A lunar surface point cloud obstacle semantic segmentation method, system, electronic device and program product

Abstract

The application discloses a lunar surface point cloud obstacle semantic segmentation method, system, electronic equipment and program product, and belongs to the cross field of computer vision, deep learning and planetary exploration environment perception technology. The application introduces a boundary enhancement mechanism based on local geometric change in the point cloud feature coding stage, strengthens the feature expression of the obstacle boundary area points, and effectively alleviates the boundary information loss problem caused by random sampling. And by introducing a global context perception mechanism on the feature transmission path between the feature coding unit and the semantic decoding unit, the long-distance semantic dependence information is compensated, so that the segmentation result of the large-scale lunar surface scene is more coherent and stable. The application effectively solves the precision decline problem caused by boundary blur and global context loss in lunar surface point cloud segmentation, significantly improves the segmentation precision of obstacles such as impact craters and rocks, maintains the high efficiency of the algorithm, and is suitable for a star-borne computing platform.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer vision, deep learning and planetary exploration environment perception technology, and specifically relates to a semantic segmentation method, system, electronic device and program product for lunar surface point cloud obstacles. Background Technology

[0002] Autonomous exploration of the lunar surface requires rovers to possess real-time and accurate perception capabilities of their surroundings. LiDAR, capable of acquiring precise three-dimensional point clouds independent of lighting conditions, is a key sensor for lunar environment perception.

[0003] However, lunar surface point clouds have the following notable characteristics:

[0004] (1) The terrain is unstructured and lacks texture information;

[0005] (2) The geometric transition between the obstacle and the background (lunar soil) is smooth, and the semantic boundaries are blurred;

[0006] (3) The point cloud is large in scale and the distribution of categories is extremely unbalanced;

[0007] (4) The spaceborne computing platform has limited resources and has extremely high requirements for the computational complexity and real-time performance of the algorithm.

[0008] In recent years, deep learning-based methods for direct point cloud processing have made significant progress. Existing point cloud segmentation methods such as PointNet, PointNet++, KPConv, and RandLA-Net have achieved good results in terrestrial scenarios. Among them, point cloud feature aggregation neural networks based on random sampling are considered more suitable for spaceborne embedded platforms due to their high computational efficiency and low memory consumption. However, these networks still have significant shortcomings in lunar applications: random sampling easily loses key geometric information at obstacle boundaries, and they lack effective modeling of global semantic consistency in large-scale scenes, leading to problems such as blurred obstacle boundaries and cross-regional semantic inconsistencies.

[0009] While some existing technologies have attempted to introduce attention mechanisms or geometric feature enhancement methods, most are designed for regular sampling or structured point clouds, failing to address the combined problem of boundary information loss and global semantic deficiency in randomly sampled point clouds on the lunar surface. Therefore, a dedicated semantic segmentation method that balances high accuracy and efficiency and is applicable to randomly sampled point cloud structures on the lunar surface is urgently needed. Summary of the Invention

[0010] To address the problems of existing random sampling point cloud neural networks in lunar surface obstacle semantic segmentation tasks, which suffer from the systematic loss of key geometric information due to irregular point cloud geometry, smooth obstacle boundary transitions, and random downsampling, resulting in blurred obstacle boundaries, inconsistent semantics across regions, and difficulty in balancing accuracy and computational efficiency, this invention proposes a lunar surface point cloud obstacle semantic segmentation method, system, electronic device, and program product.

[0011] The first aspect discloses a semantic segmentation method for obstacles in lunar surface point clouds, the method comprising:

[0012] A three-dimensional point cloud data of the lunar surface is acquired and preprocessed to obtain point cloud blocks with at least a predetermined number of points. Each point cloud block is input into a feature encoding unit. During feature encoding, a boundary enhancement feature modulation mechanism based on local geometric changes is used to adaptively enhance the point cloud features, resulting in enhanced point cloud features. By globally modeling the overall features of the point cloud blocks, channel attention weights are generated to modulate the point cloud features. The enhanced point cloud features are then calibrated at the channel level according to the channel attention weights to obtain global perception features. The global perception features are input into a semantic decoding unit to output the obstacle semantic category corresponding to each point.

[0013] The second aspect discloses a semantic segmentation system for lunar surface point cloud obstacles, the system comprising:

[0014] The data acquisition and preprocessing unit is used to acquire three-dimensional point cloud data of the lunar surface and preprocess the three-dimensional point cloud data to obtain point cloud blocks with at least a preset number of points; the feature encoding unit is used to perform random sampling feature encoding on each point cloud block, and during the feature encoding process, adaptively enhance the point cloud features based on a boundary enhancement feature modulation mechanism of local geometric changes to obtain enhanced point cloud features; the global context awareness unit is used to generate channel attention weights for modulating point cloud features by globally modeling the overall features of the point cloud blocks, and perform channel-level calibration on the enhanced point cloud features according to the channel attention weights to obtain global awareness features; the semantic decoding unit is used to output the obstacle semantic category corresponding to each point according to the global awareness features.

[0015] The third aspect discloses an electronic device including a processor and a memory, the memory storing a computer program that, when executed, implements a semantic segmentation method for lunar surface point cloud obstacles as disclosed in the first aspect or any possible implementation thereof.

[0016] The fourth aspect discloses a computer program product that, when run on a computer, causes the computer to perform a semantic segmentation method for lunar surface point cloud obstacles disclosed in the first aspect or any possible implementation of the first aspect.

[0017] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0018] This invention significantly enhances the feature representation of obstacle boundary points by introducing a boundary enhancement mechanism based on local geometric changes during the point cloud feature encoding stage, effectively mitigating the problem of missing boundary information caused by random sampling. Furthermore, by introducing a global context-aware mechanism along the feature transfer path between the feature encoding unit and the semantic decoding unit, it compensates for long-distance semantic dependencies, making the segmentation results of large-scale lunar scenes more coherent and stable. This invention effectively solves the accuracy degradation problem caused by boundary ambiguity and missing global context in lunar point cloud segmentation. While significantly improving the segmentation accuracy of obstacles such as impact craters and rocks, it maintains the algorithm's high efficiency, making it suitable for spaceborne computing platforms and providing reliable technical support for the autonomous and safe navigation of lunar rovers. Attached Figure Description

[0019] Figure 1 A flowchart illustrating the overall process of a semantic segmentation method for lunar surface point cloud obstacles provided by this invention;

[0020] Figure 2 This is a schematic diagram of a semantic segmentation system architecture for lunar surface point cloud obstacles provided by the present invention. Detailed Implementation

[0021] To make the objectives, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of the present invention will be thorough and complete.

[0022] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly on the other element or there may be an intervening element. When an element is considered to be "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," "up," "down," and similar expressions used herein are for illustrative purposes only and are not intended to indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0024] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances. The term "and / or" as used herein includes any and all combinations of one or more of the related listed items.

[0025] To address the problems of existing random sampling point cloud neural networks in lunar surface obstacle semantic segmentation tasks, which suffer from the systematic loss of key geometric information due to irregular point cloud geometry, smooth obstacle boundary transitions, and random downsampling, resulting in blurred obstacle boundaries, inconsistent semantics across regions, and difficulty in balancing accuracy and computational efficiency, this invention proposes a lunar surface point cloud obstacle semantic segmentation method, system, electronic device, and program product.

[0026] Without significantly increasing computational and storage overhead, this invention effectively improves the recognition accuracy of boundary regions and the semantic consistency of large-scale scenes in the semantic segmentation of lunar surface obstacles by introducing a synergistic boundary enhancement mechanism and a global context awareness mechanism for the existing random sampling point cloud network structure. This makes the invention applicable to resource-constrained spaceborne computing platforms and enhances the autonomous environmental perception and safe navigation capabilities of lunar rovers under resource-constrained conditions.

[0027] In one embodiment, the present invention provides a semantic segmentation method for obstacles in point clouds on the lunar surface, such as... Figure 1 As shown, this method specifically includes the following steps:

[0028] S101. Acquire three-dimensional point cloud data of the lunar surface and preprocess the three-dimensional point cloud data to obtain at least one point cloud block with a preset number of points.

[0029] Specifically, the preprocessing of 3D point cloud data mainly includes:

[0030] The 3D point cloud data is subjected to density equalization processing, and the 3D point cloud data is divided into blocks according to preset rules to obtain at least one point cloud block with a preset number of points. The preset number of points is set according to computing resources and scene complexity.

[0031] It should be noted that the raw lunar surface lidar point cloud data undergoes density equalization processing, for example, through voxelization downsampling or random sampling to reduce point cloud density and achieve density equalization. The 3D point cloud data is then divided into point cloud blocks of a predetermined number of points according to preset rules. That is, each point cloud block contains a fixed number of points, which can be set based on computing resources and scene complexity. The preset rules can be selected according to actual needs, and can be from left to right, top to bottom, etc., ranging from hundreds to thousands of points. For point cloud blocks with insufficient points, they can be supplemented through resampling or padding. The divided point cloud blocks are then used as input to the subsequent neural network.

[0032] It should be noted that, depending on the actual data requirements, operations such as unifying the point cloud coordinate scale and supplementing or cropping point cloud attribute information can be performed on the original point cloud data. Through downsampling, segmentation, and formatting, the original point cloud data can be converted into a data structure that can be directly processed by neural networks.

[0033] S102. Input each point cloud block into the point cloud feature encoding unit. During the feature encoding process, the point cloud features are adaptively enhanced based on the boundary enhancement feature modulation mechanism of local geometric changes to obtain the point cloud enhanced features.

[0034] In another embodiment, this step specifically includes:

[0035] S1021. Construct local geometric description features based on the spatial neighborhood of each sampling point in the point cloud block;

[0036] S1022. Input the local geometric description features into a nonlinear mapping function to perform feature mapping, and output the boundary modulation weights used to characterize the saliency of semantic boundaries.

[0037] S1023. The point cloud features of the sampling points are adaptively enhanced by using the boundary modulation weights in an element-wise multiplication manner to obtain the enhanced point cloud features.

[0038] The preprocessed point cloud blocks are input into the feature encoding unit. During the feature encoding process, for each sampling point, a local geometric description feature is constructed based on its spatial neighborhood, and a boundary modulation weight is generated based on the local geometric description feature. The boundary modulation weight is used to adaptively modulate the point cloud features, so that points located in the semantic boundary region of the obstacle obtain a stronger feature response, improve the feature discrimination of points located in the boundary region of the obstacle, and thus reduce the loss of boundary information caused by random sampling.

[0039] In another implementation, for each sampling point Based on its spatial neighborhood point set Constructing local geometric description features Among them, the local geometric description features are used to reflect the geometric changes of the sampling points in the local space.

[0040] Local geometric description features via nonlinear mapping function Generate boundary modulation weights Their relationship can be expressed as:

[0041] (1)

[0042] Then, the boundary modulation weights are applied. Apply element-wise weighted methods to the point cloud features corresponding to the sampling points. The point cloud features are then fused using residual connections to obtain enhanced point cloud features. Their relationship can be expressed as:

[0043] (2)

[0044] Among them, symbols This indicates element-wise multiplication.

[0045] In a specific embodiment, the local geometric description features can be one of curvature features, normal vector variation features, and point density variation features. Curvature features can be obtained through geometric analysis of the local neighborhood point set of the sampling point. For example, the principal direction feature can be calculated based on the covariance matrix of the neighborhood points, and curvature information characterizing the degree of curvature of the local surface can be obtained according to the relationship of their eigenvalues. Normal vector variation features can be obtained by calculating the angle change between the normal vectors of the sampling point and its corresponding neighborhood points, reflecting the continuity of the normal vector within a local region. Point density variation features can be obtained by statistically analyzing the number of points in the neighborhood of the sampling point or the distribution of points per unit volume, describing the spatial sparsity or density variation of the point cloud.

[0046] The following section uses local geometric description features as curvature features as an example to further illustrate the present invention.

[0047] When the local geometric description features include curvature features, the boundary enhancement feature modulation mechanism based on curvature features adaptively enhances the point cloud features, and the process of obtaining the enhanced point cloud features specifically includes:

[0048] Calculate the covariance matrix of the K neighborhood points of the sampling point;

[0049] The curvature features characterizing the degree of local surface curvature at the sampling point are obtained from the eigenvalues ​​of the covariance matrix.

[0050] The curvature features are sequentially passed through a multilayer perceptron and... The function obtains the boundary modulation weights.

[0051] Assuming given input features and corresponding point cloud coordinates For each sampling point , , To represent the point cloud features of sampling point i, construct its K nearest neighbor set M. The value of K is set based on practical experience, and the covariance matrix of its neighborhood points is calculated. The minimum eigenvalue is used as the curvature of the sampling point. The estimate is shown in the following formula:

[0052] (3)

[0053] Where j represents the index of the K value, Let j be the point cloud feature representation of a neighboring point j. It represents the centroid of the neighboring points.

[0054] For covariance matrix Perform eigenvalue decomposition to obtain eigenvalues. The curvature feature is represented by the normalized form of the minimum eigenvalue as shown in the formula:

[0055] (4)

[0056] in, The matrix normalization function represents the minimum eigenvalue.

[0057] It should be noted that curvature features are used to characterize the degree of geometric change of a sampling point in local space. The larger the curvature value, the more drastic the geometric change in the region where the point is located, and the more likely it is to exist on a terrain boundary. When the neighborhood of a sampling point is approximately coplanar, its minimum eigenvalue is close to zero; when the sampling point is located in a region of geometric abrupt change or boundary, the dispersion of neighborhood points in the normal direction increases, and the minimum eigenvalue increases accordingly. Therefore, the minimum eigenvalue or its normalized form can be used to characterize local curvature features.

[0058] Then, the curvature features of all points are... Through a lightweight multilayer perceptron and The function is mapped to the boundary attention weight graph. The boundary attention weight map quantifies the probability that each point belongs to the boundary.

[0059] (5)

[0060] in, This represents a multilayer perceptron. )express function, This represents the set of curvature features for all sampling points.

[0061] Finally, the input point cloud features are modulated using the boundary modulation weights corresponding to each point to enhance the feature representation of the boundary region. The weights are then applied to all channels via a broadcast mechanism, and the final output retains the original feature information through residual connections, specifically obtained using the following formula:

[0062] (6)

[0063] in, This represents point-by-point multiplication. This represents point cloud enhancement features, and this calculation method can preserve the original feature information.

[0064] S103. By globally modeling the overall features of the point cloud blocks, channel attention weights for modulating point cloud features are generated.

[0065] In one embodiment, this step specifically includes:

[0066] The input point cloud features are globally aggregated in terms of the number of points to obtain a context feature vector representing global semantic information.

[0067] The context feature vector is calculated using two convolutional layers of compression and dilation to obtain the channel attention weights used to modulate the point cloud features.

[0068] S104. Perform channel-level calibration on the point cloud enhancement features according to the channel attention weights used to modulate the point cloud features to obtain global perception features;

[0069] In another implementation, the point cloud enhancement features are broadcast multiplied with the channel attention weights, and the output is passed through a residual connection layer to obtain global perception features.

[0070] It should be noted that during the process of transferring point cloud features from the encoding stage to the decoding stage, the encoded point cloud features are globally aggregated to generate channel attention weights for modulating the feature channel responses, in order to compensate for the loss of long-distance semantic dependencies caused by random sampling.

[0071] The encoded point cloud features are globally aggregated along the dimension of point count to obtain a context feature vector representing global semantic information. The aggregation operation can be global averaging, global maximization, or a combination thereof.

[0072] The context feature vector is nonlinearly mapped to generate channel attention weights for modulating point cloud features, and then used to perform channel-level calibration on the point cloud features.

[0073] To further illustrate the processing procedure of the global context-aware unit, a specific embodiment is described below, given input features. First, aggregate using Global Average Pooling (GAP). The spatial information of all points is shown in formula (7):

[0074] (7)

[0075] in, This represents the point cloud feature of the i-th point. Indicates global average pooling. This is the context feature vector.

[0076] Then, a bottleneck structure is employed to achieve efficient feature transformation. Through a two-step compression-dilation convolution, the context feature vector is transformed into channel attention weights. First, a 1×1 convolution is used to reduce the dimensionality of the context feature vector, resulting in a feature dimension of... Among them, compression ratio The adjustment is made dynamically based on the input feature dimension, as shown in the following formula:

[0077] (8)

[0078] Where C represents the feature dimension, and when the feature dimension is less than 64, the compression ratio is... It equals 4, otherwise, It equals 16.

[0079] After ReLU activation, a 1×1 convolution is used to increase the dimensionality back to the original number of channels, and finally, the channel attention weights are generated using the Sigmoid function. The specific process is shown in the following formula:

[0080] (9)

[0081] in, This represents a dimension reduction convolution operation. This represents a higher-dimensional convolution operation. This represents the activation function.

[0082] The learned channel attention weights are then applied to the original input features, emphasizing important feature channels and suppressing secondary channels to inject global contextual information. Finally, residual connections are used to prevent global information from overshadowing local details, and the final result, i.e., the globally perceived features, is output. The specific process is shown in the formula below:

[0083] (10)

[0084] in, Represents broadcast multiplication at the channel dimension. This represents the activation function. Through the above processing, the integrity of multi-scale features can be preserved.

[0085] S105. Input the global perception features into the point cloud semantic decoding unit and output the obstacle semantic category corresponding to each point.

[0086] The point cloud features, after boundary enhancement and global context calibration, are input into the semantic decoding unit. Through stepwise feature recovery and fusion, the semantic category corresponding to each point on the lunar surface is output to achieve fine segmentation of obstacles such as lunar soil, impact craters, and rocks.

[0087] This invention introduces a boundary enhancement mechanism based on local geometric changes during the point cloud feature encoding stage, explicitly strengthening the feature representation of points in obstacle boundary regions and effectively alleviating the problem of missing boundary information caused by random sampling. Furthermore, by introducing a global context-aware mechanism along the feature transfer path between the feature encoding unit and the semantic decoding unit, it compensates for long-distance semantic dependencies, making the segmentation results of large-scale lunar scenes more coherent and stable. Moreover, both the boundary enhancement mechanism implementation unit and the global context-aware unit in this invention are lightweight structures, significantly improving segmentation performance without significantly increasing computational and storage overhead. In addition, this invention closely integrates with the characteristics of lunar point clouds—"geometry-dominated, boundary-blurred, and class-imbalanced"—and possesses good robustness and engineering application value.

[0088] This application also provides a lunar surface point cloud obstacle semantic segmentation system corresponding to the above-described lunar surface point cloud obstacle semantic segmentation method embodiment. Since the system embodiment is basically similar to the method embodiment, it is described simply. For details of the relevant technical features and the effects achieved, please refer to the corresponding description of the lunar surface point cloud obstacle semantic segmentation method embodiment provided above. Figure 2 This is a schematic diagram of a semantic segmentation system architecture for lunar surface point cloud obstacles, as disclosed in an embodiment of the present invention. Figure 2 As shown, the lunar surface point cloud obstacle semantic segmentation system disclosed in this invention mainly includes: a data acquisition and preprocessing unit, used to acquire three-dimensional point cloud data of the lunar surface and preprocess the three-dimensional point cloud data to obtain at least one point cloud block with a preset number of points.

[0089] The feature encoding unit is used to randomly sample and encode features for each point cloud block. During the feature encoding process, the point cloud features are adaptively enhanced based on the boundary enhancement feature modulation mechanism of local geometric changes to obtain enhanced point cloud features.

[0090] The global context awareness unit is used to generate channel attention weights for modulating point cloud features by globally modeling the overall features of point cloud blocks, and to perform channel-level calibration on the point cloud enhancement features based on the channel attention weights to obtain globally aware features.

[0091] The semantic decoding unit is used to output the semantic category of the obstacle corresponding to each point based on the global perception features.

[0092] It should be noted that the lunar surface point cloud obstacle semantic segmentation system proposed in this invention is implemented using a network structure based on RandLA-Net. The feature encoding unit employs a multi-layer random downsampling structure, and the boundary enhancement mechanism is set in the first layer of the encoding stage within the feature encoding unit to achieve boundary enhancement processing for point cloud features in all point cloud blocks. A lightweight global context-aware unit is placed on the feature transmission path between the feature encoding unit and the semantic decoding unit, effectively compensating for long-distance semantic dependencies, making the segmentation results of large-scale lunar scenes more coherent and stable, and applicable to resource-constrained spaceborne computing platforms.

[0093] This application also provides an electronic device, which includes a processor and a memory. The memory stores at least one instruction or at least one program, which is loaded and executed by the processor. This is a semantic segmentation method for lunar surface point cloud obstacles provided in the above-described method embodiments.

[0094] Furthermore, an electronic device is provided for implementing the method provided in the embodiments of this application. This device can participate in constituting or including the apparatus or system provided in the embodiments of this application. The electronic device may include one or more processors (processors may include, but are not limited to, processing devices such as microprocessors (MCUs) or programmable logic devices (FPGAs), a memory for storing data, and a transmission device for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a power supply, and / or a camera.

[0095] It should be noted that the aforementioned one or more processors and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits can be implemented wholly or partially as software, hardware, firmware, or any other combination. Furthermore, the data processing circuits can be a single, independent processing module, or wholly or partially integrated into any other element within a device (or mobile device). As involved in the embodiments of this application, the data processing circuit serves as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0096] The memory can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the method described in the embodiments of this application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the above-mentioned data processing method. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to electronic devices via a network. Examples of the above-mentioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0097] The transmission device is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the device's communication provider. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0098] The display can be, for example, a touchscreen liquid crystal display (LCD), which allows users to interact with the user interface of an electronic device (or mobile device).

[0099] This application also provides a computer program product or computer program, which includes computer instructions stored in a computer storage medium. The processor of an electronic device reads the computer instructions from the computer storage medium and executes the computer instructions, causing the electronic device to perform the semantic segmentation method for lunar surface point cloud obstacles provided in the above-described method embodiments.

[0100] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0101] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A semantic segmentation method for obstacles in lunar surface point clouds, characterized in that, The method includes: Acquire three-dimensional point cloud data of the lunar surface and preprocess the three-dimensional point cloud data to obtain point cloud blocks with at least a preset number of points; Each point cloud block is input into the feature encoding unit. During the feature encoding process, the point cloud features are adaptively enhanced by a boundary enhancement feature modulation mechanism based on local geometric changes, resulting in enhanced point cloud features. By globally modeling the overall features of the point cloud blocks, channel attention weights are generated to modulate the point cloud features. Global perception features are obtained by performing channel-level calibration on the point cloud enhancement features based on the channel attention weights. The global perception features are input into the semantic decoding unit, which outputs the obstacle semantic category corresponding to each point.

2. The method according to claim 1, characterized in that, The preprocessing of the three-dimensional point cloud data to obtain a point cloud block with at least a preset number of points includes: The 3D point cloud data is subjected to density equalization processing, and the 3D point cloud data is divided into blocks according to preset rules to obtain at least one point cloud block with a preset number of points, wherein the preset number of points is set according to computing resources and scene complexity.

3. The method according to claim 1, characterized in that, In the feature encoding process, a boundary enhancement feature modulation mechanism based on local geometric changes is used to adaptively enhance the point cloud features, resulting in enhanced point cloud features, including: Local geometric description features are constructed based on the spatial neighborhood of each sampling point in the point cloud block; The local geometric description features are input into a nonlinear mapping function for feature mapping, and the output is a boundary modulation weight used to characterize the semantic boundary saliency. The point cloud features of the sampling points are adaptively enhanced by using the boundary modulation weights in an element-wise multiplication manner to obtain the enhanced point cloud features.

4. The method according to claim 3, characterized in that, The local geometric description features include curvature features. The construction of local geometric description features based on the spatial neighborhood of each sampling point in the point cloud block includes: Calculate the covariance matrix of the K neighborhood points of the sampling point; Based on the eigenvalues ​​of the covariance matrix, curvature features that characterize the degree of curvature of the local surface at the sampling point can be obtained.

5. The method according to claim 4, characterized in that, The step of inputting the local geometric description features into a nonlinear mapping function for feature mapping and outputting boundary modulation weights for characterizing the salience of semantic boundaries includes: The curvature features are sequentially passed through a multilayer perceptron and... The function obtains the boundary modulation weights.

6. The method according to claim 1, characterized in that, The process of generating channel attention weights for modulating point cloud features by globally modeling the overall features of the point cloud blocks includes: The input point cloud features are globally aggregated in terms of the number of points to obtain a context feature vector representing global semantic information. The context feature vector is calculated using two convolutional layers of compression and dilation to obtain the channel attention weights used to modulate point cloud features.

7. The method according to claim 6, characterized in that, The step of performing channel-level calibration on the point cloud enhancement features based on the channel attention weights used to modulate the point cloud features to obtain globally perceptual features includes: The point cloud enhancement features are broadcast multiplied with the channel attention weights, and the output is passed through a residual connection layer to obtain global perception features.

8. A semantic segmentation system for obstacles in lunar surface point clouds, characterized in that, The system includes: The data acquisition and preprocessing unit is used to acquire three-dimensional point cloud data of the lunar surface and preprocess the three-dimensional point cloud data to obtain point cloud blocks with at least a preset number of points. The feature encoding unit is used to randomly sample and encode features for each point cloud block. During the feature encoding process, the point cloud features are adaptively enhanced based on the boundary enhancement feature modulation mechanism of local geometric changes to obtain enhanced point cloud features. The global context awareness unit is used to generate channel attention weights for modulating point cloud features by globally modeling the overall features of point cloud blocks, and to perform channel-level calibration on the point cloud enhancement features based on the channel attention weights to obtain globally aware features. The semantic decoding unit is used to output the semantic category of the obstacle corresponding to each point based on the global perception features.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory, the memory storing a computer program that is loaded and executed by the processor to implement the method as claimed in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program stored in a computer-readable storage medium, which a processor reads from and executes to implement the method as described in any one of claims 1 to 7.

Images