain multiple initial predicted point clouds; use a point cloud registration algorithm to align each initial predicted point cloud and the semantic point cloud to obtain multiple registered poses and corresponding registration scores; and select the registered pose with the highest registration score as the initial pose of the vehicle.

[0025] In some embodiments, the initialization positioning module is further configured to determine the candidate pose set by the following operations: obtaining a plurality of candidate pose angles with the prior position as the center; determining the candidate pose set T based on the prior position and the plurality of candidate pose angles all ={T0,T1,…,T N-1}, N represents the number of candidate pose angles, where T = [RP], P is the prior position, R is the rotation matrix and is obtained by Rodriguez formula:

[0026] R=I+(1-cosθ)*n ^ *n ^ +sinθ*n ^

[0027] where θ = ‖[α β γ] T ‖, α represents the pitch angle, β represents the roll angle, γ represents the yaw angle, n is the normalized rotation vector, n = [α β γ] T / θ=[n1 n2 n3] T , n^ represents the antisymmetric matrix of n.

[0028] In some embodiments, the pose increment determination module is further configured to obtain the pose increment by the following formula:

[0029] ΔT DR =T DR_k0 -1 *T DR_k1 ,

[0030] Where T DR_k0 represents the initial position determined by the vehicle’s dead reckoning (DR) module, T DR_k1 represents the pose determined by the DR module at the current moment, ΔT DR Indicates the position increment of the current moment relative to the initial moment.

[0031] In some embodiments, the real-time positioning module is further configured to: determine the real-time predicted pose based on the initial pose and the pose increment; apply the real-time predicted pose to the BEV point cloud of the current frame to obtain the real-time predicted point cloud; and use a point cloud registration algorithm to align the real-time predicted point cloud and the semantic point cloud to obtain the registered pose as the real-time pose of the vehicle.

[0032] The present application also proposes a device, including a processor and a memory, wherein the memory stores program instructions, and the processor runs the program instructions to implement the aforementioned laser positioning method for autonomous driving.

[0033] The technical solution of the present application reduces the dimension of the point cloud by performing BEV preprocessing on the laser point cloud, reduces the amount of calculation and storage space, and thus improves the efficiency of the algorithm. At the same time, multi-angle assumptions are used in the initial positioning to ensure positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The features, nature and advantages of the present application will become more apparent when the detailed description set forth below is understood in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals are always used for corresponding identification. It should be noted that the drawings described are only schematic and non-limiting. In the accompanying drawings, the sizes of some components may be exaggerated and are not drawn to scale for illustrative purposes.

[0035] Figure 1 An algorithm architecture diagram for laser positioning for autonomous driving is shown.

[0036] Figure 2 A laser positioning method for autonomous driving is shown.

[0037] Figure 3 An exemplary process for initializing positioning of a vehicle is shown.

[0038] Figure 4 A schematic diagram of obtaining candidate attitude angles is shown.

[0039] Figure 5 A laser positioning system for autonomous driving is shown.

[0040] Figure 6 A block diagram of a device including a laser positioning system for autonomous driving is shown. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. In the following detailed description, many specific details are set forth to provide a thorough understanding of the described exemplary embodiments. However, it is obvious to those skilled in the art that the described embodiments can be practiced without some or all of these specific details. In other exemplary embodiments, well-known structures are not described in detail to avoid unnecessarily obscuring the concepts of the present application. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. At the same time, in the absence of conflict, the various aspects described in the embodiments can be combined arbitrarily.

[0042] As mentioned above, existing laser positioning technology is quite time-consuming when performing a series of positioning calculations. When computing resources are limited, it will affect real-time performance and even make it unusable.

[0043] In order to achieve efficient laser positioning on a limited computing resource platform, the present invention proposes a lightweight technical solution.

[0044] Figure 1 An algorithm architecture diagram 100 for laser positioning for autonomous driving is shown.

[0045] like Figure 1 As shown, the laser positioning algorithm of the present invention mainly involves multiple modules related to information / data acquisition: GPS / RTK module, 3D laser radar module, DR (Dead Reckoning) module.

[0046] The GPS / RTK module is used to provide the vehicle's prior position observation information.

[0047] The laser radar module is used to scan the driving environment around the vehicle to obtain a laser point cloud.

[0048] The DR module includes a combination of a wheel speed meter and an IMU (Inertial measurement unit) to provide position increments.

[0049] In addition, the semantic map may be loaded from the vehicle locally (eg, an onboard storage device) or downloaded from a server.

[0050] After obtaining the above data, the data can be preprocessed to a certain extent.

[0051] For example, for a laser point cloud obtained by a laser radar, the laser point cloud may be preprocessed to obtain a BEV (Bird's Eye View) point cloud.

[0052] For the obtained semantic map, the semantic map can be preprocessed to obtain a semantic point cloud.

[0053] After the above preprocessing, the vehicle can be initialized and positioned based on the prior position, BEV point cloud and semantic point cloud to determine the initial posture of the vehicle.

[0054] In the present invention, the vehicle's posture includes the vehicle's position and attitude. The vehicle's position may include the vehicle's absolute position (e.g., expressed as an absolute coordinate value) or relative position (e.g., expressed as a direction and distance relative to a known position), while the vehicle's attitude may be expressed by attitude angles (yaw angle γ, pitch angle α, and roll angle β).

[0055] After the initial positioning is completed, the vehicle's posture increment can be determined by the DR module during the vehicle's driving process. Therefore, the vehicle can be tracked and positioned based on the initial posture and posture increment during the vehicle's driving process to determine the vehicle's real-time posture. The determined real-time posture can be used in downstream modules (e.g., path planning module) to assist the vehicle's automatic driving.

[0056] Figure 2 A laser positioning method 200 for autonomous driving is shown.

[0057] like Figure 2 As shown, the method 200 starts at step 205. At step 205, the prior position of the vehicle is obtained.

[0058] Before the vehicle starts to travel, the initial position of the vehicle may be obtained first.

[0059] In some embodiments of the present invention, the prior position of the vehicle can be obtained by at least one of the GPS module and the RTK module of the vehicle. GPS and RTK positioning are well known in the art and will not be described in detail here.

[0060] It should be noted that the a priori position of the vehicle obtained by the GPS / RTK module may have a certain error. Therefore, the a priori position obtained in step 205 is a rough initial position of the vehicle.

[0061] In step 210 , a multi-frame laser point cloud and a semantic map of the vehicle's driving environment are obtained.

[0062] In various embodiments of the present invention, the laser point cloud is obtained by using a laser radar of the vehicle to scan the driving environment of the vehicle.

[0063] The driving environment may be a surrounding area related to vehicle driving. For example, the driving environment may be a specific surrounding spatial area centered on the vehicle. In other examples, the driving environment may be a specific spatial area around the vehicle that is not centered on the vehicle. For example, if the vehicle is ready to move forward, the driving environment may be a specific spatial area around the vehicle and most of the area is in front of the vehicle body. In a specific implementation, those skilled in the art may set the driving environment of the vehicle as needed.

[0064] The laser radar obtains laser point cloud data of the surrounding environment by scanning the surrounding environment. It is not affected by environmental conditions such as lighting and has high distance accuracy. When the vehicle is at the initial position and has not yet started driving, the initial frame laser point cloud can be obtained by scanning the driving environment around the initial position of the vehicle. Subsequently, during the driving process of the vehicle, the subsequent frames of laser point clouds can be obtained by continuously scanning the driving environment around the vehicle. Thus, the initial frame laser point cloud in the multi-frame laser point cloud obtained in step 210 can be the laser point cloud obtained when the vehicle is at the initial position, and the subsequent frames of laser point clouds in the multi-frame laser point cloud can be the laser point clouds obtained during the driving process after the vehicle leaves the initial position.

[0065] In some implementations, the semantic map may be stored in the vehicle (eg, an onboard storage device) in advance and loaded from the storage device when in use. In other implementations, the semantic map may be downloaded from a server in real time.

[0066] In step 215 , the multi-frame laser point cloud and the semantic map are preprocessed to obtain multi-frame BEV point cloud and semantic point cloud, respectively.

[0067] In some embodiments of the present invention, preprocessing a multi-frame laser point cloud to obtain a multi-frame BEV point cloud further includes: filtering out points in the multi-frame laser point cloud whose heights are not within a preset interval; projecting the remaining points to a plane of z=0; downsampling the projected points using a voxel downsampling filter to obtain a downsampled point cloud; and performing boundary extraction on the downsampled point cloud to obtain a multi-frame BEV point cloud.

[0068] Specifically, assuming that the installation horizontal angle of the laser radar on the vehicle body is parallel to the ground or the angle of the laser point cloud has been calibrated, considering the laser point cloud of the current frame, the points whose heights are not in the interval of 0.8m-2.5m in the laser point cloud can be first screened out, thereby eliminating the ground point cloud and the high-altitude obstacle point cloud. Subsequently, the remaining points in the laser point cloud can be projected to the plane of z=0. After that, the projected point cloud is downsampled using voxel downsampling filtering to obtain a downsampled point cloud (also referred to as "downsampled point cloud" in this article). By downsampling, the number of point clouds can be reduced, thereby reducing subsequent computational consumption. Next, a boundary extraction algorithm is used to extract a more accurate obstacle outline from the downsampled point cloud, which is the BEV point cloud. The voxel downsampling filtering and boundary extraction in the above process can use the methods in the PCL (Point Cloud Library) point cloud library, which will not be repeated here.

[0069] In some embodiments of the present invention, after the laser radar obtains the laser point cloud by scanning, the overlap between the laser point cloud and the local area in the existing global map can also be determined. If the overlap is high (for example, exceeding a preset threshold), the laser point cloud is valid and can be subsequently processed. If the overlap is low (for example, not exceeding a preset threshold), it is necessary to re-scan and obtain a new laser point cloud.

[0070] For each frame of the multi-frame laser point cloud, the above preprocessing can be performed to obtain a multi-frame BEV point cloud. By preprocessing the laser point cloud, the point cloud can be reduced in dimension, thereby reducing the amount of calculation and storage space, and improving the point cloud processing efficiency.

[0071] It should be noted that although a specific preset interval of 0.8 m-2.5 m in height is shown above, this is only exemplary and not restrictive. In specific implementations, those skilled in the art may select different preset intervals according to actual conditions.

[0072] In some embodiments of the present invention, preprocessing the semantic map to obtain a semantic point cloud further includes: densely upsampling each vector element in the semantic map to obtain a semantic point cloud.

[0073] A semantic map is a 2D layer that includes multiple vector elements, which contain multiple sequential vector points. The combination of these vector points represents semantic features. For example, the four corners of a pillar in an indoor parking lot can correspond to four vector points a, b, c, and d. The combination of these four vector points represents an obstacle element (i.e., a pillar).

[0074] In subsequent point cloud processing (e.g., point cloud registration), data in point cloud format is required. Since the vector points in the vector elements of the semantic map are too discrete and sparse to be directly used for point cloud registration, each vector element needs to be densely upsampled to obtain a semantic point cloud that can be used for registration.

[0075] When performing dense upsampling, assuming that a vector element contains two vector points a and b, the distance between the two vector points dist = |ab| is first calculated. As an example, if the interpolation step size d = 0.05m, the number of points to be interpolated n = dist / d. For the i-th point to be interpolated, the scale factor t = i / n is calculated, and then the interpolation point is calculated:

[0076] [(1-t)*a x +t*b x (1-t)*a y +t*b y 0] T .

[0077] By traversing all obstacle semantic vector elements in the semantic map in this way, a semantic point cloud corresponding to the entire semantic map can be obtained.

[0078] In step 220 , the vehicle is initially positioned based on the prior position, the initial frame BEV point cloud in the multi-frame BEV point cloud, and the semantic point cloud to determine the initial posture of the vehicle.

[0079] After obtaining the vehicle's prior position, multi-frame BEV point cloud, and semantic point cloud in the above steps, the vehicle can be initialized and positioned accordingly. The general process of initial positioning is as follows: determine a set of candidate poses based on the prior position, apply each candidate pose in the candidate pose set to the initial frame BEV point cloud to obtain multiple initial predicted point clouds, use a point cloud registration algorithm to register each initial predicted point cloud and semantic point cloud to obtain multiple registered poses and corresponding registration scores, and select the registered pose with the highest registration score as the vehicle's initial pose. The detailed process of initial positioning will be described in detail below. Figure 3 Provide further explanation.

[0080] In step 225, the position increment of the vehicle is obtained during the driving process of the vehicle.

[0081] During the driving process of the vehicle, the vehicle's DR module can be used to obtain the vehicle's posture increment. Specifically, the posture increment can be obtained by the following formula:

[0082] ΔT DR =T DR_k0 -1 *T DR_k1 ,

[0083] Where T DR_k0 represents the pose determined by the DR module of the vehicle at the initial time (k0), T DR_k1 represents the pose determined by the DR module at the current time (k1), ΔT DR Indicates the position increment of the current moment relative to the initial moment.

[0084] In actual implementation, if the posture increment of the previous moment relative to the initial moment is already known, the posture increment of the current moment relative to the initial moment can also be obtained by determining the posture increment of the current moment relative to the previous moment.

[0085] In step 230 , the real-time posture of the vehicle during driving is determined based on the initial posture and the posture increment.

[0086] After obtaining the initial position and position increment of the vehicle during driving, the vehicle can be tracked and positioned in real time to determine the real-time position of the vehicle. Real-time tracking and positioning mainly includes two parts: position prediction and frame registration.

[0087] Specifically, assuming that the initial position of the vehicle at the initial time (t0) is T t0 (ie, the initial posture determined in step 220). At the current time t1, the posture increment obtained by the DR module of the vehicle is ΔT DR =T DR_t0 -1 *T DR_t1 . Therefore, the current position can be predicted as: T t1_predict =T t0 *ΔT DR .

[0088] In some other embodiments, if the vehicle's position T at the previous moment is known previous , then we can also use T previous And the current moment relative to the previous moment posture increment ΔT DR ′ to predict the pose T at the current time t1 t1_predict =T previous *ΔT DR ′.

[0089] After predicting the current pose, frame registration can be performed. Specifically, the predicted pose T t1_predict BEV point cloud C applied to the current frame origin , in order to C origin Convert to the predicted coordinate system to get C predict , and then use a point cloud registration algorithm (e.g., NDT algorithm) to align C predict The real-time position T of the vehicle at the current moment is obtained by registering with the semantic point cloud. t1 .

[0090] Thus, determining the real-time posture of the vehicle during driving based on the initial posture and the posture increment may include: determining the real-time predicted posture based on the initial posture and the posture increment; applying the real-time predicted posture to the BEV point cloud of the current frame to obtain the real-time predicted point cloud; and registering the real-time predicted point cloud and the semantic point cloud using a point cloud registration algorithm to obtain the registered posture as the real-time posture of the vehicle. In this way, the real-time posture of the vehicle at the current moment can be determined.

[0091] After obtaining the real-time position and posture of the vehicle, the real-time position and posture can be used to perform related autonomous driving tasks. For example, the real-time position and posture of the vehicle can be provided to a path planning module of the vehicle system for path planning.

[0092] Figure 3 An exemplary process 300 for initially positioning a vehicle is shown.

[0093] like Figure 3As shown, the process 300 begins with obtaining a priori position (305). As described above, the a priori position of the vehicle can be obtained through the GPS / RTK module of the vehicle.

[0094] Next, a plurality of candidate posture angles may be obtained with the priori position as the center of the circle (310).

[0095] Combination Figure 4 400 is a schematic diagram of obtaining candidate attitude angles. Figure 4 As shown, taking the prior position P as the center of the circle, the area of ​​0 to 360 degrees around the center of the circle is divided into multiple parts, thereby obtaining multiple candidate posture angles.

[0096] In a preferred embodiment, 0 to 360 degrees may be evenly divided into multiple parts. For example, 0 to 360 degrees may be evenly divided into 36 parts, thereby obtaining 36 candidate attitude angles. In the case of evenly dividing into 36 parts, Figure 4 The angle ω in the equation is 10 degrees. For simplicity, Figure 4 Only the first two angles obtained by evenly dividing 0 to 360 degrees are shown, while the remaining divided angles are omitted.

[0097] return Figure 3 After obtaining multiple candidate posture angles, a candidate posture set (315) can be determined based on the prior position and the multiple candidate posture angles.

[0098] Specifically, the candidate pose set can be expressed as T all ={T0,T1,…,T N-1}, where N represents the number of candidate pose angles, T = [RP], P is the prior position, and R is the rotation matrix and is obtained by Rodriguez formula:

[0099] R=I+(1-cosθ)*n^*n^+sinθ*n^

[0100] Where I is the identity matrix, θ = ‖[α β γ] T ‖, α represents the pitch angle, β represents the roll angle, γ represents the yaw angle, n is the normalized rotation vector, n = [α β γ] T / θ=[n1 n2 n3] T , n ^ Represents the antisymmetric matrix of n:

[0101]

[0102] When the vehicle body is parallel to the ground, the pitch angle α and the roll angle β are both equal to 0. Therefore, the vehicle's attitude angle can be represented by only the yaw angle γ, thereby reducing the amount of calculation. At the same time, by using multiple candidate attitude angles, the positioning accuracy can be improved.

[0103] After obtaining the candidate pose set, each candidate pose in the candidate pose set may be applied to the initial frame BEV point cloud to obtain a plurality of initial predicted point clouds (320).

[0104] Specifically, after the candidate pose T i Applied to a point P in the initial frame BEV point cloud origin When the corresponding prediction point p predict It can be calculated as follows:

[0105] p predict =T i *P origin =R*P origin +P

[0106] By taking the candidate pose T i Applied to all points in the initial frame BEV point cloud, the corresponding initial predicted point cloud can be obtained.

[0107] Subsequently, a point cloud registration algorithm may be used to register each initial predicted point cloud with the semantic point cloud to obtain a plurality of registration poses and corresponding registration scores (325).

[0108] In some embodiments of the present invention, the point cloud registration algorithm may adopt an NDT algorithm. The NDT algorithm uses statistical probability to determine corresponding points according to the normal distribution of the point cloud, and then calculates the transformation relationship between the source point cloud and the target point cloud. The NDT algorithm is a global registration algorithm that performs registration in a discretized coordinate space by maximizing the score of the source point on the normal distribution probability density calculated after voxelization of the target point.

[0109] In some embodiments, during the iteration process of the NDT registration algorithm, the resolution of the semantic point cloud can be adjusted from coarse to fine. Specifically, the sampling resolution leaf_size can be adjusted by using voxel downsampling filtering. The accuracy can be further improved by adjusting the resolution.

[0110] It should be noted that although the NDT algorithm is described above, this is an example rather than a limitation. In a specific implementation, those skilled in the art may also adopt other point cloud registration algorithms according to actual needs, such as ICP (Iterative Closest Point) algorithm, PFH (Point Feature Histogram) algorithm, etc.

[0111] Finally, the registration pose with the highest registration score among the multiple registration poses may be selected as the initial pose of the vehicle (330).

[0112] It can be seen from process 300 that the initial positioning adopts multi-angle assumption and multi-resolution registration, thereby ensuring the positioning accuracy.

[0113] Figure 5 A laser positioning system 500 for autonomous driving is shown.

[0114] See also Figure 5 , the system 500 may include a data acquisition module 502, a preprocessing module 504, an initialization positioning module 506, a pose increment determination module 508, and a real-time positioning module 510. Each of these modules may be directly or indirectly connected or communicated with each other on one or more buses 512.

[0115] The data acquisition module 502 may be configured to: obtain the prior position of the vehicle, and obtain a multi-frame laser point cloud and a semantic map of the driving environment of the vehicle.

[0116] In various embodiments of the present invention, the a priori position may be obtained by at least one of a GPS module and an RTK module of the vehicle.

[0117] In various embodiments of the present invention, a multi-frame laser point cloud can be obtained by a laser radar of a vehicle scanning the driving environment of the vehicle, and a semantic map is stored in the vehicle or downloaded from a server.

[0118] The preprocessing module 504 may be configured to preprocess the multi-frame laser point cloud and the semantic map to obtain multi-frame BEV point cloud and semantic point cloud, respectively.

[0119] In some embodiments, the preprocessing module 504 can also be configured to preprocess the multi-frame laser point cloud by the following operations: filter out points in the multi-frame laser point cloud whose heights are not within a preset interval; project the remaining points to the plane of z=0; use voxel downsampling filtering to downsample the projected points to obtain a downsampled point cloud; and perform boundary extraction on the downsampled point cloud to obtain a multi-frame BEV point cloud.

[0120] In some embodiments, the semantic map includes a plurality of vector elements, and the pre-processing module 504 may be further configured to pre-process the semantic map by densely upsampling each vector element in the semantic map to obtain a semantic point cloud.

[0121] The initialization positioning module 506 may be configured to perform initial positioning of the vehicle based on the prior position, the initial frame BEV point cloud in the multi-frame BEV point cloud, and the semantic point cloud to determine the initial position and posture of the vehicle.

[0122] In some embodiments, the initialization positioning module 506 can also be configured to: determine a set of candidate poses based on the prior position; apply each candidate pose in the candidate pose set to the initial frame BEV point cloud to obtain multiple initial predicted point clouds; use a point cloud registration algorithm to align each initial predicted point cloud and the semantic point cloud to obtain multiple registered poses and corresponding registration scores; and select the registered pose with the highest registration score as the initial pose of the vehicle.

[0123] In some embodiments, the initialization positioning module 506 may also be configured to determine the candidate pose set by: obtaining a plurality of candidate pose angles with the prior position as the center; determining the candidate pose set T based on the prior position and the plurality of candidate pose angles; all ={T0,T1,…,T N-1}, N represents the number of candidate pose angles, where T = [RP], P is the prior position, R is the rotation matrix and is obtained by Rodriguez formula:

[0124] R=I+(1-cosθ)*n^*n^+sinθ*n^

[0125] where θ = ‖[α β γ] T ‖, α represents the pitch angle, β represents the roll angle, γ represents the yaw angle, n is the normalized rotation vector, n = [α β γ] T / θ=[n1 n2 n3] T , n ^ Represents the antisymmetric matrix of n.

[0126] The posture increment determination module 508 may be configured to obtain the posture increment of the vehicle during the vehicle's driving process.

[0127] In some embodiments, the pose increment determination module 508 may also be configured to obtain the pose increment by the following formula:

[0128] ΔT DR =T DR_k0 -1 *T DR_k1 ,

[0129] Where T DR_k0 represents the initial position determined by the DR module of the vehicle, T DR_k1 represents the pose determined by the DR module at the current moment, ΔT DR Indicates the position increment of the current moment relative to the initial moment.

[0130] The real-time positioning module 510 may be configured to determine the real-time position and posture of the vehicle during driving based on the initial position and the position and posture increment.

[0131] In some embodiments, the real-time positioning module 510 can also be configured to: determine the real-time predicted pose based on the initial pose and the pose increment; apply the real-time predicted pose to the BEV point cloud of the current frame to obtain the real-time predicted point cloud; and use a point cloud registration algorithm to align the real-time predicted point cloud and the semantic point cloud to obtain the registered pose as the real-time pose of the vehicle.

[0132] Although Figure 5 Specific modules of system 500 are shown in FIG. 5 , but it should be understood that these modules are exemplary and non-restrictive. In different implementations, one or more of these modules can be combined, split, removed, or additional modules can be added. For example, in some implementations, data acquisition module 502 and pre-processing module 504 can be combined into a single module. In some implementations, system 500 can also include additional modules, such as an output module (not shown in the figure).

[0133] Figure 6 A block diagram of a device 600 including a laser positioning system for autonomous driving is shown.

[0134] The device illustrates a general hardware environment in which the present application may be applied according to its exemplary embodiments.

[0135] Now refer to Figure 6 Device 600 is described, which is an exemplary embodiment of a hardware device that can be applied to various aspects of the present application. Device 600 can be any machine configured to perform processing and / or computing, and can be, but is not limited to, a workstation, a server, a desktop computer, a laptop computer, a tablet computer, a personal digital assistant (PDA), a smart phone, or any combination thereof.

[0136] The device 600 may include components that may be connected or in communication with the bus 612 via one or more interfaces. For example, the device 600 may include the bus 612, a processor 602, a memory 604, an input device 608, and an output device 610, among other things.

[0137] Processor 602 may be any type of processor, and may include, but is not limited to, a general-purpose processor and / or a special-purpose processor (e.g., a special processing chip), an intelligent hardware device (e.g., a general-purpose processor, a DSP, a CPU, a microcontroller, an ASIC, an FPGA, a programmable logic device, a discrete gate or transistor logic component, a discrete hardware component, or any combination thereof). In some cases, processor 602 may be configured to operate a memory array using a memory controller. In other cases, a memory controller (not shown) may be integrated into processor 602. Processor 602 may be responsible for managing the bus and general processing, including executing software stored on the memory. Processor 602 may also be configured to perform various functions described herein related to laser positioning for autonomous driving. For example, processor 602 can be configured to: obtain a priori position of the vehicle; obtain a multi-frame laser point cloud and a semantic map of the driving environment of the vehicle; pre-process the multi-frame laser point cloud and the semantic map to obtain multi-frame BEV point clouds and semantic point clouds, respectively; initialize the positioning of the vehicle based on the priori position, the initial frame BEV point cloud in the multi-frame BEV point cloud, and the semantic point cloud to determine the initial posture of the vehicle; obtain the posture increment of the vehicle during the driving process of the vehicle; and determine the real-time posture of the vehicle during driving based on the initial posture and the posture increment.

[0138] The memory 604 may be any storage device that can implement data storage. The memory 604 may include, but is not limited to, a disk drive, an optical storage device, a solid-state memory, a floppy disk, a floppy disk, a hard disk, a tape or any other magnetic medium, an optical disk or any other optical medium, a ROM (read-only memory), a RAM (random access memory), a cache memory and / or any other memory chip or box, and / or any other medium from which a computer can read data, instructions and / or codes. The memory 604 may store computer executable software 606 including computer readable instructions that, when executed, cause the processor to perform various functions described herein related to laser positioning for autonomous driving.

[0139] Input device 608 may be any type of device that can be used to input information.

[0140] Output device 610 may be any type of device for outputting information. In one embodiment, output device 610 may be any type of output device that can display information.

[0141] The technical solution of the present invention obtains BEV point cloud by preprocessing laser point cloud, thereby reducing the dimension of point cloud and reducing the amount of calculation and storage space. In addition, the use of vectorized semantic map instead of global three-dimensional point cloud map reduces the amount of data. At the same time, multi-angle hypothesis multi-resolution registration is adopted in the initialization positioning process to ensure accuracy. Compared with the existing laser SLAM technology, the technical solution of the present invention has comparable positioning accuracy, reduces the algorithm time by about 30%, and reduces the memory usage by about 60%, which greatly improves the performance of the algorithm and helps to deploy on platforms with limited computing resources.

[0142] The detailed description described above in conjunction with the accompanying drawings describes examples and does not represent all examples that can be implemented or fall within the scope of the claims. The terms "example" and "exemplary" when used in this specification mean "used as an example, instance or illustration" and do not mean "better or better than other examples".

[0143] Reference throughout this specification to "one embodiment" or "an embodiment" means that a particular feature, structure, or characteristic described in conjunction with the embodiment is included in at least one embodiment of the present application. Thus, use of these phrases may refer to more than just one embodiment. Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0144] The previous description is provided to enable any person skilled in the art to practice the various aspects described herein. Various modifications to these aspects will be readily apparent to those skilled in the art, and the universal principles defined herein may be applied to other aspects. Therefore, the claims are not intended to be limited to the aspects shown herein, but should be granted the full scope consistent with the claims in language, wherein the singular reference to the elements is not intended to mean "there is and only one", but "one or more", unless otherwise specifically stated. Unless otherwise specifically stated, the term "some" refers to one or more. The elements of the various aspects described throughout this application are all structurally and functionally equivalent schemes currently or hereafter known to those of ordinary skill in the art, and are expressly incorporated herein by reference, and are intended to be covered by the claims.

[0145] It should also be noted that these embodiments may be described as a process depicted as a flow chart, flow diagram, structure diagram, or block diagram. Although the flow chart may describe the operations as sequential processes, many of these operations can be performed in parallel or concurrently. In addition, the order of these operations can be rearranged.

[0146] Although various embodiments have been illustrated and described, it should be understood that the embodiments are not limited to the precise configuration and components described above. Various modifications, substitutions and improvements obvious to those skilled in the art may be made in the arrangement, operation and details of the devices disclosed herein without departing from the scope of the claims.

Claims

1. A laser positioning method for autonomous driving, comprising: Obtain the prior position of the vehicle; Obtaining a multi-frame laser point cloud and a semantic map of the driving environment of the vehicle; Preprocessing the multi-frame laser point cloud and the semantic map to obtain a multi-frame BEV point cloud and a semantic point cloud respectively; Initializing the positioning of the vehicle based on the prior position, the initial frame BEV point cloud in the multi-frame BEV point cloud, and the semantic point cloud to determine the initial position and posture of the vehicle; Obtaining a position increment of the vehicle during the driving of the vehicle; as well as The real-time posture of the vehicle during driving is determined based on the initial posture and the posture increment.

2. The method according to claim 1, wherein: The a priori position is obtained by at least one of a GPS module and an RTK module of the vehicle.

3. The method according to claim 1, wherein: The multi-frame laser point cloud is obtained by scanning the driving environment of the vehicle by the laser radar of the vehicle, and the semantic map is stored in the vehicle or downloaded from a server.

4. The method according to claim 1, wherein: Preprocessing the multi-frame laser point cloud to obtain a multi-frame BEV point cloud further includes: Screening out points whose heights are not within a preset range in the multi-frame laser point cloud; Project the remaining points onto the plane z=0; Downsample the projected points using voxel downsampling filtering to obtain a downsampled point cloud; Boundary extraction is performed on the downsampled point cloud to obtain the multi-frame BEV point cloud.

5. The method according to claim 1, wherein: The semantic map includes a plurality of vector elements, and preprocessing the semantic map to obtain a semantic point cloud further includes: Each vector element in the semantic map is densely upsampled to obtain the semantic point cloud.

6. The method according to claim 1, wherein: Initializing the positioning of the vehicle based on the prior position, the initial frame BEV point cloud and the semantic point cloud to determine the initial position of the vehicle further includes: Determine a set of candidate poses based on the prior position; Applying each candidate pose in the candidate pose set to the initial frame BEV point cloud to obtain a plurality of initial predicted point clouds; Using a point cloud registration algorithm to register each initial predicted point cloud with the semantic point cloud to obtain a plurality of registration poses and corresponding registration scores; and The registered pose with the highest registration score is selected as the initial pose of the vehicle.

7. The method according to claim 6, wherein: Determining a candidate pose set based on the prior position further comprises: Obtaining a plurality of candidate attitude angles with the prior position as the center of the circle; Determine the candidate pose set T based on the prior position and the plurality of candidate pose angles all ={T0,T1,…,T N-1 }, N represents the number of candidate attitude angles, Where T = [RP], P is the prior position, R is the rotation matrix and is obtained by Rodriguez formula: R=I+(1-cosθ)*n ^ *n ^ +sinθ*n ^ where θ = ‖[αβγ] T ‖, α represents the pitch angle, β represents the roll angle, γ represents the yaw angle, n is the normalized rotation vector, n = [αβγ] T / θ=[n1n2n3] T , n ^ Represents the antisymmetric matrix of n.

8. The method according to claim 1, wherein: The posture increment is obtained by the following formula: ΔT DR =T DR_k0 -1 *T DR_k1 , Where T DR_k0 represents the initial position determined by the dead reckoning (DR) module of the vehicle, T DR_k1 represents the pose determined by the DR module at the current moment, ΔT DR Indicates the position increment of the current moment relative to the initial moment.

9. The method according to claim 1, wherein: Determining the real-time posture of the vehicle during driving based on the initial posture and the posture increment further includes: Determine a real-time predicted pose based on the initial pose and the pose increment; Applying the real-time predicted pose to the BEV point cloud of the current frame to obtain a real-time predicted point cloud; and The real-time predicted point cloud and the semantic point cloud are registered using a point cloud registration algorithm to obtain a registered pose as the real-time pose of the vehicle.

10. A laser positioning system for autonomous driving using the laser positioning method according to any one of claims 1 to 9, comprising: A data acquisition module is configured to: Obtain the prior position of the vehicle; Obtaining a multi-frame laser point cloud and a semantic map of the driving environment of the vehicle; A preprocessing module, preprocessing the multi-frame laser point cloud and the semantic map to obtain a multi-frame BEV point cloud and a semantic point cloud respectively; An initialization positioning module is used to initialize the positioning of the vehicle based on the prior position, the initial frame BEV point cloud in the multi-frame BEV point cloud, and the semantic point cloud to determine the initial position and posture of the vehicle; A posture increment determination module, which determines the posture increment of the vehicle during the driving process of the vehicle; as well as A real-time positioning module determines the real-time position and posture of the vehicle during driving based on the initial position and posture increment.

11. A device comprising a processor and a memory, wherein the memory stores program instructions, and the processor executes the program instructions to implement the laser positioning method for autonomous driving as claimed in any one of claims 1 to 9.