Point cloud registration method and system for unmanned driving assisted positioning

Abstract

A point cloud registration method for unmanned driving assisted positioning, comprising: 1) segmenting a source point cloud and a target point cloud to be matched into multiple super points through the following iterative process: randomly selecting a point P that does not belong to any super point; defining a new super point as a point located at a fixed radius R centered at the point P; sphere 1) A set of points in a sphere; 2) Project each superpoint into a depth map; 3) Use a deep neural network to compress the depth map of each superpoint into a 5×2 feature matrix; 4) Match candidate point selection and rough registration; 5) Perform fine registration through the ICP algorithm. The present invention has a fast registration speed, improves the problem that the ICP algorithm is difficult to start with when facing disordered point clouds, and improves the traditional point cloud registration algorithm in terms of accuracy and efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of point cloud registration, and in particular relates to a point cloud registration method and a system thereof for unmanned driving assisted positioning. Background Art

[0002] Point cloud registration technology is a basic technology in the field of 3D vision. Its goal is to obtain perfect coordinate transformation through calculation, and integrate point cloud data at different perspectives into a specified coordinate system through rigid transformations such as rotation and translation. With the promotion and application of LiDAR in the fields of robots and unmanned vehicles, point cloud registration technology has been increasingly valued as a core module in the fields of high-precision maps and high-precision positioning. However, unlike the rapid development of point cloud recognition and detection technology, there have not been many breakthroughs in point cloud registration. At present, the mainstream solution in the field of unmanned vehicles is still the ICP algorithm and its derivative algorithms. The main reasons are as follows: 1) ICP and its derivative algorithms are very accurate and difficult to replace; 2) The complexity of the unmanned vehicle problem itself. Compared with general point cloud registration technology, in the unmanned vehicle environment, point cloud registration technology not only faces worse initial values ​​(poor GNSS signals), but also needs to deal with the interference of dynamic objects. These problems have posed new challenges to point cloud registration technology.

[0003] Generally speaking, there will be a certain overlap between point clouds collected from different perspectives. How to find matching pairs for registration from the overlapping areas is the key to the matching method. Therefore, the point cloud registration problem can be transformed into the problem of quickly finding matching pairs. Traditional methods for finding matching pairs mostly rely on the similarity of geometric features. The geometric features used include Euclidean distance, normal vector in Euclidean space, and covariance matrix, etc. However, due to reasons such as collection perspective and occlusion, most of the collected point clouds are incomplete, which means that these low-level geometric features will not be particularly accurate, affecting the accuracy of the entire matching method. Summary of the invention

[0004] Based on this, in order to solve the above technical problems, a point cloud registration method and system for unmanned driving assisted positioning are provided.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A point cloud registration method for unmanned driving assisted positioning, comprising:

[0007] 1) The source point cloud and the target point cloud to be matched are divided into multiple super points through the following iterative process:

[0008] Randomly select a point P that does not belong to any superpoint; define a new superpoint as a point with a fixed radius R centered at the point P. sphereA set of points in a sphere, where

[0009] V local is the volume of the sphere containing the point cloud, m is the number of matching pairs used for matching, and m is at least 3;

[0010] 2) Project each superpoint into a depth map;

[0011] 3) Using a deep neural network to compress the depth map of each superpoint into a 5×2 feature matrix;

[0012] 4) Matching candidate point selection and rough registration:

[0013] Calculate the Euclidean distance between each feature matrix of the source point cloud and the nearest N feature matrices in the target point cloud. If the Euclidean distance of the latter feature matrix in the source point cloud suddenly increases compared with the Euclidean distance of the former feature matrix, then take the super point corresponding to the former feature matrix as the candidate point, and the super point corresponding to the latter feature matrix as the invalid point.

[0014] Using a random sampling consensus algorithm to iteratively find at least three point cloud registration matrices between the candidate points and super points of the target point cloud;

[0015] 5) Using the point cloud registration matrix as the initial transformation matrix, perform precise registration using the ICP algorithm to obtain the optimal transformation matrix.

[0016] The m is 5.

[0017] The projecting each superpoint into a depth map further comprises:

[0018] a. Establish a normalized local coordinate system: set the centroid of each superpoint as the origin of the coordinate system, calculate the covariance matrix of each superpoint, and use singular value decomposition to obtain the corresponding x-axis, y-axis, and z-axis of the coordinate system;

[0019] b. Project each superpoint into the two-dimensional image along the x-axis and y-axis directions, use the height of the z-axis to represent the color of the depth map, and cut the circular boundary to obtain a square two-dimensional image.

[0020] The size of the two-dimensional image is 64×64 pixels, and the size of the two-dimensional picture is 32×32 pixels.

[0021] The deep neural network adopts DAE neural network.

[0022] The present invention also relates to a point cloud registration system for unmanned driving assisted positioning, comprising a storage module, wherein the storage module comprises a plurality of instructions loaded and executed by a processor:

[0023] 1) The source point cloud and the target point cloud to be matched are divided into multiple super points through the following iterative process:

[0024] Randomly select a point P that does not belong to any superpoint; define a new superpoint as a point with a fixed radius R centered at the point P. sphere A set of points in a sphere, where

[0025] V local is the volume of the sphere containing the point cloud, m is the number of matching pairs used for matching, and m is at least 3;

[0026] 2) Project each superpoint into a depth map;

[0027] 3) Using a deep neural network to compress the depth map of each superpoint into a 5×2 feature matrix;

[0028] 4) Matching candidate point selection and rough registration:

[0029] Calculate the Euclidean distance between each feature matrix of the source point cloud and the nearest N feature matrices in the target point cloud. If the Euclidean distance of the latter feature matrix in the source point cloud suddenly increases compared with the Euclidean distance of the former feature matrix, then take the super point corresponding to the former feature matrix as the candidate point, and the super point corresponding to the latter feature matrix as the invalid point.

[0030] Using a random sampling consensus algorithm to iteratively find at least three point cloud registration matrices between the candidate points and super points of the target point cloud;

[0031] 5) Using the point cloud registration matrix as the initial transformation matrix, perform precise registration using the ICP algorithm to obtain the optimal transformation matrix.

[0032] The m is 5.

[0033] The projecting each superpoint into a depth map further comprises:

[0034] a. Establish a normalized local coordinate system: set the centroid of each superpoint as the origin of the coordinate system, calculate the covariance matrix of each superpoint, and use singular value decomposition to obtain the corresponding x-axis, y-axis, and z-axis of the coordinate system;

[0035] b. Project each superpoint into the two-dimensional image along the x-axis and y-axis directions, use the height of the z-axis to represent the color of the depth map, and cut the circular boundary to obtain a square two-dimensional image.

[0036] The size of the two-dimensional image is 64×64 pixels, and the size of the two-dimensional picture is 32×32 pixels.

[0037] The deep neural network adopts DAE neural network.

[0038] The present invention has a fast registration speed, improves the problem that the ICP algorithm is difficult to operate when facing disordered point clouds, and improves the traditional point cloud registration algorithm in terms of accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments:

[0040] Figure 1 is a flow chart of the present invention;

[0041] Figure 2 It is a schematic diagram of a depth map obtained after the super point is projected according to the present invention;

[0042] Figure 3 It is a schematic diagram of the DAE neural network model of the present invention;

[0043] Figure 4 A schematic diagram of the visualized changes of the depth image during the compression process of the present invention;

[0044] Figure 5 Schematic diagram of the transformation relationship between the six degrees of freedom between point clouds. DETAILED DESCRIPTION

[0045] like Figure 1 As shown, a point cloud registration method for unmanned driving assisted positioning includes:

[0046] S101 . Segment the source point cloud and the target point cloud to be matched into multiple super points through the following iterative process.

[0047] First, a point P that does not belong to any superpoint is randomly selected; then a new superpoint is defined as a point with a fixed radius R centered at point P. sphere A set of points in a sphere, where

[0048] V local is the volume of the sphere containing the point cloud. When segmenting the source point cloud, V local is the volume of the sphere containing the source point cloud. When segmenting the target point cloud, V local is the volume of the sphere containing the target point cloud, m is the number of matching pairs used for matching, and m is at least 3.

[0049] In this embodiment, since five matching matrix pairs are to be selected, m is 5.

[0050] In the process of autonomous driving, the point cloud data collected in advance and marked with positions is called the target point cloud, while the point cloud collected in real time during autonomous driving is called the source point cloud. Rough registration is to quickly estimate a rough point cloud registration matrix (as the initial matrix for ICP precise registration, including translation and rotation matrices) when the relative position between the target point cloud and the source point cloud is completely unknown. The entire registration process requires a relatively high calculation speed, but does not require too high a precision of the calculation results.

[0051] The present invention introduces the concept of superpoint and uses it as the basic unit of the matching process. Each superpoint is a subset of points describing a local surface and allows two superpoints to have common parts directly (i.e., a point can be contained in several superpoints). Through iteration, the vast majority of scan points in the point cloud (about 95%) can be included.

[0052] S102, project each superpoint into a depth map. The specific process is as follows:

[0053] a. Establish a normalized local coordinate system: Set the centroid of each superpoint as the origin of the coordinate system, calculate the covariance matrix of each superpoint, and use singular value decomposition to obtain the corresponding x-axis, y-axis, and z-axis of the coordinate system.

[0054] For point cloud data, each point has three-dimensional information, that is, each point can be expressed in the form of P(x, y, z). Then, for a superpoint containing several point clouds, the mean of each dimension can be calculated separately, and then the covariance matrix of each superpoint is obtained as:

[0055] Among them, cov is the covariance calculation function.

[0056] After obtaining the covariance matrix, the singular value decomposition of the matrix can be expressed as follows: C = UΣV T , where U and V are both unit orthogonal matrices, becoming left singular matrices and right singular matrices, Σ has values ​​only on the main diagonal, called singular values, and the other elements are all 0,

[0057]

[0058] The superpoint is segmented along the radial direction and the average height is calculated. The largest segmented part is the direction of the X-axis; the Z-axis points to the direction of the smallest eigenvector; the Y-axis is perpendicular to the XZ plane. This coordinate system maintains the geometric properties of the superpoint while not changing the original position and direction of the superpoint.

[0059] b. Project each superpoint into a two-dimensional image along the x-axis and y-axis directions, use the height of the z-axis to represent the color of the depth map, and cut the circular boundary to obtain a square two-dimensional image, see Figure 2 .

[0060] In this embodiment, the size of the two-dimensional image is 64×64 pixels, and the size of the two-dimensional picture is 32×32 pixels, that is, the two-dimensional image of 64×64 pixels is cut into a size of 32×32 pixels.

[0061] The purpose of projecting into a depth map is to reduce the influence of factors such as point cloud distribution intensity and random noise. While preserving the original point cloud image geometric features and point cloud distribution features, the depth image converts the original 3D point cloud data into image information, which facilitates the subsequent use of neural networks for image matching.

[0062] S103, using a deep neural network to compress the depth map of each superpoint into a 5×2 pixel image, which can be represented as a 5×2 feature matrix.

[0063] In this embodiment, the deep neural network adopts a DAE neural network to improve the compression speed and compression quality of the image.

[0064] The DAE neural network architecture consists of two stages: encoding and decoding. The encoding stage starts from the input layer and then connects to the hidden layer, gradually decreasing until the desired image size is reached. The decoding stage starts with a compact representation of the data, and the dimension of each subsequent hidden layer increases at a time until the output layer dimension reaches the input dimension. The loss function is defined as the pixel error between the input layer and the output layer. The optimization network achieves the best compact representation of the image. A hidden layer with 4 full connections is used. A sigmoid nonlinear activation function is used between each layer of the network and the input dropout. Finally, each depth image is compressed into a 5×2 pixel image for registration. Figure 3 It is a neural network model, which is divided into an input layer, an output layer and four intermediate layers (hidden layers); Figure 4 Shows the visualization of the depth image during compression.

[0065] S104, matching candidate point selection and rough registration:

[0066] Calculate the Euclidean distance between each feature matrix of the source point cloud and the nearest N feature matrices in the target point cloud. If the Euclidean distance of the latter feature matrix in the source point cloud suddenly increases compared with the Euclidean distance of the previous feature matrix, the super point corresponding to the previous feature matrix is ​​taken as a candidate point, and the super point corresponding to the latter feature matrix is ​​an invalid point.

[0067] A random sampling consensus algorithm is used to iteratively find at least three point cloud registration matrices between candidate points and superpoints of the target point cloud.

[0068] For example, if the Euclidean distance of the i-th feature matrix suddenly increases compared to the Euclidean distance of the i-1-th feature matrix, then the super point corresponding to the i-1-th feature matrix is ​​taken as a candidate point, and the super point corresponding to the i-th feature matrix is ​​regarded as an invalid point.

[0069] In this embodiment, N = 3. Analyzing the experimental data, for a source point cloud containing 10 million points and a target point cloud containing 500,000 points, the latter can be segmented into about 200 super points. When N = 3, there are about 550 candidate matching objects.

[0070] A random sampling consensus algorithm is used to iteratively find at least three point cloud registration matrices between candidate points and superpoints of the target point cloud.

[0071] The random sampling consensus algorithm (RANSAC) can provide an initial transformation matrix for point cloud precision registration and delete incorrect registration pairs. The basic idea of ​​the algorithm is to first randomly sample samples, then estimate the parameters of the sampled samples, and use the estimated parameters to divide all data into inliers (sample points that meet the estimated parameters) and outliers (sample points that do not meet the estimated parameters). Finally, the largest set of inliers is used to reversely recalculate the model parameters. Assume that the best initial registration matrix M is found from the target point cloud through iteration:

[0072]

[0073] Among them, R is the rotation matrix and T is the translation matrix.

[0074] In order to find the transformation relationship between the six degrees of freedom between point clouds (i.e., up and down (UP, DOWN), front and back (FORWARD, BACK), left and right (RIGHT, LEFT) translation and pitch (PITCH), roll (ROLL), and yaw (YAW), see Figure 5 ), at least 3 matching pairs are required. In order to enhance the robustness of the algorithm, this embodiment adopts 6 matching pairs.

[0075] S105: Using the above point cloud registration matrix as an initial transformation matrix, perform precise registration through an ICP algorithm to obtain an optimal transformation matrix.

[0076] The iterative closest point (ICP) algorithm consists of three stages: 1) obtaining the registration point pairs; 2) registering and estimating the optimal initial transformation matrix; 3) iterative registration to obtain the optimal transformation matrix.

[0077] Compared with the traditional method that uses a single point cloud data as the matching benchmark, the present invention uses a "super point" containing several point cloud data as the basic processing unit in the matching process, thereby improving the registration speed. In addition, unlike the traditional method that directly searches for the point with the most similar geometric features as the matching point, the present invention uses the similarity between the nearest neighbor features of the data after dimensionality reduction processing, and obtains the corresponding matching pairs based on the similarity, and then obtains multiple optimal transformation matrices based on the matching pairs to obtain a rough matching result. Finally, the ICP algorithm is used for precise matching, and the optimal point cloud registration matrix is ​​obtained through optimal evaluation. The present invention not only improves the problem that the ICP algorithm is difficult to start with when facing an unordered point cloud, but also improves the traditional point cloud registration algorithm in terms of accuracy and efficiency.

[0078] The present invention also relates to a point cloud registration system for unmanned driving assisted positioning, comprising a storage module, wherein the storage module comprises a plurality of instructions loaded and executed by a processor:

[0079] S101 . Segment the source point cloud and the target point cloud to be matched into multiple super points through the following iterative process.

[0080] First, a point P that does not belong to any superpoint is randomly selected; then a new superpoint is defined as a point with a fixed radius R centered at point P. sphere A set of points in a sphere, where

[0081] V local is the volume of the sphere containing the point cloud. When segmenting the source point cloud, V local is the volume of the sphere containing the source point cloud. When segmenting the target point cloud, V local is the volume of the sphere containing the target point cloud, m is the number of matching pairs used for matching, and m is at least 3.

[0082] In this embodiment, since five matching matrix pairs are to be selected, m is 5.

[0083] In the process of autonomous driving, the point cloud data collected in advance and marked with positions is called the target point cloud, while the point cloud collected in real time during autonomous driving is called the source point cloud. Rough registration is to quickly estimate a rough point cloud registration matrix (as the initial matrix for ICP precise registration, including translation and rotation matrices) when the relative position between the target point cloud and the source point cloud is completely unknown. The entire registration process requires a relatively high calculation speed, but does not require too high a precision of the calculation results.

[0084] The present invention introduces the concept of superpoint and uses it as the basic unit of the matching process. Each superpoint is a subset of points describing a local surface and allows two superpoints to have common parts directly (i.e., a point can be contained in several superpoints). Through iteration, the vast majority of scan points in the point cloud (about 95%) can be included.

[0085] S102, project each superpoint into a depth map. The specific process is as follows:

[0086] b. Establish a normalized local coordinate system: Set the centroid of each superpoint as the origin of the coordinate system, calculate the covariance matrix of each superpoint, and use singular value decomposition to obtain the corresponding x-axis, y-axis, and z-axis of the coordinate system.

[0087] For point cloud data, each point has three-dimensional information, that is, each point can be expressed in the form of P(x, y, z). Then, for a superpoint containing several point clouds, the mean of each dimension can be calculated separately, and then the covariance matrix of each superpoint is obtained as:

[0088] Among them, cov is the covariance calculation function.

[0089] After obtaining the covariance matrix, the singular value decomposition of the matrix can be expressed as follows: C = UΣV T , where U and V are both unit orthogonal matrices, becoming left singular matrices and right singular matrices, Σ has values ​​only on the main diagonal, called singular values, and the other elements are all 0,

[0090]

[0091] The superpoint is segmented along the radial direction and the average height is calculated. The largest segmented part is the direction of the X-axis; the Z-axis points to the direction of the smallest eigenvector; the Y-axis is perpendicular to the XZ plane. This coordinate system maintains the geometric properties of the superpoint while not changing the original position and direction of the superpoint.

[0092] b. Project each superpoint into a two-dimensional image along the x-axis and y-axis directions, use the height of the z-axis to represent the color of the depth map, and cut the circular boundary to obtain a square two-dimensional image, see Figure 2 .

[0093] In this embodiment, the size of the two-dimensional image is 64×64 pixels, and the size of the two-dimensional picture is 32×32 pixels, that is, the two-dimensional image of 64×64 pixels is cut into a size of 32×32 pixels.

[0094] The purpose of projecting into a depth map is to reduce the influence of factors such as point cloud distribution intensity and random noise. While preserving the original point cloud image geometric features and point cloud distribution features, the depth image converts the original 3D point cloud data into image information, which facilitates the subsequent use of neural networks for image matching.

[0095] S103, using a deep neural network to compress the depth map of each superpoint into a 5×2 pixel image, which can be represented as a 5×2 feature matrix.

[0096] In this embodiment, the deep neural network adopts a DAE neural network to improve the compression speed and compression quality of the image.

[0097] The DAE neural network architecture consists of two stages: encoding and decoding. The encoding stage starts from the input layer and then connects to the hidden layer, gradually decreasing until the desired image size is reached. The decoding stage starts with a compact representation of the data, and the dimension of each subsequent hidden layer increases at a time until the output layer dimension reaches the input dimension. The loss function is defined as the pixel error between the input layer and the output layer. The optimization network achieves the best compact representation of the image. A hidden layer with 4 full connections is used. A sigmoid nonlinear activation function is used between each layer of the network and the input dropout. Finally, each depth image is compressed into a 5×2 pixel image for registration. Figure 3 It is a neural network model, which is divided into an input layer, an output layer and four intermediate layers (hidden layers); Figure 4 Shows the visualization of the depth image during compression.

[0098] S104, matching candidate point selection and rough registration:

[0099] Calculate the Euclidean distance between each feature matrix of the source point cloud and the nearest N feature matrices in the target point cloud. If the Euclidean distance of the latter feature matrix in the source point cloud suddenly increases compared with the Euclidean distance of the previous feature matrix, the super point corresponding to the previous feature matrix is ​​taken as a candidate point, and the super point corresponding to the latter feature matrix is ​​an invalid point.

[0100] A random sampling consensus algorithm is used to iteratively find at least three point cloud registration matrices between candidate points and superpoints of the target point cloud.

[0101] For example, if the Euclidean distance of the i-th feature matrix suddenly increases compared to the Euclidean distance of the i-1-th feature matrix, then the super point corresponding to the i-1-th feature matrix is ​​taken as a candidate point, and the super point corresponding to the i-th feature matrix is ​​regarded as an invalid point.

[0102] In this embodiment, N = 3. Analyzing the experimental data, for a source point cloud containing 10 million points and a target point cloud containing 500,000 points, the latter can be segmented into about 200 super points. When N = 3, there are about 550 candidate matching objects.

[0103] A random sampling consensus algorithm is used to iteratively find at least three point cloud registration matrices between candidate points and superpoints of the target point cloud.

[0104] The random sampling consensus algorithm (RANSAC) can provide an initial transformation matrix for point cloud precision registration and delete incorrect registration pairs. The basic idea of ​​the algorithm is to first randomly sample samples, then estimate the parameters of the sampled samples, and use the estimated parameters to divide all data into inliers (sample points that meet the estimated parameters) and outliers (sample points that do not meet the estimated parameters). Finally, the largest set of inliers is used to reversely recalculate the model parameters. Assume that the best initial registration matrix M is found from the target point cloud through iteration:

[0105]

[0106] Among them, R is the rotation matrix and T is the translation matrix.

[0107] In order to find the transformation relationship between the six degrees of freedom between point clouds (i.e., up and down (UP, DOWN), front and back (FORWARD, BACK), left and right (RIGHT, LEFT) translation and pitch (PITCH), roll (ROLL), and yaw (YAW), see Figure 5 ), at least 3 matching pairs are required. In order to enhance the robustness of the algorithm, this embodiment adopts 6 matching pairs.

[0108] S105: Using the above point cloud registration matrix as an initial transformation matrix, perform precise registration through an ICP algorithm to obtain an optimal transformation matrix.

[0109] The iterative closest point (ICP) algorithm consists of three stages: 1) obtaining the registration point pairs; 2) registering and estimating the optimal initial transformation matrix; 3) iterative registration to obtain the optimal transformation matrix.

[0110] Compared with the traditional method that uses a single point cloud data as the matching benchmark, the present invention uses a "super point" containing several point cloud data as the basic processing unit in the matching process, thereby improving the registration speed. In addition, unlike the traditional method that directly searches for the point with the most similar geometric features as the matching point, the present invention uses the similarity between the nearest neighbor features of the data after dimensionality reduction processing, and obtains the corresponding matching pairs based on the similarity, and then obtains multiple optimal transformation matrices based on the matching pairs to obtain a rough matching result. Finally, the ICP algorithm is used for precise matching, and the optimal point cloud registration matrix is ​​obtained through optimal evaluation. The present invention not only improves the problem that the ICP algorithm is difficult to start with when facing an unordered point cloud, but also improves the traditional point cloud registration algorithm in terms of accuracy and efficiency.

[0111] However, those skilled in the art should recognize that the above embodiments are only used to illustrate the present invention, and are not intended to limit the present invention. As long as they are within the spirit of the present invention, any changes or modifications to the above embodiments will fall within the scope of the claims of the present invention.

Claims

1. A point cloud registration method for unmanned driving assisted positioning, characterized in that: include: 1) The source point cloud and the target point cloud to be matched are divided into multiple super points through the following iterative process: Randomly select a point P that does not belong to any superpoint; define a new superpoint as a point with a fixed radius R centered at the point P. sphere A set of points in a sphere, where V local is the volume of the sphere containing the point cloud, m is the number of matching pairs used for matching, and m is at least 3; 2) Project each superpoint into a depth map; 3) Using a deep neural network to compress the depth map of each superpoint into a 5×2 feature matrix; 4) Matching candidate point selection and rough registration: Calculate the Euclidean distance between each feature matrix of the source point cloud and the nearest N feature matrices in the target point cloud. If the Euclidean distance of the latter feature matrix in the source point cloud suddenly increases compared with the Euclidean distance of the former feature matrix, then take the super point corresponding to the former feature matrix as the candidate point, and the super point corresponding to the latter feature matrix as the invalid point. Using a random sampling consensus algorithm to iteratively find at least three point cloud registration matrices between the candidate points and super points of the target point cloud; 5) Using the point cloud registration matrix as the initial transformation matrix, perform precise registration using the ICP algorithm to obtain the optimal transformation matrix.

2. The point cloud registration method for unmanned driving assisted positioning according to claim 1, characterized in that: The m is 5.

3. A point cloud registration method for unmanned driving assisted positioning according to claim 1 or 2, characterized in that: The projecting each superpoint into a depth map further comprises: a. Establish a normalized local coordinate system: set the centroid of each superpoint as the origin of the coordinate system, calculate the covariance matrix of each superpoint, and use singular value decomposition to obtain the corresponding x-axis, y-axis, and z-axis of the coordinate system; b. Project each superpoint into the two-dimensional image along the x-axis and y-axis directions, use the height of the z-axis to represent the color of the depth map, and cut the circular boundary to obtain a square two-dimensional image.

4. The point cloud registration method for unmanned driving assisted positioning according to claim 3, characterized in that: The size of the two-dimensional image is 64×64 pixels, and the size of the two-dimensional picture is 32×32 pixels.

5. The point cloud registration method for unmanned driving assisted positioning according to claim 4, characterized in that: The deep neural network adopts DAE neural network.

6. A point cloud registration system for unmanned driving assisted positioning, characterized in that: A storage module is included, wherein the storage module includes a plurality of instructions loaded and executed by a processor: 1) The source point cloud and the target point cloud to be matched are divided into multiple super points through the following iterative process: Randomly select a point P that does not belong to any superpoint; define a new superpoint as a point with a fixed radius R centered at the point P. sphere A set of points in a sphere, where V local is the volume of the sphere containing the point cloud, m is the number of matching pairs used for matching, and m is at least 3; 2) Project each superpoint into a depth map; 3) Using a deep neural network to compress the depth map of each superpoint into a 5×2 feature matrix; 4) Matching candidate point selection and rough registration: Calculate the Euclidean distance between each feature matrix of the source point cloud and the nearest N feature matrices in the target point cloud. If the Euclidean distance of the latter feature matrix in the source point cloud suddenly increases compared with the Euclidean distance of the former feature matrix, then take the super point corresponding to the former feature matrix as the candidate point, and the super point corresponding to the latter feature matrix as the invalid point. Using a random sampling consensus algorithm to iteratively find at least three point cloud registration matrices between the candidate points and super points of the target point cloud; 5) Using the point cloud registration matrix as the initial transformation matrix, perform precise registration using the ICP algorithm to obtain the optimal transformation matrix.

7. The point cloud registration system for unmanned driving assisted positioning according to claim 6, characterized in that: The m is 5.

8. A point cloud registration system for unmanned driving assisted positioning according to claim 6 or 7, characterized in that: The projecting each superpoint into a depth map further comprises: a. Establish a normalized local coordinate system: set the centroid of each superpoint as the origin of the coordinate system, calculate the covariance matrix of each superpoint, and use singular value decomposition to obtain the corresponding x-axis, y-axis, and z-axis of the coordinate system; b. Project each superpoint into the two-dimensional image along the x-axis and y-axis directions, use the height of the z-axis to represent the color of the depth map, and cut the circular boundary to obtain a square two-dimensional image.

9. The point cloud registration system for unmanned driving assisted positioning according to claim 8, characterized in that: The size of the two-dimensional image is 64×64 pixels, and the size of the two-dimensional picture is 32×32 pixels.

10. The point cloud registration system for unmanned driving assisted positioning according to claim 9, characterized in that: The deep neural network adopts DAE neural network.

Images