Real-time local high-precision map construction method and device and storage medium

Abstract

The invention discloses a real-time local high-precision map construction method and device and a storage medium, and the method comprises the steps of real-time detection, feature preprocessing, matching and optimization, and local high-precision map generation. The method has the following characteristics: 1, the real-time detected road condition and speedometer pose are used, and through preprocessing, fusion optimization and map processing, the real-time local high-precision map is constructed; and constructing a real-time local high-precision map. 2, preprocessing operations such as time synchronization, filtering, fitting, sub-map construction, confidence degree design and the like can be carried out on multi-path detection results; 3, global matching and optimization are carried out on the smooth sub-map and the coarse prior map, and a high-precision map outside the detection range is output. And 4, carrying out topological construction on the detection result by using the coarse prior map topological information, and generating a local map required by planning. 5, a foreground and background high-precision map reading thread is designed, and when a local map is updated, quick and non-inductive map switching can be achieved.

Description

Technical Field

[0001] The present invention relates to the field of high-level assisted driving and autonomous driving, and in particular to a method, device and storage medium for constructing a real-time local high-precision map. Background Art

[0002] With the rapid development of autonomous driving technology, high-precision maps are increasingly used in autonomous driving. Pre-built high-precision maps contain rich road information and semantic information, and have a fairly high accuracy, which improves positioning accuracy and stability, reduces planning difficulty, and greatly improves the stability and safety of autonomous driving. However, the disadvantages of pre-building high-precision maps are also obvious: high-precision maps need to be updated in a timely manner, mapping is expensive, and there are data privacy and security issues; therefore, a real-time local high-precision map construction method is urgently needed to solve the above problems. Summary of the invention

[0003] In order to solve the above problems of the prior art solutions, the present invention proposes a real-time local high-precision map construction method, device and storage medium.

[0004] In order to achieve the above object, the present invention adopts the following technical solution: a method for constructing a real-time local high-precision map, comprising the following steps:

[0005] S1, real-time detection, including but not limited to one or more of periscopic camera detection, fisheye camera detection, and laser camera detection, to obtain feature detection results;

[0006] S2, feature preprocessing, restore 3D information from various feature detection results, then perform time alignment through odometer, design confidence through feature distribution, Ransac fitting, build submap, MeanShift, voxel filtering, B-spline fitting operations to complete the smooth construction of smooth submap (also called smooth local semantic submap, the same below);

[0007] S3, matching and optimization, matching and optimizing the smooth sub-map with the coarse prior map, specifically by transforming the coordinates of the features in the non-detection range in the coarse prior map and splicing them with the smooth sub-map to obtain an extended local point cloud map;

[0008] S4, generate a local high-precision map, use the rough prior map topology information to generate the topology of the above-mentioned extended local point cloud map, and obtain a local high-precision map that can be used for planning. Finally, the fast reading and writing module of the local map can be realized through the front and back threads.

[0009] The present invention has the following characteristics: 1. Using the real-time detected road conditions and odometer posture, a real-time local high-precision map is constructed through preprocessing, fusion optimization and map processing. 2. Preprocessing operations such as time synchronization, filtering, fitting, sub-map construction, and confidence design can be performed on the multi-channel detection results. 3. By globally matching and optimizing the smooth sub-map and the coarse prior map, a high-precision map outside the detection range is output. 4. Using the coarse prior map topology information, the detection results are topologically constructed to generate the local map required for planning. 5. The front-end and back-end high-precision map reading threads are designed, and when the local map is updated, the map can be switched quickly and seamlessly.

[0010] Furthermore, the specific steps of the S2 feature preprocessing are as follows:

[0011] S21, perform depth recovery on the road conditions including visual lane lines / curbs, and obtain one or both of visual 3D information and laser 3D information; in practical applications, the depth recovery methods used include: laser-assisted depth recovery, that is, giving depth information to the projection point after the laser is projected onto the image; lane lines and curbs are both on the ground, and the depth information is obtained by assuming that the ground is flat; monocular and binocular depth recovery methods, etc. Finally, 3D information of various elements is obtained.

[0012] S22, when the visual 3D information and the laser 3D information are received at the same time, the odometer posture time synchronization is used for the visual 3D information and the laser 3D information, and the formula is as follows:

[0013]

[0014] Among them, T is the 4*4 transformation matrix corresponding to the time odometer posture.

[0015] S23, perform nearest neighbor search on the synchronized 3D information, and the repeated area has a high confidence level; perform clustering and Ransac fitting on the lane lines and curbs, and filter outliers to remove obvious outliers and reduce the thickness of the detection results;

[0016] S24, constructing a coarse sub-map with a range of N meters (200-500m) and a maximum number of point clouds by the method of step S22;

[0017] S25, performing MeanShift, downsampling and B-spline fitting operations on the coarse sub-map again to obtain a smooth sub-map.

[0018] Furthermore, in S3, the smooth submap and the coarse prior map are matched and optimized to obtain the optimized pose value, and then the coarse prior map is matched. Figure 1 The map elements within a certain range are converted into a smooth sub-map to obtain an extended local point cloud map.

[0019] Furthermore, in S3, the steps for matching and optimizing the smooth submap and the coarse prior map are as follows:

[0020] S31, read the map features and construct a binary search tree of the Kd-tree point cloud for storage;

[0021] S32, using the initial posture to perform a radius search in the Kd-tree to find map feature points within a certain range;

[0022] S33, classifying the sub-map feature points and the feature points in the map to perform Euclidean distance nearest neighbor matching as a set of candidate matching point pairs, and performing certain threshold screening and Ransac fitting and outlier filtering operations at the same time to obtain a set of nearest matching point pairs;

[0023] S34, setting the maximum number of iterations and the iteration termination condition, the initial posture T of each iteration is the posture after optimization in the previous iteration;

[0024] S35, using all matching point pairs Match{plidar (smooth submap), pmap (coarse prior map)'} and pose Pose, construct the following residual:

[0025]

[0026] S36, use Levenberg–Marquardt to optimize Pose and minimize the residual e of the above steps;

[0027] S37, continuously repeat steps S33, S34, and S35 until the iteration condition is met, and output the optimized pose.

[0028] Furthermore, after steps S31-S37 are completed, the features in the coarse prior map that are beyond the detection range are transformed in coordinates through the optimized Pose to obtain an extended map, and finally the smoothed sub-map and the extended map are spliced ​​to obtain the final extended local point cloud map.

[0029] Furthermore, in S4, a binary search tree is constructed with coarse prior map elements, and the nearest adjacent pairs are found using the extended local point cloud map and given topological attributes to generate a local high-precision map that can be used for planning.

[0030] The present invention also provides a device for constructing a real-time local high-precision map, comprising:

[0031] A real-time detection module, including but not limited to one or more of a panoramic camera detection, a fisheye camera detection, and a laser camera detection, to obtain feature detection results;

[0032] The feature preprocessing module restores 3D information from various feature detection results, then performs time alignment through the odometer, designs confidence through feature distribution, Ransac fitting, builds submaps, MeanShift, voxel filtering, and B-spline fitting operations to complete the construction of smooth submaps (also called smooth local semantic submaps, the same below);

[0033] The matching and optimization module matches and optimizes the smooth sub-map with the coarse prior map. Specifically, the coordinates of the features in the non-detection range of the coarse prior map are transformed and spliced ​​with the smooth sub-map to obtain an extended local point cloud map.

[0034] The local high-precision map generation module uses the rough prior map topology information to generate the topology of the above-mentioned extended local point cloud map to obtain a local high-precision map that can be used for planning. Finally, the fast reading and writing module of the local map can be realized through the front and back threads.

[0035] Furthermore, the specific steps of feature preprocessing by the feature preprocessing module are as follows:

[0036] S21, perform depth recovery on the road conditions including visual lane lines / curbs, and obtain one or both of visual 3D information and laser 3D information; the depth recovery methods that can be used in practical applications include: laser-assisted depth recovery, that is, giving depth information to the projection point after the laser is projected onto the image; lane lines and curbs are on the ground, and the depth information is obtained by assuming that the ground is flat; monocular and binocular depth recovery methods, etc. Finally, 3D information of various elements is obtained.

[0037] S22, when the visual 3D information and the laser 3D information are received at the same time, the odometer posture time synchronization is used for the visual 3D information and the laser 3D information, and the formula is as follows:

[0038]

[0039] Among them, T is the 4*4 transformation matrix corresponding to the time odometer posture.

[0040] S23, perform nearest neighbor search on the synchronized 3D information, and the repeated area has a high confidence level; perform clustering and Ransac fitting on the lane lines and curbs, and filter outliers to remove obvious outliers and reduce the thickness of the detection results;

[0041] S24, constructing a coarse sub-map with a range of N meters (200-500m) and a maximum number of point clouds by the method of step S22;

[0042] S25, performing MeanShift, downsampling and B-spline fitting operations on the coarse sub-map again to obtain a smooth sub-map.

[0043] Furthermore, the matching and optimization module performs matching optimization on the smooth sub-map and the coarse prior map to obtain the optimized pose value, and then converts the coarse prior map into Figure 1 The map elements within a certain range are converted into a smooth sub-map to obtain an extended local point cloud map;

[0044] The steps for matching and optimizing the smooth submap and the coarse prior map are as follows:

[0045] S31, read the map features and construct a binary search tree of the Kd-tree point cloud for storage;

[0046] S32, using the initial posture to perform a radius search in the Kd-tree to find map feature points within a certain range;

[0047] S33, classifying the sub-map feature points and the feature points in the map to perform Euclidean distance nearest neighbor matching as a set of candidate matching point pairs, and performing certain threshold screening and Ransac fitting and outlier filtering operations at the same time to obtain a set of nearest matching point pairs;

[0048] S34, setting the maximum number of iterations and the iteration termination condition, the initial posture T of each iteration is the posture after optimization in the previous iteration;

[0049] S35, using all matching point pairs Match{plidar (smooth submap), pmap (coarse prior map)'} and pose Pose, construct the following residual:

[0050]

[0051] S36, use Levenberg–Marquardt to optimize Pose and minimize the residual e of the above steps;

[0052] S37, continuously repeating steps S33, S34, and S35 until the iteration condition is met, and outputting the optimized pose;

[0053] After steps S31-S37 are completed, the features beyond the detection range in the coarse prior map are transformed into coordinates through the optimized Pose to obtain the extended map. Finally, the smoothed sub-map and the extended map are spliced ​​to obtain the final extended local point cloud map.

[0054] Furthermore, the local high-precision map generation module constructs a binary search tree with coarse prior map elements, uses the extended local point cloud map to find the nearest adjacent pairs and assigns topological attributes, thereby generating a local high-precision map that can be used for planning.

[0055] The present invention also provides a computer-readable storage medium containing a computer program, which, when executed by one or more processors, implements any of the above-mentioned methods for constructing a real-time local high-precision map. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A schematic diagram of a method for constructing a real-time local high-precision map according to the present invention;

[0057] Figure 2 It is a schematic diagram of feature preprocessing process;

[0058] Figure 3 A schematic diagram of the process of matching and optimizing the smoothed submap and the coarse prior map;

[0059] Figure 4 This is a schematic diagram of the real-time detection fusion results of the curb module;

[0060] Figure 5a and Figure 5b To expand the local point cloud map schematic;

[0061] Figure 6 This is a schematic diagram of a real-time local high-precision map. DETAILED DESCRIPTION

[0062] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0063] like Figure 1 As shown, a real-time local high-precision map construction method includes the following steps:

[0064] S1, real-time detection, including but not limited to one or more of periscopic camera detection, fisheye camera detection, and laser camera detection, to obtain feature detection results;

[0065] S2, feature preprocessing, restore 3D information from various feature detection results, then perform time alignment through odometer, design confidence through feature distribution, Ransac fitting, build submap, MeanShift, voxel filtering, B-spline fitting operations to complete the smooth construction of smooth submap (also called smooth local semantic submap, the same below);

[0066] S3, matching and optimization, matching and optimizing the smooth sub-map with the coarse prior map, specifically by transforming the coordinates of the features in the non-detection range in the coarse prior map and splicing them with the smooth sub-map to obtain an extended local point cloud map;

[0067] S4, generate a local high-precision map, use the rough prior map topology information to generate the topology of the above-mentioned extended local point cloud map, and obtain a local high-precision map that can be used for planning. Finally, the fast reading and writing module of the local map can be realized through the front and back threads.

[0068] The present invention has the following characteristics: 1. Using the real-time detected road conditions and odometer posture, a real-time local high-precision map is constructed through preprocessing, fusion optimization and map processing. 2. Preprocessing operations such as time synchronization, filtering, fitting, sub-map construction, and confidence design can be performed on the multi-channel detection results. 3. By globally matching and optimizing the smooth sub-map and the coarse prior map, a high-precision map outside the detection range is output. 4. Using the coarse prior map topology information, the detection results are topologically constructed to generate the local map required for planning. 5. The front-end and back-end high-precision map reading threads are designed, and when the local map is updated, the map can be switched quickly and seamlessly.

[0069] The input of the present invention is the detection results of the camera and / or radar, a coarse prior map or a simple topological map, and the odometer posture. The output is a real-time local high-precision map that can be used for planning, and it can also include a local map fast reading and writing module.

[0070] The present invention includes feature preprocessing, matching and optimization of smooth submaps and coarse prior maps, map resampling, and local map fast reading and writing modules. The feature preprocessing part restores 3D information from the results of various feature detections, then performs time alignment through the odometer, designs confidence through feature distribution, Ransac fitting, constructs submaps, MeanShift, voxel filtering, B-spline fitting and other operations to complete the construction of a smooth local semantic submap; next, by matching and optimizing the smooth submap and the coarse prior map, the features within the non-detection range in the coarse prior map are transformed in coordinates, and spliced ​​with the smooth submap to obtain an extended local point cloud map; the coarse prior map topology information is used to generate topology for the above-mentioned extended local point cloud map to obtain a local high-precision map that can be used for planning. Each part is explained in detail below.

[0071] like Figure 2 As shown, the specific steps of the S2 feature preprocessing are as follows:

[0072] S21, perform depth recovery on the road conditions including visual lane lines / curbs, and obtain one or both of visual 3D information and laser 3D information; in practical applications, the depth recovery methods used include: laser-assisted depth recovery, that is, giving depth information to the projection point after the laser is projected onto the image; lane lines and curbs are both on the ground, and the depth information is obtained by assuming that the ground is flat; monocular and binocular depth recovery methods, etc. Finally, 3D information of various elements is obtained.

[0073] S22, when the visual 3D information and the laser 3D information are received at the same time, the odometer posture time synchronization is used for the visual 3D information and the laser 3D information, and the formula is as follows:

[0074]

[0075] Among them, T is the 4*4 transformation matrix corresponding to the time odometer posture.

[0076] S23, perform nearest neighbor search on the synchronized 3D information, and the repeated area has a high confidence level; perform clustering and Ransac fitting on the lane lines and curbs, and filter outliers to remove obvious outliers and reduce the thickness of the detection results;

[0077] S24, constructing a coarse sub-map with a range of N meters (e.g., set to 200-500 meters) and a maximum number of point clouds through the method of step S22;

[0078] S25, performing MeanShift, downsampling and B-spline fitting operations on the coarse sub-map again to obtain a smooth sub-map.

[0079] like Figure 3 As shown in S3, the smooth submap and the coarse prior map are matched and optimized to obtain the optimized pose value, and then the coarse prior map is matched. Figure 1 The map elements within a certain range are converted into a smooth sub-map to obtain an extended local point cloud map. The certain range mentioned above can be set in actual operation, such as 200 meters before and after or other suitable distance ranges; map elements include lane lines, traffic lights, crosswalks, intersections, slopes, etc.

[0080] Specifically, in S3, the steps for matching and optimizing the smooth submap and the coarse prior map are as follows:

[0081] S31, read the map features and construct a binary search tree of the Kd-tree point cloud for storage;

[0082] S32, using the initial posture to perform a radius search in the Kd-tree to find map feature points within a certain range; the certain range can be set in actual operation, such as 200 meters before and after or other suitable distance ranges, which are consistent with the aforementioned range.

[0083] S33, classify the sub-map feature points and the feature points in the map to perform Euclidean distance nearest neighbor matching as a set of candidate matching point pairs, and perform certain threshold screening and Ransac fitting and outlier filtering operations to obtain the nearest matching point pair set; the threshold can be set as needed. In actual operation, this threshold is generally set to 20 cm; when a point exceeding 20 cm is matched with the current point, the reliability may be reduced.

[0084] S34, setting the maximum number of iterations and the iteration termination condition, the initial posture T of each iteration is the posture after optimization in the previous iteration;

[0085] S35, using all matching point pairs Match{plidar (smooth submap), pmap (coarse prior map)'} and pose Pose, construct the following residual:

[0086]

[0087] S36, use Levenberg–Marquardt to optimize Pose and minimize the residual e of the above steps;

[0088] S37, continuously repeat steps S33, S34, and S35 until the iteration condition is met, and output the optimized pose.

[0089] After steps S31-S37 are completed, the features beyond the detection range in the coarse prior map are transformed into coordinates through the optimized Pose to obtain the extended map. Finally, the smoothed sub-map and the extended map are spliced ​​to obtain the final extended local point cloud map.

[0090] Finally, in S4, a binary search tree is constructed with coarse prior map elements, and the extended local point cloud map is used to find the nearest adjacent pairs and assign topological attributes to generate a local high-precision map that can be used for planning.

[0091] The schematic diagram of the real-time detection fusion result in the present invention (only the curb module is shown) is as follows Figure 4 As shown, green and pink are visual detection results, and blue is laser detection results.

[0092] The schematic diagram of expanding the local point cloud map in the present invention is as follows Figure 5a and Figure 5b As shown, the red part is the extended submap, and the pink and blue parts are the smooth submaps.

[0093] The schematic diagram of the real-time local high-precision map used for planning in the present invention is as follows: Figure 6 As shown,

[0094] The method of the present invention can construct a local high-precision map in real time. When the autonomous driving scene changes or there is an error in the high-precision map, there is no need to collect data specifically, spend a lot of manpower and material resources to update the high-precision map, or even to produce a high-precision high-precision map.

[0095] The present invention also provides a device for constructing a real-time local high-precision map, comprising:

[0096] A real-time detection module, including but not limited to one or more of a panoramic camera detection, a fisheye camera detection, and a laser camera detection, to obtain feature detection results;

[0097] The feature preprocessing module restores 3D information from various feature detection results, then performs time alignment through the odometer, designs confidence through feature distribution, Ransac fitting, builds submaps, MeanShift, voxel filtering, and B-spline fitting operations to complete the construction of smooth submaps (also called smooth local semantic submaps, the same below);

[0098] The matching and optimization module matches and optimizes the smooth sub-map with the coarse prior map. Specifically, the coordinates of the features in the non-detection range of the coarse prior map are transformed and spliced ​​with the smooth sub-map to obtain an extended local point cloud map.

[0099] The local high-precision map generation module uses the coarse prior map topology information to perform topology generation on the above-mentioned extended local point cloud map to obtain a local high-precision map that can be used for planning.

[0100] Furthermore, the specific steps of feature preprocessing by the feature preprocessing module are as follows:

[0101] S21, perform depth recovery on the road conditions including visual lane lines / curbs, and obtain one or both of visual 3D information and laser 3D information; the depth recovery methods that can be used in practical applications include: laser-assisted depth recovery, that is, giving depth information to the projection point after the laser is projected onto the image; lane lines and curbs are on the ground, and the depth information is obtained by assuming that the ground is flat; monocular and binocular depth recovery methods, etc. Finally, 3D information of various elements is obtained.

[0102] S22, when the visual 3D information and the laser 3D information are received at the same time, the odometer posture time synchronization is used for the visual 3D information and the laser 3D information, and the formula is as follows:

[0103]

[0104] Among them, T is the 4*4 transformation matrix corresponding to the time odometer posture.

[0105] S23, perform nearest neighbor search on the synchronized 3D information, and the repeated area has a high confidence level; perform clustering and Ransac fitting on the lane lines and curbs, and filter outliers to remove obvious outliers and reduce the thickness of the detection results;

[0106] S24, constructing a coarse sub-map with a range of N meters (200-500m) and a maximum number of point clouds by the method of step S22;

[0107] S25, performing MeanShift, downsampling and B-spline fitting operations on the coarse sub-map again to obtain a smooth sub-map.

[0108] Furthermore, the matching and optimization module performs matching optimization on the smooth sub-map and the coarse prior map to obtain the optimized pose value, and then converts the coarse prior map into Figure 1 The map elements within a certain range are converted into a smooth sub-map to obtain an extended local point cloud map;

[0109] The steps for matching and optimizing the smooth submap and the coarse prior map are as follows:

[0110] S31, read the map features and construct a binary search tree of the Kd-tree point cloud for storage;

[0111] S32, using the initial posture to perform a radius search in the Kd-tree to find map feature points within a certain range;

[0112] S33, classifying the sub-map feature points and the feature points in the map to perform Euclidean distance nearest neighbor matching as a set of candidate matching point pairs, and performing certain threshold screening and Ransac fitting and outlier filtering operations at the same time to obtain a set of nearest matching point pairs;

[0113] S34, setting the maximum number of iterations and the iteration termination condition, the initial posture T of each iteration is the posture after optimization in the previous iteration;

[0114] S35, using all matching point pairs Match{plidar (smooth submap), pmap (coarse prior map)'} and pose Pose, construct the following residual:

[0115]

[0116] S36, use Levenberg–Marquardt to optimize Pose and minimize the residual e of the above steps;

[0117] S37, continuously repeating steps S33, S34, and S35 until the iteration condition is met, and outputting the optimized pose;

[0118] After steps S31-S37 are completed, the features beyond the detection range in the coarse prior map are transformed into coordinates through the optimized Pose to obtain the extended map. Finally, the smoothed sub-map and the extended map are spliced ​​to obtain the final extended local point cloud map.

[0119] Furthermore, the local high-precision map generation module constructs a binary search tree with coarse prior map elements, uses the extended local point cloud map to find the nearest adjacent pairs and assigns topological attributes, thereby generating a local high-precision map that can be used for planning.

[0120] The present invention also provides a computer-readable storage medium containing a computer program, which, when executed by one or more processors, implements any of the above-mentioned methods for constructing a real-time local high-precision map.

[0121] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for constructing a real-time local high-precision map, characterized in that: The following steps are involved: S1, real-time detection, including but not limited to one or more of periscopic camera detection, fisheye camera detection, and laser camera detection, to obtain feature detection results; S2, feature preprocessing, restore 3D information from various feature detection results, then perform time alignment through odometer, design confidence through feature distribution, Ransac fitting, build submap, MeanShift, voxel filtering, and B-spline fitting operations to complete the smooth construction of smooth submap; S3, matching and optimization, matching and optimizing the smooth sub-map with the coarse prior map, specifically by transforming the coordinates of the features in the non-detection range in the coarse prior map and splicing them with the smooth sub-map to obtain an extended local point cloud map; S4, generating a local high-precision map, using the coarse prior map topology information to perform topology generation on the above-mentioned extended local point cloud map to obtain a local high-precision map that can be used for planning.

2. The method for constructing a real-time local high-precision map according to claim 1, characterized in that: The specific steps of the S2 feature preprocessing are as follows: S21, performing deep recovery of road conditions including visual lane lines / curbs to obtain one or both of visual 3D information and laser 3D information; S22, when the visual 3D information and the laser 3D information are received at the same time, the odometer posture time synchronization is used for the visual 3D information and the laser 3D information, and the formula is as follows: Among them, T is the 4*4 transformation matrix corresponding to the time odometer posture. S23, perform nearest neighbor search on the synchronized 3D information, and the repeated area has a high confidence level; perform clustering and Ransac fitting on the lane lines and curbs, and filter outliers to remove obvious outliers and reduce the thickness of the detection results; S24, constructing a coarse sub-map with a range of N meters (200-500m) and a maximum number of point clouds by the method of step S22; S25, performing MeanShift, downsampling and B-spline fitting operations on the coarse sub-map again to obtain a smooth sub-map.

3. The method for constructing a real-time local high-precision map according to claim 2, characterized in that: In S3, the smooth submap and the coarse prior map are matched and optimized to obtain the optimized pose value, and then the map elements within a certain range of the coarse prior map are converted to the smooth submap to obtain the extended local point cloud map.

4. The method for constructing a real-time local high-precision map according to claim 3, characterized in that: In S3, the steps for matching and optimizing the smooth submap and the coarse prior map are as follows: S31, read the map features and construct a binary search tree of the Kd-tree point cloud for storage; S32, using the initial posture to perform a radius search in the Kd-tree to find map feature points within a certain range; S33, classifying the sub-map feature points and the feature points in the map to perform Euclidean distance nearest neighbor matching as a set of candidate matching point pairs, and performing certain threshold screening and Ransac fitting and outlier filtering operations at the same time to obtain a set of nearest matching point pairs; S34, setting the maximum number of iterations and the iteration termination condition, the initial posture T of each iteration is the posture after optimization in the previous iteration; S35, using all matching point pairs Match{pl idar (smooth submap), pmap (coarse prior map)'} and pose Pose, construct the following residual: S36, use Levenberg–Marquardt to optimize Pose and minimize the residual e of the above steps; S37, continuously repeat steps S33, S34, and S35 until the iteration condition is met, and output the optimized pose.

5. The method for constructing a real-time local high-precision map according to claim 4, characterized in that: After steps S31-S37 are completed, the features beyond the detection range in the coarse prior map are transformed into coordinates through the optimized Pose to obtain the extended map. Finally, the smoothed sub-map and the extended map are spliced ​​to obtain the final extended local point cloud map.

6. The method for constructing a real-time local high-precision map according to claim 5, characterized in that: In S4, a binary search tree is constructed with coarse prior map elements, and the nearest adjacent pairs are found using the extended local point cloud map and given topological attributes to generate a local high-precision map that can be used for planning.

7. A device for constructing a real-time local high-precision map, characterized in that: include: A real-time detection module, including but not limited to one or more of a panoramic camera detection, a fisheye camera detection, and a laser camera detection, to obtain feature detection results; The feature preprocessing module restores 3D information from various feature detection results, then performs time alignment through the odometer, designs confidence through feature distribution, performs Ransac fitting, builds submaps, MeanShift, voxel filtering, and B-spline fitting operations to complete the smooth construction of smooth submaps; The matching and optimization module matches and optimizes the smooth sub-map with the coarse prior map. Specifically, the coordinates of the features in the non-detection range of the coarse prior map are transformed and spliced ​​with the smooth sub-map to obtain an extended local point cloud map. The local high-precision map generation module uses the coarse prior map topology information to perform topology generation on the above-mentioned extended local point cloud map to obtain a local high-precision map that can be used for planning.

8. The device for constructing a real-time local high-precision map according to claim 7, characterized in that: The specific steps of feature preprocessing by the feature preprocessing module are as follows: S21, performing deep recovery of road conditions including visual lane lines / curbs to obtain one or both of visual 3D information and laser 3D information; S22, when the visual 3D information and the laser 3D information are received at the same time, the odometer posture time synchronization is used for the visual 3D information and the laser 3D information, and the formula is as follows: Among them, T is the 4*4 transformation matrix corresponding to the time odometer posture. S23, perform nearest neighbor search on the synchronized 3D information, and the repeated area has a high confidence level; perform clustering and Ransac fitting on the lane lines and curbs, and filter outliers to remove obvious outliers and reduce the thickness of the detection results; S24, constructing a coarse sub-map with a range of N meters (200-500m) and a maximum number of point clouds by the method of step S22; S25, performing MeanShift, downsampling and B-spline fitting operations on the coarse sub-map again to obtain a smooth sub-map.

9. The device for constructing a real-time local high-precision map according to claim 8, characterized in that: The matching and optimization module matches and optimizes the smooth submap and the coarse prior map to obtain the optimized pose value, and then converts the map elements within a certain range of the coarse prior map into the smooth submap to obtain an extended local point cloud map; The steps for matching and optimizing the smooth submap and the coarse prior map are as follows: S31, read the map features and construct a binary search tree of the Kd-tree point cloud for storage; S32, using the initial posture to perform a radius search in the Kd-tree to find map feature points within a certain range; S33, classifying the sub-map feature points and the feature points in the map to perform Euclidean distance nearest neighbor matching as a set of candidate matching point pairs, and performing certain threshold screening and Ransac fitting and outlier filtering operations at the same time to obtain a set of nearest matching point pairs; S34, setting the maximum number of iterations and the iteration termination condition, the initial posture T of each iteration is the posture after optimization in the previous iteration; S35, using all matching point pairs Match{pl idar (smooth submap), pmap (coarse prior map)'} and pose Pose, construct the following residual: S36, use Levenberg–Marquardt to optimize Pose and minimize the residual e of the above steps; S37, continuously repeating steps S33, S34, and S35 until the iteration condition is met, and outputting the optimized pose; After steps S31-S37 are completed, the features beyond the detection range in the coarse prior map are transformed into coordinates through the optimized Pose to obtain the extended map. Finally, the smoothed sub-map and the extended map are spliced ​​to obtain the final extended local point cloud map.

10. The device for constructing a real-time local high-precision map according to claim 5, characterized in that: The local high-precision map generation module constructs a binary search tree with coarse prior map elements, uses an extended local point cloud map to find the nearest adjacent pairs and assigns topological attributes, thereby generating a local high-precision map that can be used for planning.

11. A computer-readable storage medium containing a computer program, characterized in that: When the computer program is executed by one or more processors, the real-time local high-precision map construction method described in any one of claims 1 to 6 is implemented.

Images