Point cloud descriptor matching method and system based on characteristics of included angles of complex vectors

Abstract

A point cloud descriptor matching method and system based on characteristics of included angles of complex vectors. The method comprises: performing dimension reduction on existing complex descriptor structures, and constructing derivatives of complex descriptors; using a database to perform quick retrieval and derivative filtering on constructed derivatives of the complex descriptors; and by using the characteristics of various included angles of complex vectors, performing similarity comparison on the complex descriptors acquired after quick retrieval and derivative filtering, so as to obtain complex descriptors with set precision. Possible complex descriptors are quickly retrieved and filtered out by means of constructed derivatives, and finally, precise matching of the same position is then completed by using the characteristics of various included angles of complex vectors. Therefore, the features of original complex descriptors are taken into consideration, and derivatives corresponding to different features are acquired, thereby improving the usage efficiency of original complex descriptors.

Description

A point cloud descriptor matching method and system based on complex vector angle characteristics

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to Chinese patent application number 202311438446.8, filed with the China Patent Office on October 31, 2023, entitled “A Point Cloud Descriptor Matching Method and System Based on Complex Vector Angle Characteristics,” the entire contents of which are incorporated by reference into this application. Technical Field

[0003] The present invention belongs to the field of high-precision positioning for autonomous driving, and specifically relates to a point cloud descriptor matching method and system based on the angle characteristics of complex vectors. Background Art

[0004] Autonomous driving, or advanced driver assistance systems, play a crucial role in current and future transportation. By autonomously sensing the driving environment and making decisions to control safe driving, they not only effectively free the hands of human drivers but also achieve high traffic efficiency through information exchange and data sharing between vehicles. Position information is a crucial component of autonomous driving, and accurate position information provides a secure foundation for decision-making and control. In complex urban environments, due to multipath effects, position information obtained solely by GNSS can experience significant offsets, leading to vehicle positioning failures. Currently, most positioning solutions use high-precision maps. Accurate vehicle position information can be obtained by matching vehicle sensors with pre-loaded maps. However, building high-precision maps is complex, and matching real-time data with map information consumes significant computing resources.

[0005] To address the time and cost of map construction, existing solutions leverage crowdsourcing, dividing a map into multiple sub-areas and utilizing multiple mapping devices to complete the map construction. Assembling the maps of each sub-area requires map stitching and location recognition. Existing location recognition methods improve detection efficiency by mapping key information from each measurement into a compact descriptor. Furthermore, to achieve stronger environmental representation, some methods utilize complex numbers to store more measurement results, forming cassava descriptors. Identification of the same location is achieved through descriptor matching. Existing methods often directly brute-force descriptor matching based on similarity, which not only fails to achieve satisfactory recall but also consumes significant computational resources. When faced with large-scale map construction or vehicle positioning in large scenarios, traditional location recognition can lead to a series of problems, such as detection failures or false detections. This makes it impossible to meet the requirements of map construction and vehicle alignment, ultimately hindering the proper operation of autonomous or assisted driving. Summary of the Invention

[0006] The purpose of the present invention is to provide a point cloud descriptor matching method and system based on the complex vector angle characteristics, so as to overcome the problem that the existing technology obtains descriptors to construct maps, which may lead to detection failure or erroneous detection.

[0007] A point cloud descriptor matching method based on complex vector angle characteristics, characterized by comprising the following steps:

[0008] S1, reduce the dimension of the existing complex descriptor structure and construct the derivative of the complex descriptor;

[0009] S2, using the database to quickly search and filter derivatives of the constructed complex descriptors;

[0010] S3, using multiple angle characteristics of complex vectors to perform similarity comparison on the complex descriptors obtained by fast retrieval and filtering derivatives, to obtain a complex descriptor with a set accuracy.

[0011] Preferably, the existing complex descriptor structure is reduced in terms of columns and space.

[0012] Preferably, the existing complex descriptor structure is reduced in dimension from the column to obtain a fast retrieval derivative with rotation invariance.

[0013] Preferably, for the determined complex point cloud descriptor C ij =R ij +I ij (i≤Nr,j≤Ns) is a 2D complex matrix with Nr rows and Ns columns. It corresponds to the multiple curved trapezoids after the radial and circumferential division of the lidar. The real and imaginary parts store the two main characteristics of these curved trapezoids respectively. The size of the modulus of each complex unit is calculated as follows:

[0014] Among them, M ij It is a real number matrix with the same structure and size as the complex descriptor, which becomes the module matrix and stores the module size of the original complex matrix; it degenerates the original 2D matrix into an Nr-dimensional column vector K = [k1, k2, ..., k Nr ], thereby further compressing the original feature information:

[0015] By calculating the average value of each row vector of the modular matrix and using this average value to reflect the original row information, the original 2-dimensional matrix is ​​compressed into a 1-dimensional vector.

[0016] Preferably, feature dimensionality reduction is performed on the existing complex descriptor structure in space to obtain derivatives with feature distribution characteristics.

[0017] Preferably, the following formula is used to calculate the logical derivative to reflect the feature distribution of the original descriptor:

[0018] The derivative is a logical matrix with exactly the same structure and size as the original complex descriptor, which is used to record the characteristic distribution of the original descriptor by calculating the characteristic modulus of each part.

[0019] Preferably, the rotationally invariant derivatives are used to quickly retrieve complex descriptors that meet a set threshold, and derivatives with the same feature distribution are used to quickly filter results with low feature similarity among the complex descriptors that meet the set threshold, and the optimal complex descriptor is retained.

[0020] A point cloud descriptor matching system based on complex vector angle characteristics includes a derivative construction module, a retrieval and filtering module, and a matching module:

[0021] The derivative construction module reduces the dimensionality of the existing complex descriptor structure and constructs the derivative of the complex descriptor;

[0022] A retrieval and filtering module, which uses a database to quickly retrieve and filter derivatives of the constructed complex descriptors;

[0023] The matching module uses multiple angle characteristics of complex vectors to perform similarity comparison on the complex descriptors obtained by fast retrieval and filtering derivatives to obtain complex descriptors with a set accuracy.

[0024] Preferably, the existing complex descriptor structure is reduced in dimension from the column and space perspectives respectively; the existing complex descriptor structure is reduced in dimension from the column perspective to obtain a fast retrieval derivative with rotation invariance.

[0025] Preferably, for the determined complex point cloud descriptor C ij =R ij +I ij (i≤Nr,j≤Ns) is a 2D complex matrix with Nr rows and Ns columns. It corresponds to the multiple curved trapezoids after the radial and circumferential division of the lidar. The real and imaginary parts store the two main characteristics of these curved trapezoids respectively. The size of the modulus of each complex unit is calculated as follows:

[0026] Among them, M ij It is a real number matrix with the same structure and size as the complex descriptor, which becomes the module matrix and stores the module size of the original complex matrix; it degenerates the original 2D matrix into an Nr-dimensional column vector K = [k1, k2, ..., k Nr ], thereby further compressing the original feature information:

[0027] By calculating the average value of each row vector of the modular matrix and using this average value to reflect the original row information, the original 2-dimensional matrix is ​​compressed into a 1-dimensional vector.

[0028] Compared with the prior art, the present invention has the following beneficial technical effects:

[0029] The present invention discloses a point cloud descriptor matching method based on the angular characteristics of complex vectors. The method reduces the dimensionality of an existing complex descriptor structure, constructs derivatives of the complex descriptor, uses a database to quickly retrieve and filter the derivatives of the constructed complex descriptor, uses multiple angular characteristics of the complex vectors to perform similarity comparison on the complex descriptors obtained by the rapid retrieval and filtering of the derivatives, and obtains complex descriptors of a set accuracy. With the help of the constructed derivatives, possible complex descriptors are quickly retrieved and filtered out. Finally, the characteristics of the multiple angular characteristics of the complex vectors are used to complete accurate matching of the same position. The characteristics of the original complex descriptor are taken into consideration to obtain derivatives corresponding to different features, thereby improving the utilization efficiency of the original complex descriptor.

[0030] Preferably, the present invention takes into account the inefficiency of traditional brute force search and uses the derivatives of descriptors to quickly retrieve and filter out possible candidates from the candidate data in a hierarchical manner.

[0031] Preferably, the present invention utilizes the different characteristics of various angles of complex vectors to enhance the information utilization of descriptors and improve the accuracy of position recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] FIG1 is a flow chart of a point cloud descriptor matching method based on complex vector angle characteristics in an embodiment of the present invention.

[0033] FIG2 is a schematic diagram of a complex search derivative according to an embodiment of the present invention.

[0034] FIG3 is a schematic diagram of a characteristic distribution derivative according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described 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 making creative efforts should fall within the scope of protection of the present invention.

[0036] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0037] As shown in Figure 1, the present invention provides a point cloud descriptor matching method based on the angle characteristics of complex vectors. By constructing derivatives of the complex descriptors, different characteristics of the original descriptors are obtained. The constructed derivatives are used to quickly search and filter possible candidates. Finally, the characteristics of multiple angles of the complex vectors are used to achieve accurate matching of the same location. The present invention can not only improve the efficiency of location recognition, but also improve the accuracy of location recognition. Specifically, it includes the following steps:

[0038] S1, reduce the dimension of the existing complex descriptor structure and construct the derivative of the complex descriptor;

[0039] S2, using the database to quickly search and filter derivatives of the constructed complex descriptors;

[0040] S3, using multiple angle characteristics of complex vectors to perform similarity comparison on the complex descriptors obtained by fast retrieval and filtering derivatives, to obtain a complex descriptor with a set accuracy.

[0041] The specific steps include:

[0042] The existing complex descriptor structure is reduced in terms of column and space dimensions.

[0043] S1-1: Reduce the dimensionality of the existing complex descriptor structure from the column perspective to obtain fast retrieval derivatives with rotation invariance, as shown in Figure 2.

[0044] S1-2: Perform spatial dimension reduction on the existing complex descriptor structure to obtain derivatives with characteristic distribution characteristics, as shown in Figure 3.

[0045] Although the descriptor has extracted key information from the original environmental measurements, alleviating the problems of large amounts of original measurement data and difficulty in preservation, there are still problems such as low efficiency when searching for locations. The present invention proceeds from two aspects and performs feature extraction of different dimensions on complex descriptors to obtain fast retrieval keys with rotation invariance and derivatives with characteristic distribution characteristics. Specifically, the modulus of the complex descriptor is first calculated, and then the average calculation is performed on each row to obtain the average value of each row. Secondly, the modulus is mapped to a logical matrix of the same size as the original descriptor using logical mapping to reflect its characteristic distribution.

[0046] The specific process is as follows:

[0047] S1-1 Get fast retrieval derivatives with rotation invariance:

[0048] For the complex point cloud descriptor C that has been determined ij =R ij +I ij (i≤Nr,j≤Ns) is a 2D complex matrix with Nr rows and Ns columns. It corresponds to the multiple curved trapezoids after the radial and circumferential division of the lidar. The real and imaginary parts respectively store the two main characteristics of these curved trapezoids; for each complex unit, the magnitude of its modulus is calculated:

[0049] Among them, M ij It is a real number matrix with the same structure and size as the complex descriptor. It becomes a modular matrix that stores the modular size of the original complex matrix and can reflect the strength of its main features. Then the original 2D matrix is ​​degenerated into an Nr-dimensional column vector K = [k1, k2, ..., k Nr ], thereby further compressing the original feature information:

[0050] By calculating the average value of each row vector of the modular matrix and using this average value to reflect the original row information, the original 2-dimensional matrix is ​​compressed into a 1-dimensional vector. Since the original complex matrix row stores the information of the measurement point around the circle, the average of each row information can achieve rotational invariance of the heading.

[0051] S1-2 Obtain derivatives with characteristic distribution properties:

[0052] The existing complex descriptor is converted into a column vector. The present invention uses the following formula to calculate the logical derivative to reflect the feature distribution of the original descriptor:

[0053] The derivative is a logical matrix with exactly the same structure and size as the original complex descriptor, which is used to record the characteristic distribution of the original descriptor by calculating the characteristic modulus of each part.

[0054] The derivatives of the constructed complex descriptor are quickly retrieved and filtered using a database, and the complex descriptor having the derivatives quickly retrieved and filtered is obtained, that is, a possible candidate.

[0055] Through step S1, derivatives of the original descriptor are constructed from different aspects, and then these derivatives are used to efficiently retrieve and filter out the most likely candidates from the dataset to be judged:

[0056] S2-1 uses rotation-invariant derivatives to quickly retrieve multiple possible candidates, that is, to obtain complex descriptors that meet the set threshold through rapid retrieval:

[0057] The present invention constructs the descriptors to be searched as a column vector K with rotation invariance, constructs a KD-Tree as the search object, and for each candidate to be searched, selects the nearest N candidates as the initial candidates based on the KD-Tree:

[0058] Φ c =KD-Tree(K current ,N) (4)

[0059] where Φ c With the current K current The set of the N closest candidates.

[0060] S2-2 quickly filters out wrong candidates with the help of logical derivatives:

[0061] Although the initial candidates obtained by S2-1 have some similarity with the current measurement value, directly using the corresponding descriptors for accurate matching will undoubtedly waste a lot of time. The present invention uses derivatives with the same feature distribution to quickly filter out the results with low feature similarity among these initial candidates, retaining the best candidates, thereby reducing the pressure of subsequent accurate matching and improving the efficiency of position recognition as a whole:

[0062] in They represent the kth point cloud descriptor in the candidate set and the current point cloud descriptor respectively.

[0063] An exclusive-or operation is performed on the logical descriptors corresponding to the current descriptor and the initial candidate to calculate the point cloud candidate that is most similar to the current measured feature distribution. However, in practice, due to the possibility of different headings during revisits, errors may occur in the similarity calculation. The candidate is column-shifted. Since the original complex descriptor is divided according to radial and circumferential directions, its different columns represent different headings. The present invention achieves the effect of rotating the heading angle by shifting the columns, and obtains the optimal matching result by rotating the heading angle:

[0064] Where shift is the heading difference between the two scans.

[0065] S3 uses the various angle characteristics of complex vectors for precise matching:

[0066] The vector angle is used to evaluate the distance between two vectors, but when faced with a matrix using complex descriptors, the distance cannot be solved using the vector distance in the real field. This method proposes a precise matching method for complex descriptors by analyzing the characteristics of different attached angles. For the vector angle in the real space, the present invention uses the inner product of the vector divided by the product of the corresponding modules to calculate, and evaluates the distance or similarity between the two vectors by the calculated angle. There are real vectors a = (a1, a2, ..., an) and b = (b1, b2, ..., bn) (where ai, bi∈R), then the angle between a and b can be expressed as:

[0067] Where (a,b) R Represents the inner product of two vectors a and b in real space. The specific calculation is as follows:

[0068] Assume there exists a complex vector:

[0069] c=(c1,c2,...,c n )

[0070] d=(d1,d2,...,d n ) (11)

[0071] The same operation as in the real number field can be obtained:

[0072] Will It is called the complex angle between complex vectors a and b; the inner product of complex vectors a and b in complex space is:

[0073] The cosine value of this complex angle is actually a complex number, which can be expanded using Euler's formula:

[0074] Where ρ is the modulus of the complex cosine value of the complex angle, and φ is the corresponding argument. When ρ≤1:

[0075] Θ H (a, b) is called the Hermitian angle between the complex vectors a and b, and φ = φ(a, b) is called the pseudo angle between the complex vectors a and b. C and possible point cloud candidate descriptors C H , use the following formula to calculate their real part, imaginary part, and total distance respectively. Comprehensively judge the similarity of complex descriptors.

[0076] Total score:

[0077] Real part score:

[0078] Imaginary part score:

[0079] The present invention obtains the different characteristics of the original descriptor by constructing derivatives of the complex descriptor. Secondly, the constructed derivatives are used to quickly retrieve and filter out possible candidates. Finally, the characteristics of the multiple angles of the complex vector are used to complete the precise matching of the same position. The present invention takes into account the characteristics of the original complex descriptor, obtains derivatives corresponding to different features, and improves the efficiency of the use of the original complex descriptor; considering the inefficiency of traditional brute force search, the derivatives of the descriptor are used to quickly retrieve and filter out possible candidates from the candidate data in a hierarchical manner. With the help of the different characteristics of the multiple angles of the complex vector, the information utilization of the descriptor is improved and the accuracy of position recognition is improved.

Claims

1. A point cloud descriptor matching method based on complex vector angle characteristics, characterized in that: The following steps are involved: S1, reduce the dimension of the existing complex descriptor structure and construct the derivative of the complex descriptor; S2, using the database to quickly retrieve and filter derivatives of the constructed complex descriptors; S3, using multiple angle characteristics of complex vectors to compare the similarity of complex descriptors obtained by fast retrieval and filtering derivatives, to obtain complex descriptors with set accuracy.

2. The point cloud descriptor matching method based on complex vector angle characteristics according to claim 1, characterized in that: The existing complex descriptor structure is reduced in terms of column and space dimensions.

3. The point cloud descriptor matching method based on complex vector angle characteristics according to claim 2, characterized in that: The existing complex descriptor structure is reduced in terms of columns to obtain fast retrieval derivatives that are rotationally invariant.

4. The point cloud descriptor matching method based on complex vector angle characteristics according to claim 3 is characterized in that: For the complex point cloud descriptor C that has been determined ij =R ij +I ij (i≤Nr,j≤Ns) is a 2D complex matrix with Nr rows and Ns columns, which corresponds to the multiple curved trapezoids after radial and annular division of the lidar. The real and imaginary parts store the two main features of these curved trapezoids respectively. The size of the modulus is calculated for each complex unit: Among them, M ij It is a real number matrix with the same structure and size as the complex descriptor, which becomes a modular matrix and stores the modular size of the original complex matrix; it degenerates the original 2D matrix into an Nr-dimensional column vector K = [k1, k2, ..., k Nr ], thereby further compressing the original feature information: By calculating the average value of each row vector of the module matrix, the average value is used to reflect the original row information, thereby compressing the original 2-dimensional matrix into a 1-dimensional vector.

5. The point cloud descriptor matching method based on complex vector angle characteristics according to claim 2, characterized in that: The existing complex descriptor structure is spatially dimensionally reduced to obtain derivatives with characteristic distribution characteristics.

6. The point cloud descriptor matching method based on complex vector angle characteristics according to claim 5, characterized in that: The logical derivative is calculated using the following formula to reflect the feature distribution of the original descriptor: The derivative is a logical matrix with exactly the same structure and size as the original complex descriptor, and is used to record the feature distribution of the original descriptor by calculating the feature modulus of each part.

7. The point cloud descriptor matching method based on complex vector angle characteristics according to claim 1, characterized in that: In addition, the rotation-invariant derivatives are used to quickly retrieve complex descriptors that meet the set threshold, and the derivatives with the same feature distribution are used to quickly filter the results with low feature similarity among the complex descriptors that meet the set threshold, and the optimal complex descriptors are retained.

8. A point cloud descriptor matching system based on complex vector angle characteristics, characterized in that: Includes derivative building modules, search filtering modules and matching modules: The derivative construction module reduces the dimension of the existing complex descriptor structure and constructs the derivatives of the complex descriptor; A retrieval and filtering module, which uses a database to quickly retrieve and filter derivatives of the constructed complex descriptors; The matching module uses multiple angle characteristics of complex vectors to compare the similarity of complex descriptors obtained by fast retrieval and filtering derivatives to obtain complex descriptors with a set accuracy.

9. The point cloud descriptor matching system based on complex vector angle characteristics according to claim 8, characterized in that: The existing complex descriptor structure is dimensionally reduced in terms of columns and space respectively; the existing complex descriptor structure is dimensionally reduced in terms of columns to obtain fast retrieval derivatives with rotation invariance.

10. The point cloud descriptor matching system based on complex vector angle characteristics according to claim 9, characterized in that: For the complex point cloud descriptor C that has been determined ij =R ij +I ij (i≤Nr,j≤Ns) is a 2D complex matrix with Nr rows and Ns columns, which corresponds to the multiple curved trapezoids after radial and annular division of the lidar. The real and imaginary parts store the two main features of these curved trapezoids respectively. The size of the modulus is calculated for each complex unit: Among them, M ij It is a real number matrix with the same structure and size as the complex descriptor, which becomes a modular matrix and stores the modular size of the original complex matrix; it degenerates the original 2D matrix into an Nr-dimensional column vector K = [k1, k2, ..., k Nr ], thereby further compressing the original feature information: By calculating the average value of each row vector of the module matrix, the average value is used to reflect the original row information, thereby compressing the original 2-dimensional matrix into a 1-dimensional vector.

Images