3D target detection method and medium for automatic driving-oriented fused high-precision map

Abstract

The application relates to a 3D target detection method for automatic driving and fusion of high-precision maps, which comprises the following steps: collecting a city high-precision map; calculating a drivable area mask graph and a mask conversion matrix; acquiring point clouds when an intelligent vehicle is driving; acquiring the intelligent vehicle pose by a high-precision positioning algorithm; calculating a conversion matrix for converting the point clouds to a city coordinate system, converting the point clouds to the city coordinate system, and obtaining city point clouds; projecting the city point clouds to the drivable area mask graph based on the mask conversion matrix, and calculating point cloud masks; extracting the point clouds in the drivable area; performing data enhancement on the point clouds in the drivable area; detecting the enhanced point clouds by using a 3D target detection model, and obtaining the center position of an object detection frame, the size of the detection frame and a yaw angle. Compared with the prior art, the application has the advantages of improving the inference speed of model detection, reducing the delay of intelligent vehicle system decision response and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automatic driving perception, and in particular to a 3D target detection method for automatic driving based on fusion of high-precision maps and a medium. BACKGROUND

[0002] 3D target detection is one of the key technologies in the field of automatic driving perception, and is related to the real-time dynamic planning of intelligent vehicle routes, decision-making and behavior implementation, including emergency braking, emergency obstacle avoidance and other behaviors. The accuracy of 3D target detection directly affects the accuracy of vehicle system decision-making and behavior implementation, and the speed of detection directly affects the response delay of the vehicle system. This influence is particularly magnified in high-speed driving, and is directly related to the safety of passengers. Using a reasonable and effective detection method can accurately and quickly detect the environment around the intelligent vehicle, thereby improving the safety of automatic driving and protecting the safety of passengers.

[0003] In the 3D target detection method based on laser radar point cloud, there are a large number of redundant background points in the point cloud data, such as buildings in urban environments. If the collected data is detected completely, it may be affected by redundant background points, reducing the recognition accuracy and reasoning speed. In addition, the intelligent vehicle pays different attention to different areas, and full point cloud prediction will to some extent distract the attention of the model.

[0004] High-precision maps provide high-precision road and regional information, which can assist 3D target detection. However, in the method of fusing high-precision maps, the mainstream method is to use attention mechanisms and feature fusion mechanisms to combine the features of high-precision maps and point clouds together for full point cloud detection, still retaining a large number of redundant points, slowing down the reasoning speed of the model detection and increasing the response delay of the intelligent vehicle system decision-making. SUMMARY

[0005] The purpose of the present application is to improve the reasoning speed of the detection model, a 3D target detection method for automatic driving based on fusion of high-precision maps and a medium.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] A 3D target detection method for automatic driving based on fusion of high-precision maps, comprising the following steps:

[0008] Collecting a high-precision city map;

[0009] Based on the high-precision map, calculating a drivable area mask map and a mask conversion matrix;

[0010] Obtaining point cloud data when the intelligent vehicle is driving;

[0011] An absolute pose reference value is calculated, and a high-precision positioning algorithm is used to obtain the pose of the intelligent vehicle;

[0012] A conversion matrix for converting the point cloud data to the urban coordinate system is calculated, the point cloud data is converted to the urban coordinate system, and urban point cloud data is obtained;

[0013] Based on the mask conversion matrix, the urban point cloud data is projected onto the drivable area mask map, and a point cloud mask is calculated;

[0014] Based on the point cloud mask, the point cloud in the drivable area is extracted;

[0015] The point cloud in the drivable area is subjected to data enhancement, and enhanced point cloud is obtained;

[0016] A 3D target detection model is used to detect the enhanced point cloud, and the center position of the object detection frame, the size of the detection frame, and the yaw angle are obtained.

[0017] Further, the urban high-precision map is collected by GNSS combined with laser radar, camera and IMU.

[0018] Further, the high-precision positioning algorithm is a SLAM algorithm.

[0019] Further, the conversion matrix of the point cloud data to the urban coordinate system is:

[0020]

[0021] where SE3D t is the conversion matrix of the urban coordinate system, SE3D t .T is the transfer matrix, SE3D t .R is the rotation matrix, SE3D t .s is the scaling factor, (x r ,y t ,z t ) is the point cloud coordinate.

[0022] Further, the specific steps of extracting the point cloud in the drivable area include:

[0023] Based on the projection coordinates of each point in the point cloud mask, the mask value at that position in the mask map is obtained, which is:

[0024]

[0025] where pm i is the mask value of the i-th point cloud, D mask is the mask map, is the projection coordinates of the i-th point cloud in the mask map;

[0026] determining whether the mask value is 1 or 0, if it is 1, it indicates that it is located in the drivable area, if it is 0, it is located in the non-drivable area, the point cloud in the drivable area is:

[0027] Points d ={p i |0≤i<N,pm i =1}

[0028] wherein, Points d is the point cloud in the drivable area, p i represents the i-th point cloud, and N represents the number of point clouds.

[0029] Further, the point cloud in the drivable area is subjected to data enhancement by using random flipping, random rotation and random world scaling in turn, and the calculation expression of the enhanced point cloud is:

[0030]

[0031] wherein, is the enhanced point cloud, f rflip (*) is a random flipping function, f rrot (*) is a random rotation function, and f rscal (*) is a random world scaling function.

[0032] Further, the 3D target detection model is a CenterPoint model.

[0033] Further, it further comprises:

[0034] visualizing the center position of the object detection frame, the size of the detection frame and the yaw angle.

[0035] Further, the Rviz tool is used for visualization.

[0036] The application also provides a computer readable storage medium comprising one or more programs for execution by one or more processors of an electronic device, the one or more programs comprising instructions for performing the 3D target detection method of the fusion high-definition map for automatic driving as described above.

[0037] Compared with the prior art, the application has the following beneficial effects:

[0038] (1) The application uses a high-precision map to extract a drivable area mask map, projects the obtained point cloud data into the drivable area mask map to calculate a point cloud mask, only retains the point cloud in the drivable area according to the point cloud mask, filters out a large number of redundant background points of the point cloud, and then performs regional detection, so as to improve the reasoning speed of model detection and reduce the delay of intelligent vehicle system decision-making.

[0039] (2) This invention achieves stable acquisition of vehicle pose by calculating absolute pose reference value and integrating the high-precision positioning algorithm of SLAM method.

[0040] (3) This invention demonstrates the entire detection and fusion process and sends the detection results for visualization, providing a systematic and engineered process for using high-precision maps to process point clouds and perform 3D target detection. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0042] Figure 2 This is a flowchart of a high-precision positioning algorithm based on the fusion of lidar, IMU, and GNSS SLAM method in an embodiment of the present invention;

[0043] Figure 3 This is a network model diagram of the CenterPoint 3D target detection model in an embodiment of the present invention. Detailed Implementation

[0044] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0045] This embodiment provides a 3D target detection method based on high-precision maps for autonomous driving, such as... Figure 1 As shown, the method includes the following steps:

[0046] Step S1: Collect a high-precision map of the city.

[0047] High-precision urban maps are collected using GNSS combined with LiDAR, cameras, and IMUs, and pre-saved as vector maps using the world coordinate system. Driving areas are irregular polygonal regions, saved as continuous vertices. Vertex structure is defined as follows:

[0048] Vertex(x,y,z)

[0049] A drivable area consists of several vertices. Adjacent vertices are connected by straight lines. Connecting the first and last vertices forms a drivable area, which is assigned a unique number to identify it.

[0050] Road id ={v1,v2,v3,…,v n}

[0051] Among them, vi represents the i-th vertex in the drivable region, and id represents the number of the drivable region. The number of points n contained is not necessarily the same for different drivable regions.

[0052] Step S2: Based on the high-precision map, calculate the drivable region mask image and the mask conversion matrix.

[0053] The pseudo code of the processing process of the high-precision map drivable region is as follows:

[0054] Input: Roads is N drivable regions, s is the scaling factor

[0055] Output: drivable region mask image D mask and the corresponding conversion matrix SE2D

[0056] 1. Calculate the boundary values of the x-axis and y-axis of all drivable regions Roads: x min ,y min ,x max ,y max

[0057] 2. Calculate the width and height of the mask image:

[0058]

[0059]

[0060] 3. Calculate the conversion matrix for converting the point cloud to the mask image:

[0061] The rotation matrix is a 2×2 unit matrix: SE2D.R=E2(1)

[0062] The translation matrix is:

[0063] The scaling factor: SE2D.s=s

[0064] 4. Define the black and white mask image, single channel, initialize to fill 0:

[0065] D mask =Image(img w ,img h ,color=0)

[0066] 5. Get the mask image:

[0067] for Road in Roads:

[0068] Project to the two-dimensional image mask: roadxy img =SE2D.s*(Road.xy*(SE2D.R)T *SE2D.T

[0069] Rounding the data:

[0070] In mask image D mask Draw a polygon on top, filled with 1: DrawPolygon(D mask ,roadxy img ,fill=1)

[0071] end

[0072] Here, `Image` represents a class defined in the program, and this retrieves an instance of that class. `DrawPolygon` represents the function for drawing polygons; you can use existing packages to draw images from `Image` instances. In the drivable area mask map D... mask In the diagram, 0 indicates that the location is not a drivable area, and 1 indicates that the location is a drivable area.

[0073] Step S3: Obtain point cloud data while the intelligent vehicle is in motion.

[0074] Point cloud data is captured in real time using a 64-line LiDAR. The coordinates and laser reflection intensity of each point are retained as initial features (x, y, z, intensity). Generally, the captured point cloud features also include the line bundle number, but this feature is not used in this embodiment. The original point cloud data frame at time t is shown in the figure. t As shown below:

[0075] Points t ={p i |0i<N}

[0076] p i =(x i ,y i ,z i intensity i )

[0077] Where N represents the number of points in the point cloud data frame, p i Let i represent the i-th point.

[0078] Step S4: Calculate the absolute pose reference value and combine it with a high-precision positioning algorithm to obtain the intelligent parking position pose.

[0079] The embodiment intends to realize high-precision positioning by adopting a SLAM method of fusing a laser radar, an IMU and a GNSS. In a good GNSS signal area, the GNSS can provide an absolute pose reference value, eliminate cumulative errors and make the ego vehicle positioning not drift over time. In a poor or missing GNSS signal area, precise pose estimation can also be realized by using a laser radar and an IMU.

[0080] The flowchart of the algorithm in the embodiment is shown in FIG. 1, and the specific process is as follows. Figure 2

[0081] Firstly, positioning information is acquired by using a GPS, and laser point cloud and pose information are acquired by using a laser radar and an IMU. Subsequently, the algorithm performs IMU pre-integration, and realizes point cloud distortion removal, feature extraction and current frame and historical map matching by using the pre-integration result, so as to obtain a laser odometer. The algorithm maintains two factor graphs: an IMU bias optimization factor graph optimizes the bias of the IMU based on the pose result of the current frame, and re-integrates by using the optimized bias, thereby improving the precision of pre-integration; and a global optimization factor graph performs global ego vehicle pose optimization based on loop detection results and positioning information provided by the GNSS, so as to obtain an optimized ego vehicle pose Pose t . The precise pre-integration result can better realize point cloud distortion removal and provide a better initial value for laser point cloud matching, thereby improving the matching precision and further improving the positioning effect of the entire system.

[0082] The pose Pose t of the intelligent vehicle at the current time t is as follows:

[0083]

[0084] The pose Pose t contains two parts: a position and a rotation state in a world coordinate system. Among them, (x t , y t , z t ) represents coordinates, describes the position of the ego vehicle at time t, and R t represents a rotation matrix, describes the rotation state of the ego vehicle at time t.

[0085] Step S5: calculating a conversion matrix of converting the point cloud data to a city coordinate system, converting the point cloud data to the city coordinate system to obtain city point cloud data.

[0086] ​The point cloud data frame needs to be unified with the urban high-definition map in the same coordinate system. The point cloud data frame adopts the vehicle coordinate system, with the front of the vehicle as the x-axis, following the right-hand system, the left side as the y-axis, and upward as the z-axis. The urban high-definition map coordinate system adopts the world coordinate system. The method converts the point cloud data frame to the world coordinate system, and the corresponding coordinate conversion matrix is obtained according to the pose of the ego vehicle. In the vehicle coordinate system, the origin (0, 0, 0) represents the ego vehicle position, and after the conversion matrix, the pose Pose t of the world coordinate system is obtained. According to the pose Pose t of the world coordinate system, the conversion matrix SE3D t of the point cloud coordinate system to the urban coordinate system at the current t time is obtained:

[0087]

[0088] wherein SE3D t .T represents the transition matrix, SE3D t .R represents the rotation matrix, and SE3D t .s represents the scaling factor. Considering that no scaling is performed in the coordinate conversion process, the scaling factor is set to 1.

[0089] Since each point contains four features, namely position information and laser reflection intensity, the position of the point changes, but the laser reflection intensity of the point remains unchanged, and no coordinate conversion is required, and the consistency of the feature before and after conversion needs to be maintained. Therefore, according to the conversion matrix SE3D t , the corresponding conversion is performed on the position of each point, and the specific conversion is as follows:

[0090]

[0091]

[0092] wherein, represents the feature of the i-th point after conversion to the urban coordinate system, and the corresponding urban coordinate system point cloud data frame

[0093]

[0094] The point cloud data frame maintains the order invariance before and after conversion.

[0095] Step S6: based on the mask conversion matrix, projecting the urban point cloud data to the drivable area mask map, and calculating the point cloud mask.

[0096] According to the urban high-definition map conversion matrix SE2D, the urban point cloud data frame is converted to the mask map D maskThe above yields the mask image D corresponding to each point. mask Coordinates on:

[0097]

[0098]

[0099] That is, to Each point in Rotate, translate, and scale the x and y coordinates, and multiply them by the corresponding matrices or numerical values ​​to obtain the projected coordinates. Finally, the point cloud mask P is obtained. mask .

[0100] Step S7: Extract the point cloud in the drivable area based on the point cloud mask.

[0101] According to the point cloud mask P mask Projected coordinates of each point Obtain the mask image D mask The mask value at this position:

[0102]

[0103] The function f(*) is used to extract the mask image D. mask Middle position The mask value at that location. If the mask value at that location is pm i A value of 1 indicates If the location is within the drivable area, which is also the key area of ​​focus for the intelligent vehicle, then the original point p corresponding to that location is retained. i If the mask value at this location is pm i A value of 0 indicates that the point is located in a non-drivable area and is discarded. The point cloud of the drivable area is as follows:

[0104] Points d ={p i |0≤i<N,pm i =1}

[0105] Step S8: Perform data augmentation on the point cloud in the drivable area to obtain an augmented point cloud.

[0106] Data augmentation is performed on the point cloud of the drivable area. The resulting drivable area point cloud Ooints... d Data augmentation is performed by random flipping, random rotation, and random world scaling.

[0107] For the random flip method, the embodiment randomly flips the point cloud data along the x-axis of the point cloud, and the probability of flipping is set to 0.5. For the random rotation method, the embodiment sets the random rotation angle range to (-0.78539816, 0.78539816), and uses the rotation angle to construct a rotation matrix for rotation along the z-axis to multiply the coordinates, and then splices the rotated point cloud coordinates with the original reflection feature intensity to obtain the enhanced data after random rotation. For the random world scaling method, the embodiment sets the random scaling factor range to (0.95, 1.05), and directly multiplies the point cloud coordinates by the random scaling factor. The specific calculation formula is as follows:

[0108]

[0109] wherein, fr rflip (*) and f rrot (*) and f rscal (*) respectively represent the random flip function, the random rotation function and the random world scaling function.

[0110] Step S9: using a 3D target detection model to detect the enhanced point cloud to obtain the center position of the object detection box, the size of the detection box and the yaw angle.

[0111] The D target detection model used in the embodiment is a CenterPoint model, and the network structure of the model is as shown in the accompanying Figure 3 CenterPoint uses point representation of targets to simplify the three-dimensional target detection task, and the specific steps are as follows:

[0112] Input the processed laser point cloud data frame Voxelize the point cloud data frame; use VoxelNet to extract features from the voxelized point cloud to obtain a feature vector map, denoted as M; project M into the BEV perspective to obtain the BEV feature M B ; send the BEV feature M B to the 2D Backbone for further feature extraction to obtain the feature F; send the feature F to the CenterPoint detection head to obtain the 3D bounding box BBox; according to the 3D bounding box BBox, obtain the corresponding bounding box feature from the BEV feature M B , and perform two-stage refinement to finally obtain the optimized 3D bounding box. Finally, the center position (x, y, z) of each 3D object prediction box, the size (l, w, h) of the box and the yaw angle (yaw) are predicted.

[0113] Step S10: send the detection result and visualize. The detection result of the current frame is attached to the current timestamp t to form an information stream sent to other modules of the subsequent vehicle-mounted system, and the detection result is visualized. In this embodiment, the publication of the information stream uses ROS (Robot Operating System). A detection result node is created using ROS to convert the detection result into its corresponding data format and publish it to the node. Other modules of the vehicle-mounted system can subscribe to the node through ROS to obtain the detection result and perform corresponding processing. The visualization module uses the visualization tool Rviz provided by ROS to visualize point clouds, images and other data.

[0114] To verify the effectiveness of the above method, this embodiment tests the publicly available Argoverse2 dataset, which is the largest lidar dataset to date, with 20,000 30-second sequences, including lidar sensor data, high-definition maps, and ego vehicle poses. Each scene contains its own high-definition map, which contains 3D lane, crosswalk, and drivable area geometry, and the data comes from data captured in six different cities in the United States. The dataset is randomly divided into training, validation, and test sets with 16,000, 2,000, and 2,000 scenes, respectively.

[0115] Table 1 Comparison of the number of class annotation boxes in Argoverse2 and other datasets

[0116]

[0117] The Argoverse2 dataset not only contains high-definition map data, has a large amount of data, and is rich in categories, but also has a large number of 3D annotation boxes corresponding to each category. For the three categories of vehicles, pedestrians, and cyclists that are mainly concerned about autonomous driving, the comparison between Argoverse2 and other datasets is shown in Table 1:

[0118] In the above method, the high-definition map is used to process the point cloud collected by the lidar to extract the point cloud in the drivable area. To verify the effectiveness of this step, the number of points in the point cloud before and after extraction in some scenes is compared in Table 2. The difference in the number of points in the point cloud before and after extraction is related to the distribution of the drivable area. In scenes with dense road areas, the proportion of point clouds in the drivable area is relatively high; in scenes with dense buildings, the proportion of point clouds in the drivable area is low. Through testing, the average proportion of point clouds in the drivable area in the total point cloud is about 23.14%. As can be seen from Table 2, the method proposed in this embodiment can greatly remove redundant background information and reduce the size of the input of the detection model.

[0119] Table 2 Comparison of the number of points in the point cloud and the drivable area in some scenes

[0120]

[0121] Table 3 Comparison of CenterPoint on drivable area point cloud and full point cloud detection effect

[0122]

[0123] Table 4 CenterPoint effect on Argoverse2 validation set

[0124]

[0125] In order to verify the performance of the embodiment method, CenterPoint is taken as a 3D target detection model, and training is performed on the Argoverse2 dataset. The method of the present application is verified on the Argoverse2 dataset in Table 3. As can be seen from Table 3, compared with detecting the full point cloud, the FPS index of the method proposed in the present application is more optimal, and the detection speed is faster. The precision of the detection is shown in Table 4, and as can be seen from Table 4, compared with other methods, the detection precision of the present application is higher.

[0126] The above functions, if realized in the form of a software function unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on such an understanding, the technical solutions of the present application or parts of the present application that essentially contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0127] Those skilled in the art will appreciate that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages, such as object-oriented programming languages Java and interpreted scripting language JavaScript.

[0128] The present application is described in reference to the flowchart and / or block diagram of the method, apparatus (system) and computer program product according to an embodiment of the present application. It will be understood that each block of the flowchart and / or block diagram, and combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0129] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0130] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0131] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those of skill in the art once they have the benefit of the present disclosure without departing from the spirit and scope of the application. Accordingly, the attached claims are intended to cover all such variations and modifications as falling within the scope of the application.

[0132] 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 detection method for automatic driving oriented fusion high-definition map, characterized in that, The method comprises the following steps: Collecting a high-precision map of a city; Based on the high-precision map, calculating a drivable area mask graph and a mask conversion matrix; Obtaining point cloud data when the intelligent vehicle is driving; Calculating an absolute pose reference value, and obtaining the pose of the intelligent vehicle in combination with a high-precision positioning algorithm; Calculating a conversion matrix for converting the point cloud data to a city coordinate system, converting the point cloud data to the city coordinate system to obtain city point cloud data, and the conversion matrix for converting the point cloud data to the city coordinate system is: wherein, is a conversion matrix for the urban coordinate system, is a transfer matrix, is a rotation matrix, is a scaling factor, is a point cloud coordinate; Based on the mask conversion matrix, projecting the city point cloud data to the drivable area mask graph, and calculating a point cloud mask; Based on the point cloud mask, extracting point clouds in the drivable area, and the specific steps of extracting point clouds in the drivable area include: Based on the projection coordinates of each point in the point cloud mask, obtaining the mask value of the position in the mask graph, which is: wherein, is a mask value for the i th point cloud, is a mask map, is a projection coordinate for the i th point cloud in the mask map; Determining whether the mask value is 1 or 0, if it is 1, it indicates that it is located in the drivable area, if it is 0, it is located in the non-drivable area, and the point cloud in the drivable area is: wherein, is a point cloud in the drivable area, p i denotes the i point cloud, N denotes the number of point clouds; Performing data enhancement on the point cloud in the drivable area to obtain enhanced point cloud; Using a 3D target detection model to detect the enhanced point cloud to obtain the center position of the object detection frame, the size of the detection frame, and the yaw angle.

2. The 3D target detection method for a fusion high-definition map oriented to automatic driving according to claim 1, characterized in that, The city high-precision map is collected by GNSS combined with laser radar, camera and IMU.

3. The 3D target detection method for a fusion high-definition map oriented to automatic driving according to claim 1, characterized in that, The high-precision positioning algorithm is a SLAM algorithm.

4. The 3D target detection method for a fusion high-definition map oriented to automatic driving according to claim 1, characterized in that, The point cloud in the drivable area is sequentially enhanced by random flipping, random rotation and random world scaling, and the calculation expression of the enhanced point cloud is: wherein, is an enhanced point cloud, , , are a random flip function, a random rotation function and a random world scale function, respectively.

5. The 3D target detection method for automatic driving-oriented fusion high-definition map according to claim 1, characterized in that, The 3D target detection model is a CenterPoint model.

6. The 3D target detection method for a fusion high-definition map oriented to automatic driving according to claim 1, characterized in that, It also includes: Visualizing the center position of the object detection frame, the size of the detection frame, and the yaw angle.

7. The 3D target detection method for a fusion high-definition map oriented to automatic driving according to claim 6, characterized in that, Visualizing by using the Rviz tool.

8. A computer-readable storage medium, characterized in that, One or more programs for one or more processors of an electronic device, the one or more programs comprising instructions for performing the 3D target detection method for the fusion high-precision map for autonomous driving according to any one of claims 1-7.

Images