A global positioning method and system based on virtual descriptor

Abstract

The present invention relates to a global positioning method and system based on virtual descriptors, the method comprising: preprocessing laser radar point cloud data and extracting feature point clouds; dynamically generating virtual descriptors in sequence with arbitrary positions as the center for key frame feature point clouds; building a sparse descriptor map with adjustable density in combination with positioning information obtained from a GNSS / INS combined positioning system; building a point cloud map using original point cloud information at the same time; extracting SC descriptors from query frame feature point clouds, loading sparse descriptor maps for two-stage weighted search and matching, and obtaining candidate frames and matching initial values; loading point cloud information within a set range around the candidate frames, building a local point cloud map, matching the query frame point cloud with the local point cloud map, and obtaining global positioning information. Compared with the prior art, the present invention can simultaneously have rotation and translation invariance, and has the advantages of small calculation amount, small map acquisition workload, high precision, and strong environmental applicability.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving vehicle positioning, and in particular to a global positioning method and system based on virtual descriptors. Background Art

[0002] In recent years, with the rapid development of autonomous driving technology, the positioning requirements of intelligent vehicles have become increasingly higher. Among them, autonomous driving in fixed areas and autonomous driving in specific scenarios are recognized by the industry as the most promising application scenarios to be implemented first, and the global positioning algorithm based on LiDAR has been widely used in the above scenarios because of its good positioning accuracy and stability.

[0003] At present, the global positioning methods based on LiDAR at home and abroad mainly include: 1. Global positioning method based on point cloud matching. Figure 1 Generally, the method is large in size, resulting in a huge workload and poor real-time performance, and simple algorithms often cannot complete the complex task of global positioning. 2. The global positioning method based on descriptor matching is more efficient and accurate, but the global descriptors (such as SC descriptors) currently proposed only have rotation invariance but not translation invariance. When the global positioning trajectory and the map trajectory have translation, sparse environmental features, or high wall occlusion, position recognition often fails. In addition, real vehicle experiments have shown that the maximum map collection interval allowed for successful positioning based on the SC descriptor is 3m. To this end, it is required to densely collect maps in the vehicle's drivable area, which makes the point cloud map data huge, and easily leads to problems such as low map loading rate and high memory consumption. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a global positioning method and system based on virtual descriptors, which can have both rotation invariance and translation invariance, thereby improving the positioning success rate and accuracy.

[0005] The object of the present invention can be achieved by the following technical solution: A global positioning method based on virtual descriptors comprises the following steps:

[0006] S1. Obtain laser radar point cloud data, and perform point cloud preprocessing and feature point cloud extraction operations in sequence to obtain key frame feature point cloud and query frame feature point cloud;

[0007] S2, dynamically generate virtual descriptors in sequence with any position as the center for the key frame feature point cloud;

[0008] Combined with the positioning information obtained from the GNSS / INS combined positioning system, a sparse descriptor map with adjustable density is constructed;

[0009] At the same time, the original point cloud information is used to construct a point cloud map;

[0010] S3, extracting SC descriptors from the query frame feature point cloud, and then loading the sparse descriptor map to perform two-stage weighted search and matching to obtain candidate frames and matching initial values;

[0011] S4. Load the point cloud information within a set range around the candidate frame to construct a local point cloud map, match the query frame point cloud with the local point cloud map, and obtain global positioning information.

[0012] Furthermore, the step S2 specifically includes the following steps:

[0013] S21, for a key frame feature point cloud, down-sampling using a grid filter of a set size;

[0014] Then, with any position as the center, use a fan-shaped grid to divide the space area;

[0015] S22, traversing the point cloud in each grid, calculating the maximum height feature value corresponding to each grid to obtain the initial value of the virtual descriptor;

[0016] S23, performing an environmental consistency check on each virtual descriptor to determine the final value of the characteristic value of each grid in the virtual descriptor and its environmental consistency weight;

[0017] S24, combining the positioning information obtained from the GNSS / INS combined positioning system, and using the virtual descriptors after the environmental consistency check in step S23, a sparse descriptor map is constructed;

[0018] At the same time, the original point cloud information is used to construct a point cloud map.

[0019] Furthermore, the specific process of using the sector grid to divide the space area in step S21 is: divide one grid at every set angle A along the circumference, a circle of 360 degrees, and a total of (360 / A) grids in the circumference are obtained;

[0020] Divide the grid into a grid every set length B in the radial direction, with the division range from 0 to C meters, and obtain a total of (C / B) grids in the radial direction;

[0021] Finally, we get ((360 / A)*(C / B)) sector grids.

[0022] Furthermore, the initial value of the virtual descriptor in step S22 is specifically:

[0023]

[0024] Among them, I virtual is the initial value of the virtual descriptor, H ijis the maximum height characteristic value of the grid, is a real number, N r is the number of radial grids, N s is the number of grids in the circumferential direction.

[0025] Furthermore, the step S23 specifically adopts a template descriptor method to compare the virtual descriptor with the template descriptor to determine the final value of the feature value of each grid in the virtual descriptor and its environmental consistency weight.

[0026] Furthermore, the final value of the characteristic value of each grid in the virtual descriptor and its environmental consistency weight are specifically:

[0027] If the distance between the point cloud and the virtual descriptor generation center becomes smaller, then:

[0028]

[0029]

[0030] If the distance between the point cloud and the virtual descriptor generation center becomes larger, then:

[0031]

[0032] Where r is the maximum radius of the grid, is the maximum pitch angle of laser scanning, H virtual is the virtual descriptor I virtual The initial value of the grid eigenvalue, H templatek Template descriptor I templatek The grid feature value of For I virtual The final value of the grid eigenvalue after the environmental consistency test, α is the corresponding environmental consistency weight.

[0033] Furthermore, the step S3 specifically includes the following steps:

[0034] S31, taking the laser radar as the center, extracting the corresponding SC descriptor from the query frame feature point cloud;

[0035] S32, loading a sparse descriptor map for location recognition, for the extracted SC descriptors, summing and averaging the rows to obtain a (C / B)-dimensional vector, called a ring key, wherein the ring key of the sparse descriptor map is used to construct a KD tree, and searching for similar keys according to the ring key of the query frame;

[0036] S33, in the fast matching and judgment stage, for each candidate frame and query frame, the similarity between each candidate frame and the query frame is judged by a distance function: the two are compared column by column, the cosine distance of two column vectors with the same index in the two is calculated, the cosine distances of all columns are added and normalized, and a distance value used to describe the similarity of two descriptors is obtained;

[0037] The minimum distance value is screened out to determine the optimal column shift number and distance, where the optimal column shift number is the output matching initial value.

[0038] Furthermore, the searching of similar keys in step S32 specifically uses the weighted L1 distance as a similarity evaluation index to search for a set number of similar keys, wherein the specific calculation formula of the similarity evaluation index is:

[0039]

[0040] Among them, H p is the grid feature value of the query frame descriptor, N r Represents the dimension of the ring bond, that is, the number of radial grids.

[0041] Furthermore, the optimal column shift number in step S33 is specifically:

[0042]

[0043]

[0044]

[0045] Among them, I q ,I c are the descriptors obtained from the query frame and the candidate frame respectively, are the column vectors with the same index from the query frame and the candidate frame, respectively. is a candidate frame descriptor shifted by n columns, N s Represents the dimension of the descriptor column vector, that is, the number of grids in the circumferential direction, d(I q ,I c ) is I q with I c The similarity distance between them.

[0046] Furthermore, the step S4 specifically performs ICP (Iterative Closest Point) matching on the local point cloud map and the query frame point cloud, that is, using the matching initial value obtained in step S3 as the ICP matching initial value to calculate the corresponding global positioning information.

[0047] A global positioning system based on virtual descriptors, comprising a processing and extraction module, a mapping module and a global positioning module, wherein the input end of the processing and extraction module is connected to a laser radar, the output end of the processing and extraction module is respectively connected to the mapping module and the global positioning module, the input end of the mapping module is also connected to a GNSS / INS combined positioning system, the output end of the mapping module is connected to the global positioning module, and the processing and extraction module is used to pre-process the point cloud data collected by the laser radar and extract feature point clouds therefrom;

[0048] The mapping module is used to extract virtual descriptors from the collected single-lane laser point cloud information with any position of the point cloud as the center, and to construct a multi-lane sparse descriptor map and point cloud map with adjustable output density in combination with the positioning information of the GNSS / INS combined positioning system;

[0049] The global positioning module extracts the SC descriptor of the query frame point cloud with the laser radar as the center, performs position recognition by loading the sparse descriptor map, uses the two-stage weighted search matching to output the candidate frame and the initial matching value, and then loads the local point cloud map within the set range around the candidate frame to perform ICP matching with the query frame point cloud to calculate the global positioning information.

[0050] Compared with the prior art, the present invention dynamically generates virtual descriptors in sequence with arbitrary positions as the center for key frame feature point clouds. Since virtual descriptors can be dynamically generated at arbitrary positions in a single frame point cloud, the problem of positioning failure of the existing global descriptors in the case where the global positioning trajectory and the map trajectory have translation, the environmental features are sparse, etc. is solved, and the method has the advantage of translation invariance. Especially in common multi-lane scenarios, in order to achieve successful positioning, the existing methods need to collect map data lane by lane, while the method of the present invention only needs to collect single lane data, which greatly reduces the workload of map collection and data processing, and improves the positioning success rate and accuracy.

[0051] The present invention constructs a sparse descriptor map with adjustable density, extracts SC descriptors from the query frame feature point cloud, and then loads the sparse descriptor map for position identification. That is, only the descriptor map is loaded during matching without loading the entire point cloud map. When the position identification is successful, the local point cloud map near the candidate frame is loaded for precise matching. Since the memory of the descriptor map is much smaller than the memory of the point cloud map, the memory consumption and loading time are greatly reduced, the operating efficiency of the algorithm is improved, and the requirements for device performance are reduced.

[0052] The present invention takes into account the laser radar scanning principle. Each frame of point cloud is collected with the laser radar itself as the center. When the descriptor generation center is translated, the real observable point cloud information will change, but it is not actually collected. Instead, the descriptor is extracted using the original collected point cloud information. Therefore, the problem of uncertainty in the consistency between the virtual descriptor and the environment will occur. Therefore, the present invention is designed to perform environmental consistency test on each virtual descriptor; in addition, the present invention also uses two-stage weighted search matching to output candidate frames and matching initial values. Therefore, on the one hand, by performing environmental consistency test on the virtual descriptor, the credibility of the virtual descriptor is improved; at the same time, by using a two-stage weighted search, the environmental consistency weight can be used, that is, only the descriptor of the part consistent with the environment is used to participate in the search, and the weight of the part of the consistency uncertainty is reduced, so as to realize the function of selecting only the accurate overlapping area for search and matching, and effectively remove the descriptor area that will cause matching failure, thereby effectively improving the matching success rate, and finally improving the accuracy and precision of global positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0054] Figure 2 A schematic diagram of grid division in the embodiment;

[0055] Figure 3 This is a schematic diagram of environmental consistency testing in an embodiment;

[0056] Figure 4 It is a schematic diagram of the system structure principle of the present invention. DETAILED DESCRIPTION

[0057] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] Example

[0059] like Figure 1 As shown, a global positioning method based on virtual descriptors includes the following steps:

[0060] S1. Obtain laser radar point cloud data, and perform point cloud preprocessing and feature point cloud extraction operations in sequence to obtain key frame feature point cloud and query frame feature point cloud;

[0061] S2, dynamically generate virtual descriptors in sequence with any position as the center for the key frame feature point cloud;

[0062] Combined with the positioning information obtained from the GNSS / INS combined positioning system, a sparse descriptor map with adjustable density is constructed;

[0063] At the same time, the original point cloud information is used to construct a point cloud map;

[0064] S3, extracting SC descriptors from the query frame feature point cloud, and then loading the sparse descriptor map to perform two-stage weighted search and matching to obtain candidate frames and matching initial values;

[0065] S4. Load the point cloud information within a set range around the candidate frame to construct a local point cloud map, match the query frame point cloud with the local point cloud map, and obtain global positioning information.

[0066] This embodiment applies the above method, and its specific application process includes:

[0067] 1) Collect laser radar point cloud data, perform point cloud preprocessing and feature point cloud extraction;

[0068] 2) For the key frame feature point cloud, virtual descriptors are dynamically generated in sequence with arbitrary positions as the center. The distance between each center can be flexibly set, and the environmental consistency check is performed based on the template descriptor. Combined with the positioning information of the high-precision GNSS / INS combined positioning system, a sparse descriptor map with adjustable density is constructed. At the same time, a high-precision point cloud map is constructed using the point cloud information.

[0069] The specific process is as follows:

[0070] A frame of point cloud is downsampled using a 0.5-meter grid filter; then a sector grid is used to divide the space area with any position as the center. Figure 2 As shown in the figure, the specific division method is to divide the circumferential direction into a grid every 6 degrees, and a circle of 360 degrees has 60 grids in total; the radial direction is divided into a grid every 4 meters, and the division range is 0 to 80 meters, with a total of 20 grids. The final fan-shaped grid has a total of 1200 (20*60); then traverse the point cloud in each grid and calculate the eigenvalue H of each grid ij : Maximum height value, get a 20*60 array, get the virtual descriptor I virtual The initial value of is calculated as follows:

[0071]

[0072] Considering the scanning principle of LiDAR, each frame of point cloud is collected with LiDAR itself as the center. When the descriptor generation center is translated, the real observable point cloud information will change, but it is not actually collected. Instead, the descriptor is extracted using the original collected point cloud information. Therefore, there will be a problem of uncertainty in the consistency between the virtual descriptor and the environment. Therefore, it is necessary to perform environmental consistency check on each virtual descriptor (such as Figure 3 As shown), this technical solution is combined with the template descriptor I templatek The comparative analysis is used to determine the final characteristic value of each grid in the virtual descriptor and its environmental consistency weight. The formula for calculating the characteristic value of the grid in the template descriptor is:

[0073]

[0074] If the point cloud is closer to the center of the virtual descriptor generation:

[0075]

[0076] If the point cloud is farther away from the center of the virtual descriptor generation:

[0077]

[0078] Where r is the maximum radius of the grid, is the maximum pitch angle of the laser scan; H virtual isI virtual The initial value of the grid eigenvalue; H templatek isI templatek The grid feature value of isI virtual The final value of the grid eigenvalue after environmental consistency test; α is its environmental consistency weight.

[0079] Then, combined with the high-precision positioning information of the GNSS / INS combined positioning system, the extracted virtual descriptors are used to construct a sparse descriptor map. Since the spacing between the centers can be adjusted when the virtual descriptors are generated, the density of the generated sparse descriptor map can be adjusted; at the same time, the original point cloud information is used to construct a high-precision point cloud map.

[0080] 3) Extract SC descriptors from the query frame feature point cloud, load the sparse descriptor map for two-stage weighted search and matching, and obtain candidate frames and matching initial values ​​after successful position recognition. The specific process is as follows:

[0081] With the laser radar as the center, the SC descriptor of the query frame point cloud is extracted; then the constructed sparse descriptor map is loaded for position recognition, and the extracted descriptors are summed and averaged by row to obtain a 20-dimensional vector, called a ring key, where the ring key of the descriptor map is used to construct a KD tree, and the ring key of the query frame is used to find similar keys. In this embodiment, the 50 most similar keys are searched, and the weighted L1 distance is used as the similarity evaluation index. The specific formula is:

[0082]

[0083] In the fast matching and judgment stage, for each candidate frame and query frame, a distance function is used to judge the similarity between each candidate frame and the query frame. The two are compared column by column, the cosine distance of the two column vectors with the same index is calculated, the cosine distance of all columns is added and normalized, and the distance value used to describe the similarity of the two descriptors is obtained. The distance function is:

[0084]

[0085] When the vehicle rotates in place, the descriptor shows that the value of each column remains unchanged, but it will be translated by column. In order to achieve rotation invariance of positioning, the distance between the two descriptors after all possible column displacements is calculated, and the minimum distance is found. This distance is used to evaluate the similarity between the two descriptors. The smaller the distance, the more similar they are. The formula for determining the optimal column displacement number and distance is as follows, where the optimal column displacement number is the output matching initial value:

[0086]

[0087]

[0088] Among them, H p is the grid feature value of the query frame descriptor; I q ,I c They are the descriptors obtained from the query frame and the candidate frame respectively; denotes the column vectors with the same index from the query frame and the candidate frame respectively; is a candidate frame descriptor shifted by n columns, N s represents the dimension of the descriptor column vector, which is 60 in this embodiment.

[0089] 4) After obtaining the candidate frame and the initial matching value, load the point cloud information near the candidate frame to build a local point cloud map, and perform ICP matching with the query frame point cloud. Specifically, use the initial matching value output in step 3) as the initial ICP matching value to calculate and obtain high-precision global positioning information.

[0090] like Figure 4 As shown, the present embodiment also provides a global positioning system based on virtual descriptors, wherein the system includes a processing and extraction module, a mapping module and a global positioning module, wherein the input end of the processing and extraction module is connected to the laser radar, the output end of the processing and extraction module is respectively connected to the mapping module and the global positioning module, the input end of the mapping module is also connected to the GNSS / INS combined positioning system, and the output end of the mapping module is connected to the global positioning module.

[0091] In practical applications, the processing and extraction module is used to pre-process the point cloud data collected by the lidar and extract feature point clouds from it;

[0092] The mapping module is used to extract virtual descriptors from the collected single-lane laser point cloud information with any position of the point cloud as the center, and combine the positioning information of the GNSS / INS combined positioning system to construct a multi-lane sparse descriptor map and point cloud map with adjustable output density;

[0093] The global positioning module extracts the SC descriptor of the query frame point cloud with the lidar as the center, performs location recognition by loading the sparse descriptor map, uses the two-stage weighted search and matching to output the candidate frame and the initial matching value, and then loads the local point cloud map within the set range around the candidate frame to perform ICP matching with the query frame point cloud to calculate the global positioning information.

[0094] In summary, this technical solution solves the problem that global positioning based on descriptors does not have translation invariance, and has the advantages of small computational complexity, small map collection workload, high accuracy, and strong environmental applicability.

Claims

1. A global positioning method based on virtual descriptors, characterized in that: The following steps are involved: S1. Obtain laser radar point cloud data, and perform point cloud preprocessing and feature point cloud extraction operations in sequence to obtain key frame feature point cloud and query frame feature point cloud; S2, dynamically generate virtual descriptors in sequence with any position as the center for the key frame feature point cloud; Combined with the positioning information obtained from the GNSS / INS combined positioning system, a sparse descriptor map with adjustable density is constructed; At the same time, the original point cloud information is used to construct a point cloud map; S3, extracting SC descriptors from the query frame feature point cloud, and then loading the sparse descriptor map to perform two-stage weighted search and matching to obtain candidate frames and matching initial values; S4, loading the point cloud information within a set range around the candidate frame to construct a local point cloud map, matching the query frame point cloud with the local point cloud map to obtain global positioning information; The step S2 specifically includes the following steps: S21, for a key frame feature point cloud, down-sampling using a grid filter of a set size; Then, with any position as the center, use a fan-shaped grid to divide the space area; S22, traversing the point cloud in each grid, calculating the maximum height feature value corresponding to each grid to obtain the initial value of the virtual descriptor; S23, performing an environmental consistency check on each virtual descriptor to determine the final value of the characteristic value of each grid in the virtual descriptor and its environmental consistency weight; S24, combining the positioning information obtained from the GNSS / INS combined positioning system, and using the virtual descriptors after the environmental consistency check in step S23, a sparse descriptor map is constructed; At the same time, the original point cloud information is used to construct a point cloud map; The step S3 specifically comprises the following steps: S31, taking the laser radar as the center, extracting the corresponding SC descriptor from the query frame feature point cloud; S32, loading a sparse descriptor map for location recognition, for the extracted SC descriptors, summing and averaging the rows to obtain a (C / B)-dimensional vector, called a ring key, wherein the ring key of the sparse descriptor map is used to construct a KD tree, and searching for similar keys according to the ring key of the query frame; S33, in the fast matching and judgment stage, for each candidate frame and query frame, the similarity between each candidate frame and the query frame is judged by a distance function: the two are compared column by column, the cosine distance of two column vectors with the same index in the two is calculated, the cosine distances of all columns are added and normalized, and a distance value used to describe the similarity of two descriptors is obtained; The minimum distance value is screened out to determine the optimal column shift number and distance, where the optimal column shift number is the output matching initial value.

2. A global positioning method based on virtual descriptors according to claim 1, characterized in that: The specific process of using the sector grid to divide the space area in step S21 is: divide one grid at every set angle A along the circumference, and a circle of 360 degrees, so that a total of (360 / A) grids in the circumference are obtained; Divide the grid into a grid every set length B in the radial direction, with the division range from 0 to C meters, and obtain a total of (C / B) grids in the radial direction; Finally, we get ((360 / A)*(C / B)) sector grids.

3. A global positioning method based on virtual descriptors according to claim 2, characterized in that: The initial value of the virtual descriptor in step S22 is specifically: Among them, I virtual is the initial value of the virtual descriptor, H ij is the maximum height characteristic value of the grid, is a real number, N r is the number of radial grids, N s is the number of grids in the circumferential direction.

4. A global positioning method based on virtual descriptors according to claim 3, characterized in that: The step S23 specifically uses the template descriptor to compare the virtual descriptor with the template descriptor to determine the final value of the feature value of each grid in the virtual descriptor and its environmental consistency weight: If the distance between the point cloud and the virtual descriptor generation center becomes smaller, then: If the distance between the point cloud and the virtual descriptor generation center becomes larger, then: Where r is the maximum radius of the grid, is the maximum pitch angle of laser scanning, H virtual is the virtual descriptor I virtual The initial value of the grid eigenvalue, H templatek Template descriptor I templatek The grid feature value of For I virtual The final value of the grid eigenvalue after the environmental consistency test, α is the corresponding environmental consistency weight.

5. A global positioning method based on virtual descriptors according to claim 4, characterized in that: The specific method of searching for similar keys in step S32 is to use the weighted L1 distance as a similarity evaluation index to search for a set number of similar keys, wherein the specific calculation formula of the similarity evaluation index is: Among them, H p is the grid feature value of the query frame descriptor, N r Represents the dimension of the ring bond, that is, the number of radial grids.

6. A global positioning method based on virtual descriptors according to claim 5, characterized in that: The optimal column shift number in step S33 is specifically: Among them, I w ,I c are the descriptors obtained from the query frame and the candidate frame respectively, are the column vectors with the same index from the query frame and the candidate frame, respectively. is a candidate frame descriptor shifted by n columns, N s Represents the dimension of the descriptor column vector, that is, the number of grids in the circumferential direction, d(I q ,I c ) is I q with I c The similarity distance between them.

7. A global positioning method based on virtual descriptors according to claim 1, characterized in that: The step S4 specifically performs ICP matching on the local point cloud map and the query frame point cloud, that is, using the initial matching value obtained in step S3 as the initial ICP matching value to calculate the corresponding global positioning information.

8. A global positioning system based on virtual descriptors, using a global positioning method based on virtual descriptors as claimed in claim 1, characterized in that: It includes a processing and extraction module, a mapping module and a global positioning module. The input end of the processing and extraction module is connected to the laser radar, and the output end of the processing and extraction module is connected to the mapping module and the global positioning module respectively. The input end of the mapping module is also connected to the GNSS / INS combined positioning system, and the output end of the mapping module is connected to the global positioning module. The processing and extraction module is used to pre-process the point cloud data collected by the laser radar and extract feature point clouds from it; The mapping module is used to extract virtual descriptors from the collected single-lane laser point cloud information with any position of the point cloud as the center, and to construct a multi-lane sparse descriptor map and point cloud map with adjustable output density in combination with the positioning information of the GNSS / INS combined positioning system; The global positioning module extracts the SC descriptor of the query frame point cloud with the laser radar as the center, performs position recognition by loading the sparse descriptor map, uses the two-stage weighted search matching to output the candidate frame and the initial matching value, and then loads the local point cloud map within the set range around the candidate frame to perform ICP matching with the query frame point cloud to calculate the global positioning information.

Images