A point cloud segmentation method and device

Abstract

The application discloses a point cloud segmentation method and device, and the method comprises the following steps: acquiring laser point cloud data, wherein the laser point cloud data comprises three-dimensional point cloud data; using a point cloud segmentation network to perform coarse-grained prediction prior on the laser point cloud data to obtain a coarse-grained prior result; reducing the dimension of the coarse-grained prior result to convert the coarse-grained prior result into one-dimensional reduced dimension point cloud data; extracting object local point cloud data from the reduced dimension point cloud data; and performing clustering segmentation on the object local point cloud data to obtain a point cloud segmentation object. The above implementation scheme converts the coarse-grained prior result into one-dimensional reduced dimension point cloud data, which can greatly reduce the subsequent data calculation amount and improve the data processing speed on the premise of retaining the three-dimensional data characteristics; subsequently, the object local data is quickly extracted by using the reduced dimension point cloud data, and the object local data is subjected to clustering segmentation, so that the segmentation accuracy of the laser point cloud object is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to laser point cloud technology, more particularly, to a point cloud segmentation method and device. BACKGROUND

[0002] With the wide application of 3D sensors such as laser radar in the fields of unmanned driving, high-precision map construction, three-dimensional reconstruction, etc., object segmentation using laser point cloud has become a research focus. The point cloud data obtained by laser radar scanning technology has the characteristics of mass, disorder and non-structure, which makes the laser point cloud object segmentation a research difficulty.

[0003] In the prior art, a point cloud index is established in advance to record the spatial information of the collected original laser point cloud data. This method needs to know the content of the point cloud in advance and record the spatial information of each point, and the implementation process is complicated and cannot be applied to real-time semantic segmentation object scenarios. SUMMARY

[0004] Therefore, the present application provides the following technical solutions:

[0005] A point cloud segmentation method comprises:

[0006] Obtaining laser point cloud data, the laser point cloud data comprising three-dimensional point cloud data;

[0007] Using a point cloud segmentation network to perform coarse-grained prediction prior on the laser point cloud data to obtain coarse-grained prior results, wherein the coarse-grained prior results comprise four-dimensional information;

[0008] Reducing the dimension of the coarse-grained prior results to obtain one-dimensional reduced dimension point cloud data;

[0009] Extracting object local point cloud data from the reduced dimension point cloud data;

[0010] Performing clustering segmentation on the object local point cloud data to obtain point cloud segmentation objects.

[0011] Optionally, the three-dimensional point cloud data of the laser point cloud data comprises an [N, 3] format data matrix of N point clouds, wherein N represents the number of point cloud points, and 3 represents the point cloud coordinate values of the point cloud points in three spatial dimensions.

[0012] Using a point cloud segmentation network to perform coarse-grained prediction prior on the laser point cloud data to obtain coarse-grained prior results, comprising:

[0013] Classify each point cloud point in the data matrix through a laser point cloud segmentation network to obtain a coarse-grained prior result in the format of [N, 4], wherein 4 represents point cloud coordinate values of the point cloud point in three spatial dimensions and the category of the point cloud point.

[0014] Optionally, the classification of each point cloud point in the data matrix through the laser point cloud segmentation network to obtain the coarse-grained prior result in the format of [N, 4] comprises:

[0015] Input the point cloud data in the format of [N, 3] into the laser point cloud segmentation network.

[0016] The laser point cloud segmentation network classifies each point cloud point in the received point cloud data in a single laser point category and outputs the coarse-grained prior result in the format of [N, 4].

[0017] Optionally, the dimension reduction of the coarse-grained prior result into one-dimensional reduced point cloud data comprises:

[0018] Based on the interval length resolution of each dimension set, interval positioning of each point cloud point in the coarse-grained prior result in the format of [N, 4] is performed in three spatial dimensions to obtain three interval resolutions corresponding to the three spatial dimensions.

[0019] The three interval resolutions are alternately interpolated and combined in dimension reduction to obtain one-dimensional reduced point cloud data of three-dimensional point cloud data.

[0020] Optionally, every 3-bit value in the one-dimensional reduced point cloud data represents a local rectangular body in the point cloud space.

[0021] Optionally, after the alternately interpolated and combined three interval resolutions are dimensionally reduced to obtain one-dimensional reduced point cloud data of three-dimensional point cloud data, the method further comprises:

[0022] The one-dimensional reduced point cloud data is fused into the coarse-grained prior result in the format of [N, 4] to obtain data in the format of [N, 5], wherein 5 represents point cloud coordinate values of the point cloud point in three spatial dimensions, one-dimensional reduced point cloud data, and the category of the point cloud point.

[0023] Optionally, the interval positioning of each point cloud point in the coarse-grained prior result in the format of [N, 4] in three spatial dimensions based on the interval length resolution of each dimension set to obtain three interval resolutions corresponding to the three spatial dimensions comprises:

[0024] In a parent interval represented by an effective boundary distance value of each spatial dimension, it is determined whether the point cloud point is located in a front half interval or a rear half interval, and if in the front half interval, a first identifier is marked, and if in the rear half interval, a second identifier is marked.

[0025] The front half interval or the rear half interval where the point cloud point is located is taken as the latest parent interval, and the above step of determining whether the point cloud point is located in the front half interval or the rear half interval is repeated until the number of identification bits obtained is the same as the number of identification bits corresponding to the interval length resolution.

[0026] Optionally, the extracting the object local point cloud data from the reduced dimension point cloud data comprises:

[0027] According to the point cloud point category in the data in the [N, 5] format, all point cloud data with the same front bit number are extracted from the one-dimensional reduced dimension point cloud data, and all point cloud values in a space region represented by the three-dimensional point cloud data are obtained.

[0028] The object local point cloud data is obtained according to the all point cloud values, wherein the front bit number includes the first three bit numbers of the one-dimensional reduced dimension point cloud data.

[0029] Optionally, the clustering and segmenting the object local point cloud data to obtain the point cloud segmented object comprises:

[0030] The object local point cloud data is clustered and calculated by using a K-means clustering algorithm K-means to obtain clustered point cloud data.

[0031] The clustered point cloud data is matched with an object category based on the category of the point cloud point in the data in the [N, 5] format to obtain the point cloud segmented object.

[0032] A point cloud segmentation device comprises:

[0033] A data acquisition module is configured to acquire laser point cloud data, wherein the laser point cloud data comprises three-dimensional point cloud data.

[0034] A network prior module is configured to use a point cloud segmentation network to perform coarse-grained prediction prior on the laser point cloud data to obtain a coarse-grained prior result, wherein the coarse-grained prior result comprises four-dimensional information.

[0035] A data reduction module is configured to reduce and convert the coarse-grained prior result into one-dimensional reduced dimension point cloud data.

[0036] A local extraction module is configured to extract object local point cloud data from the reduced dimension point cloud data.

[0037] A clustering and segmentation module is configured to cluster and segment the object local point cloud data to obtain a point cloud segmented object.

[0038] As can be seen from the above technical solution, compared with the prior art, the embodiments of the present invention disclose a point cloud segmentation method and apparatus. The method includes: acquiring laser point cloud data, wherein the laser point cloud data includes three-dimensional point cloud data; using a point cloud segmentation network to perform coarse-grained prediction prior on the laser point cloud data to obtain a coarse-grained prior result, wherein the coarse-grained prior result includes four-dimensional information; reducing the dimensionality of the coarse-grained prior result into one-dimensional reduced point cloud data; extracting local point cloud data of an object from the reduced point cloud data; and performing clustering segmentation on the local point cloud data of the object to obtain point cloud segmented objects. The above implementation scheme reduces the dimensionality of the coarse-grained prior result into one-dimensional reduced point cloud data, which can greatly reduce the amount of subsequent data calculation and improve the data processing speed while retaining the three-dimensional data features; subsequently, the local data of the object is quickly extracted using the reduced point cloud data, and the local data of the object is clustered and segmented to improve the segmentation accuracy of the laser point cloud object. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0040] Figure 1 This is a flowchart of a point cloud segmentation method disclosed in an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the range of three-dimensional point cloud data disclosed in an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram illustrating the principle of the point cloud segmentation network disclosed in an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram of the network structure of the PointNet network disclosed in an embodiment of the present invention;

[0044] Figure 5 This is a flowchart of a data dimensionality reduction process disclosed in an embodiment of the present invention;

[0045] Figure 6 This is a schematic diagram illustrating the dimensionality reduction processing principle disclosed in an embodiment of the present invention;

[0046] Figure 7A This is a schematic diagram illustrating the principle of dimensionality reduction of three-dimensional data into one-dimensional data as disclosed in an embodiment of the present invention;

[0047] Figure 7BA process principle schematic diagram before 3M data is fused to a label dimension disclosed by the embodiment of the present application is disclosed.

[0048] Figure 8 A schematic diagram of a local object point cloud data quickly extracted disclosed by the embodiment of the present application is disclosed.

[0049] Figure 9 A clustering segmentation process schematic diagram disclosed by the embodiment of the present application is disclosed.

[0050] Figure 10 A structure schematic diagram of a point cloud segmentation device disclosed by the embodiment of the present application is disclosed. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0052] The point cloud segmentation method described in the embodiment of the present application relates to a laser point cloud semantic segmentation technology. The laser point cloud semantic segmentation is a point-by-point classification of each object in a point cloud scene to achieve the purpose of segmentation. Figure 1 A flowchart of a point cloud segmentation method disclosed by the embodiment of the present application is disclosed, referring to Figure 1 As shown in the figure, the point cloud segmentation method can include:

[0053] Step 101: acquiring laser point cloud data, the laser point cloud data including three-dimensional point cloud data.

[0054] To perform point cloud object segmentation, laser point cloud data needs to be acquired first. The implementation is to use a laser radar to collect point cloud data, and record the effective boundary limit values X, Y and Z of the point cloud data, wherein X is the effective boundary interval value of the forward direction of the collection radar; Y is the effective boundary interval value of the transverse direction perpendicular to the forward direction of the collection radar; and Z is the effective boundary interval value of the vertical direction perpendicular to the forward direction of the collection radar. Figure 2 A three-dimensional point cloud data range schematic diagram disclosed by the embodiment of the present application is disclosed, which can be combined with Figure 2 To understand the foregoing X, Y and Z, the forward direction can be understood as the moving direction of the carrier vehicle of the laser radar; and the effective boundary interval value refers to the height of the laser radar point cloud data, which can be set according to actual requirements.

[0055] Step 102: using a point cloud segmentation network to perform coarse-grained prediction prior on the laser point cloud data to obtain a coarse-grained prior result.

[0056] The coarse-grained initial verification result includes four-dimensional information, which includes the point cloud coordinates of the point cloud points in three spatial dimensions and the category of the point cloud points.

[0057] For example, in one embodiment, a coarse-grained prior for point cloud semantic segmentation can be obtained by using the PointNet 3D point cloud network. The PointNet network is used to pre-segment the point cloud scene with coarse grain. After segmentation, there will be areas of imprecise segmentation in the point cloud data; these imprecise segments serve as the segmentation prior. Subsequent processing of the coarse-grained prior results will then yield accurate object segmentation results. The coarse-grained prior results include the category of the point cloud points, which refers to the object category scanned by the LiDAR, such as people, vehicles, or other targets.

[0058] The specific implementation of coarse-grained prior prediction for laser point cloud data will be described in detail in the later embodiments, and will not be elaborated on here.

[0059] Step 103: Dimensionally reduce the coarse-grained prior results into one-dimensional point cloud data.

[0060] When dealing with multiple point cloud points, calculating the distance between points requires multiple square root operations, resulting in high computational cost and low efficiency. Therefore, in this embodiment of the invention, the three-dimensional data of the spatial point cloud is reduced to a one-dimensional representation, which can greatly reduce the workload of subsequent data processing and facilitate the rapid execution of the entire point cloud segmentation process. The specific implementation of data dimensionality reduction will be described in detail in later embodiments.

[0061] Step 104: Extract local point cloud data of the object from the reduced point cloud data.

[0062] In this implementation, rapid data extraction is performed on the dimensionality-reduced point cloud data to obtain local point cloud data of the object. The local point cloud data refers to the point cloud data of the object itself and some points around the object.

[0063] Step 105: Perform clustering and segmentation on the local point cloud data of the object to obtain point cloud segmented objects.

[0064] Specifically, the K-means clustering algorithm can be used to cluster the local point cloud data of the object, and then the object can be further segmented to obtain the point cloud segmented object.

[0065] Figure 3 This is a schematic diagram of the point cloud segmentation network principle disclosed in the embodiments of the present invention, combined with... Figure 3As shown, by using the PointNet network to perform point cloud semantic segmentation prior to achieve a coarse-grained prior result of the scene object, the spatial point cloud three-dimensional dimension data is reduced to one-dimensional data for fast extraction of object local point cloud data information, and the extracted object local data is used to improve the segmentation accuracy of the point cloud by using the K-means clustering method.

[0066] In the point cloud segmentation method described in the embodiment, the coarse-grained prior result is reduced to one-dimensional reduced point cloud data, which can greatly reduce the subsequent data calculation amount and improve the data processing speed under the premise of retaining three-dimensional data characteristics; the object local data is quickly extracted using the reduced point cloud data, and the object local data is clustered and segmented to improve the segmentation accuracy of the laser point cloud object.

[0067] In the above embodiment, the three-dimensional point cloud data of the laser point cloud data includes an [N, 3] format data matrix of N point clouds, where N represents the number of point cloud points, and 3 represents the point cloud coordinate values of the point cloud points in three spatial dimensions. The coarse-grained prediction prior of the point cloud segmentation network on the laser point cloud data can include: classifying each point cloud point in the data matrix by the laser point cloud segmentation network to obtain a coarse-grained prior result in the format of [N, 4], where 4 represents the point cloud coordinate values of the point cloud points in three spatial dimensions and the class of the point cloud points.

[0068] For example, a point cloud data graph in the format of [N, 3] can be input into the PointNet network for semantic segmentation prior. [N, 3] represents an N-row, 3-column data matrix, where N represents the number of point cloud points, and 3 represents the point cloud coordinate values in the X, Y, and Z directions. The PointNet network can classify the input point cloud into a single laser point class, and the network outputs a prior segmentation point cloud data in the format of [N, 4] (the data structure is similar to the aforementioned [N, 3]), where N is the number of point cloud points in a point cloud data graph, and 4 represents four columns of data, specifically x, y, z, and label, where x, y, and z represent point cloud coordinate values, and label represents the class of the point cloud point. Therefore, the classification of each point cloud point in the data matrix by the laser point cloud segmentation network to obtain a coarse-grained prior result in the format of [N, 4] can include: inputting the point cloud data in the format of [N, 3] into the laser point cloud segmentation network; the laser point cloud segmentation network classifies each point cloud point in the received point cloud data into a single laser point class and outputs a coarse-grained prior result in the format of [N, 4].

[0069] The network structure diagram of PointNet is as follows: Figure 4As shown in the figure. Among them, inputtransform, feature transform align the input points with the feature points; feature fusion combines local information and global information; MLP (Multi-Layer Perceptron) extracts point cloud features.

[0070] Figure 5 A flowchart of data dimension reduction processing disclosed by an embodiment of the present application, Figure 5 The steps shown can be performed for each point cloud point. Referring to Figure 5 As shown in the figure, the dimension reduction of the coarse-grained priori result into one-dimensional dimension reduction point cloud data can include:

[0071] Step 501: Based on the set interval length resolution of each dimension, the interval positioning of each point cloud point in the coarse-grained priori result in the format of [N, 4] is performed in three spatial dimensions respectively, and three interval resolutions corresponding to three spatial dimensions are obtained.

[0072] For example, in some embodiments, the interval positioning of the point cloud point in the coarse-grained priori result in the format of [N, 4] based on the set interval length resolution of each dimension is performed in three spatial dimensions respectively, and three interval resolutions corresponding to three spatial dimensions are obtained, which can include: in the parent interval represented by the effective boundary interval value of each spatial dimension, it is determined that the point cloud point is located in the first half interval or the second half interval, if in the first half interval, marked with a first identifier, for example, marked with 0, if in the second half interval, marked with a second identifier, for example, marked with 1; the first half interval or the second half interval where the point cloud point is located is taken as the latest parent interval, and the above step of determining whether the point cloud point is located in the first half interval or the second half interval is repeated until the number of identification bits obtained is the same as the number of identification bits corresponding to the interval length resolution.

[0073] For example, Figure 6 The dimension reduction processing principle diagram disclosed by an embodiment of the present application is shown in the figure. Referring to Figure 6 For the coarse-grained priori result [N, 4], each point is composed of four-dimensional information, and now the dimension reduction processing is performed on each laser point data, and it is judged point by point whether the point cloud point x is located in the [0, X / 2] interval (where point x represents the x coordinate value of a point in the point cloud; capital X is Figure 3If x is located in the interval [0, X / 2], then continue to judge whether x is located in [0, X / 4] or [X / 4, X / 2], and if x is located in [0, X / 4], then mark 0 in the second bit of the interval length, otherwise if x is located in [X / 4, X / 2], then mark 1 in the second bit of the interval length; if x is located in the interval [X / 2, X] first, then continue to judge whether x is located in [X / 2, 3X / 4] or [3X / 4, X], and if x is located in [X / 2, 3X / 4], then mark 0 in the second bit of the interval length, otherwise if x is located in [3X / 4, X], then mark 1 in the second bit of the interval length; judge x according to the above cycle until the specified M interval length resolution is reached (where M represents the number of marks 0 / 1, if the value of M is larger, the number of marks 0 / 1 is more, that is, the interval in which the point cloud point is divided is more accurate, and then the point cloud position represented by the M interval length is more accurate).

[0074] As shown in Figure 7A , the same is true for the judgment and marking of y in the Y interval, and the interval length resolution is also M; the judgment and marking of z in the Z interval, and the interval length resolution is also M. Step 502: The three interval resolutions are alternately interpolated and combined to obtain one-dimensional reduced dimension point cloud data of three-dimensional point cloud data.

[0075] Among them, every 3 bits of the one-dimensional reduced dimension point cloud data represent a local rectangular body of the point cloud space.

[0076] Step 502: The three interval resolutions are alternately interpolated and combined to obtain one-dimensional reduced dimension point cloud data of three-dimensional point cloud data.

[0077] As shown in Figure 7B , the three interval resolutions M of x, y, and z obtained after judgment are alternately interpolated and combined to realize the conversion of three-dimensional data to one-dimensional data.

[0078] Step 503: The one-dimensional reduced dimension point cloud data is fused into the coarse-grained prior result of [N, 4] format to obtain data of [N, 5] format, where 5 represents the point cloud coordinate value of the point cloud point in three spatial dimensions, the one-dimensional reduced dimension point cloud data, and the category of the point cloud point.

[0079] For example, before the data of the three interval resolutions M of x, y, and z are fused into the label dimension, construct the output [N, 5] data.

[0080] In the above embodiments, the step of extracting local point cloud data of an object from the reduced-dimensional point cloud data may include: extracting all point cloud data with the same first digits from the one-dimensional reduced-dimensional point cloud data according to the point cloud point categories in the [N, 5] format data, and obtaining all point cloud values ​​within the spatial region represented by the three-dimensional point cloud data; obtaining local point cloud data of the object based on the all point cloud values, wherein the first digits include the first three digits of the one-dimensional reduced-dimensional point cloud data.

[0081] Figure 8 This is a schematic diagram illustrating the rapid extraction of local object point cloud data disclosed in an embodiment of the present invention. Figure 8 As shown, a small rectangle represents the space where a local object's point cloud data is located. Based on the label information in the [N, 5] data, the local point cloud data of the object is extracted. Then, point cloud extraction is performed in the list data according to category. This extraction utilizes 3M data for fast data extraction. Each 3-digit value of 3M represents a local rectangle in the point cloud space (according to the previous reduction model, "each 3-digit value" represents the precision of the spatial cube). The later the position, the smaller the local rectangle in the point cloud space. For example, in 3M data 110100101, 110 represents a spatial range greater than 100 or 101; it is a precision division, where the preceding position represents a larger spatial area than the following position. Using the 3M data representation, all point cloud data with the same first digit are extracted to obtain all point cloud values ​​within the spatial region.

[0082] like Figure 9 As shown, the step of clustering and segmenting the local point cloud data of the object to obtain point cloud segmented objects can include: using the K-means clustering algorithm to perform clustering calculations on the local point cloud data of the object to obtain clustered point cloud data; and matching the object category of the clustered point cloud data based on the category of the point cloud points in the [N, 5] format data to obtain point cloud segmented objects. Using x, y, z data for K-means clustering has high computational complexity. This paper utilizes the K-means algorithm to perform clustering calculations on the extracted local information data of the object point cloud, and applies the labels in [N, 5] to the clustered data for specific object category matching, thereby achieving the purpose of laser point cloud segmentation of the object.

[0083] In this invention, the PointNet laser point cloud segmentation network is used to perform coarse-grained prediction priors of objects and backgrounds in the point cloud scene. The laser point cloud after the priors is then subjected to dimensionality reduction data transformation. Local object point cloud data is quickly extracted from the transformed data, and the K-means clustering algorithm is used to apply the dimensionality reduction data to perform fast clustering and segmentation of the local point cloud data. This allows for object segmentation with faster computation speed and improves the segmentation accuracy of laser point cloud objects.

[0084] For each method embodiment described above, for the sake of simple description, it is expressed as a combination of a series of actions, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the present application, certain steps can be performed in other order or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily necessary for the present application.

[0085] The method is described in detail in the above disclosed embodiments of the present application, and the method of the present application can be implemented in various forms of devices, so the present application also discloses a device, and the following specific embodiments are given to describe in detail.

[0086] Figure 10 A structure diagram of a point cloud segmentation device disclosed in an embodiment of the present application is shown in Figure 10 As shown in the figure, the point cloud segmentation device 100 can include:

[0087] A data acquisition module 101 is configured to acquire laser point cloud data, and the laser point cloud data includes three-dimensional point cloud data.

[0088] A network prior module 102 is configured to use a point cloud segmentation network to perform coarse-grained prediction prior on the laser point cloud data to obtain a coarse-grained prior result, wherein the coarse-grained prior result includes four-dimensional information.

[0089] A data dimension reduction module 103 is configured to reduce the dimension of the coarse-grained prior result to one-dimensional reduced dimension point cloud data.

[0090] A local extraction module 104 is configured to extract object local point cloud data from the reduced dimension point cloud data.

[0091] A clustering segmentation module 105 is configured to perform clustering segmentation on the object local point cloud data to obtain a point cloud segmentation object.

[0092] In the implementation process of the point cloud segmentation device described in the embodiment, the coarse-grained prior result is reduced to one-dimensional reduced dimension point cloud data, which can greatly reduce the subsequent data calculation amount and improve the data processing speed on the premise of retaining the three-dimensional data characteristics; the object local data is quickly extracted using the reduced dimension point cloud data, and the object local data is clustered and segmented, thereby improving the segmentation accuracy of the laser point cloud object.

[0093] The specific implementation of the point cloud segmentation device and each module contained therein and other possible implementation manners can be referred to the content introduction of the corresponding part in the method embodiment, which will not be repeated here.

[0094] Any one of the point cloud segmentation devices in the above embodiments comprises a processor and a memory, the data acquisition module, the network priori module, the data dimension reduction module, the local extraction module, the clustering segmentation module and the like in the above embodiments are stored in the memory as program modules, and the corresponding functions are realized by the processor executing the above program modules stored in the memory.

[0095] The processor comprises a core, and the core retrieves the corresponding program modules in the memory. The core can be set as one or more, and the processing of the revisited data is realized by adjusting the core parameters.

[0096] The memory can comprise a non-permanent memory in a computer readable medium, random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM), and the memory comprises at least one memory chip.

[0097] The embodiment of the present application provides a storage medium, which stores a program, and the program is executed by a processor to realize the point cloud segmentation method in the above embodiment.

[0098] The embodiment of the present application provides a processor, which is used for running a program, wherein the program is executed to perform the point cloud segmentation method in the above embodiment.

[0099] Further, the embodiment provides an electronic device comprising a processor and a memory. The memory is used for storing executable instructions of the processor, and the processor is configured to execute the point cloud segmentation method in the above embodiment by executing the executable instructions.

[0100] In the specification, each embodiment is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the related parts can be referred to the method part.

[0101] It is also to be noted that, as used in the specification and the appended claims, the singular forms "a," "an" and "the" include plural referents unless otherwise indicated. Furthermore, to the extent that the terms "including," "includes," "having," "has," "with," or "contains" are used in either the detailed description and the claims, such terms are intended to be inclusive in a manner similar to the term "comprising" as an open transition term without precluding any additional or other elements.

[0102] The embodiments disclosed herein can each be implemented as a method, apparatus, or article of manufacture using programming instructions. The embodiments disclosed herein can be implemented using software, firmware, hardware, or a combination thereof. The various elements of the disclosed embodiments, as well as the procedural aspects of the disclosed embodiments, can be implemented using a variety of different programming techniques, such as object-oriented programming, procedural programming, or other programming techniques. In one embodiment, the various elements of the disclosed embodiments are implemented using a programming language such as C, C++, Java, or other programming language. In another embodiment, the various elements of the disclosed embodiments are implemented using a scripting language such as JavaScript, VBScript, or other scripting language. In another embodiment, the various elements of the disclosed embodiments are implemented using a hardware description language such as VHDL, Verilog, or other hardware description language. In another embodiment, the various elements of the disclosed embodiments are implemented using a hardware-agnostic description language such as PSL, SLD, or other hardware-agnostic description language. In another embodiment, the various elements of the disclosed embodiments are implemented using a hardware-specific description language such as RTL, GDSII, or other hardware-specific description language. In another embodiment, the various elements of the disclosed embodiments are implemented using a hardware description language such as VHDL, Verilog, or other hardware description language. In another embodiment, the various elements of the disclosed embodiments are implemented using a hardware-agnostic description language such as PSL, SLD, or other hardware-agnostic description language. In another embodiment, the various elements of the disclosed embodiments are implemented using a hardware-specific description language such as RTL, GDSII, or other hardware-specific description language.

[0103] The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for point cloud segmentation, characterized in that, The method comprises the following steps: acquiring laser point cloud data, wherein the laser point cloud data comprises three-dimensional point cloud data, the three-dimensional point cloud data of the laser point cloud data comprises an [N, 3] format data matrix of N point clouds, [N, 3] represents an N-row and 3-column data matrix, wherein N represents the number of point cloud points, and the three columns comprise point cloud coordinate values of the point cloud points in three spatial dimensions; performing category classification on each point cloud point in the data matrix by using a laser point cloud segmentation network to obtain a [N, 4] format coarse-grained prior result, wherein [N, 4] represents an N-row and 4-column data matrix, the four columns comprise point cloud coordinate values of the point cloud points in three spatial dimensions and a category of the point cloud points, and the coarse-grained prior result comprises four-dimensional information; performing interval positioning on each point cloud point in the [N, 4] format coarse-grained prior result in three spatial dimensions based on a set interval length resolution of each dimension to obtain three interval resolutions corresponding to the three spatial dimensions; and performing alternating interpolation dimension reduction combination on the three interval resolutions to obtain one-dimensional dimension-reduced point cloud data of the three-dimensional point cloud data; fusing the one-dimensional dimension-reduced point cloud data into the [N, 4] format coarse-grained prior result to obtain [N, 5] format data, wherein [N, 5] represents an N-row and 5-column data matrix, and the five columns comprise point cloud coordinate values of the point cloud points in three spatial dimensions, one-dimensional dimension-reduced point cloud data and a category of the point cloud points; extracting all point cloud data with the same number of front bits by using the one-dimensional dimension-reduced point cloud data according to the category of the point cloud points in the [N, 5] format data to obtain all point cloud values in a spatial region represented by the three-dimensional point cloud data; and obtaining object local point cloud data according to the all point cloud values, wherein the number of the front bits comprises the first three bits of the one-dimensional dimension-reduced point cloud data; performing clustering segmentation on the object local point cloud data to obtain a point cloud segmented object.

2. The point cloud segmentation method of claim 1, wherein, The method of performing category classification on each point cloud point in the data matrix by using the laser point cloud segmentation network to obtain the [N, 4] format coarse-grained prior result comprises the following steps: inputting the [N, 3] format point cloud data into the laser point cloud segmentation network; the laser point cloud segmentation network performs single laser point category classification on each point cloud point in the received point cloud data and outputs the [N, 4] format coarse-grained prior result.

3. The point cloud segmentation method of claim 1, wherein, Each 3-bit value in the one-dimensional dimension-reduced point cloud data represents a local rectangular body in a point cloud space.

4. The point cloud segmentation method of claim 1, wherein, The method of performing interval positioning on each point cloud point in the [N, 4] format coarse-grained prior result in three spatial dimensions based on the set interval length resolution of each dimension to obtain the three interval resolutions corresponding to the three spatial dimensions comprises the following steps: in a mother interval represented by an effective boundary interval value of each spatial dimension, determining whether a point cloud point is located in a front half interval or a rear half interval, and if the point cloud point is located in the front half interval, marking the point cloud point with a first mark, and if the point cloud point is located in the rear half interval, marking the point cloud point with a second mark. The front half interval or the rear half interval where the point cloud point is located is taken as the latest parent interval, and the above steps of determining whether the point cloud point is located in the front half interval or the rear half interval are repeated until the number of identification bits obtained is the same as the number of identification bits corresponding to the interval length resolution.

5. The point cloud segmentation method of claim 2, wherein, The object local point cloud data is obtained by clustering and segmenting the object local point cloud data based on the category of the point cloud point in the [N, 5] format data. The object local point cloud data is obtained by clustering and segmenting the object local point cloud data based on the category of the point cloud point in the [N, 5] format data. The object local point cloud data is obtained by clustering and segmenting the object local point cloud data based on the category of the point cloud point in the [N, 5] format data.

6. An apparatus for point cloud segmentation, the apparatus comprising: The object local point cloud data is obtained by clustering and segmenting the object local point cloud data based on the category of the point cloud point in the [N, 5] format data. The data acquisition module is configured to acquire laser point cloud data, wherein the laser point cloud data comprises three-dimensional point cloud data, and the three-dimensional point cloud data of the laser point cloud data comprises an [N, 3] format data matrix of N point clouds, wherein [N, 3] represents an N-row and 3-column data matrix, N represents the number of point cloud points, and the 3 columns comprise point cloud coordinate values of the point cloud points in three spatial dimensions. The network prior module is configured to perform category classification on each point cloud point in the data matrix through a laser point cloud segmentation network to obtain a [N, 4] format coarse-grained prior result, wherein [N, 4] represents an N-row and 4-column data matrix, the 4 columns comprise point cloud coordinate values of the point cloud points in three spatial dimensions and the category of the point cloud points, and the coarse-grained prior result comprises four-dimensional information. The data dimension reduction module is configured to perform interval positioning on each point cloud point in the [N, 4] format coarse-grained prior result in three spatial dimensions based on a set interval length resolution of each dimension to obtain three interval resolutions corresponding to the three spatial dimensions; and perform alternating interpolation dimension reduction combination on the three interval resolutions to obtain one-dimensional dimension-reduced point cloud data of the three-dimensional point cloud data. The local extraction module is configured to fuse the one-dimensional dimension-reduced point cloud data into the [N, 4] format coarse-grained prior result to obtain [N, 5] format data, wherein [N, 5] represents an N-row and 5-column data matrix, the 5 columns comprise point cloud coordinate values of the point cloud points in three spatial dimensions, one-dimensional dimension-reduced point cloud data, and the category of the point cloud points; all point cloud data with the same number of front bits is extracted from the one-dimensional dimension-reduced point cloud data according to the category of the point cloud point in the [N, 5] format data to obtain all point cloud values in a spatial region represented by the three-dimensional point cloud data; and the object local point cloud data is obtained based on the all point cloud values, wherein the number of front bits comprises the first three bits of the one-dimensional dimension-reduced point cloud data. The clustering and segmentation module is configured to cluster and segment the object local point cloud data to obtain point cloud segmented objects.

Images