Service robot indoor scene semantic segmentation method, device and equipment and storage medium

Abstract

The application discloses a service robot indoor scene semantic segmentation method and device, equipment and storage medium, through collecting RGB-D point cloud data, relying on a multilayer perception machine to map the RGB-D point cloud data to a high-dimensional space, and based on a grid sampling strategy, layer by layer down-sampling the point cloud to reduce the number of points in the point cloud; for the down-sampled point cloud in each layer, a local area is constructed around each point by using a KD tree, a geometric separable convolution is used to learn local features from the local area, and the point cloud feature dimension is increased; the point cloud features are interpolated by using up-sampling, the same point cloud features as the input point number are obtained, and indoor scene semantic segmentation is performed. Compared with the prior art, the grid sampling and the geometric separable convolution are used, so that the RGB-D point cloud data semantic segmentation precision and efficiency of the service robot in an indoor scene are improved.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, and relates to a semantic segmentation method for indoor scenes, and particularly to a semantic segmentation method, apparatus, device and storage medium for indoor scenes of service robots. Background Technology

[0002] Service robots are widely used in service industries such as healthcare, catering, and rehabilitation. Compared to industrial robots with fixed paths, service robots operate in more complex indoor environments, making accurate perception of these environments fundamental for their task completion. Existing depth cameras, such as the Intel RealSense depth camera, can simultaneously acquire color images and 3D point cloud data of a scene to obtain RGB-D point cloud data, which is increasingly being used in service robots. Utilizing RGB-D point cloud data to achieve semantic segmentation of indoor scenes for service robots is one of the key technologies for improving their environmental perception and intelligence levels.

[0003] Common deep learning-based semantic segmentation methods for indoor scenes fall into two main categories: voxel-based methods and point-based methods. In practical applications, the choice between voxel-based and point-based methods typically depends on the specific application scenario and task requirements. Voxel-based methods convert point cloud data into voxels, transforming it into a regular 3D voxel mesh, thus achieving greater computational efficiency. This regular structure allows for natural integration with traditional 3D convolutional neural networks, easily capturing both local and global features. It offers advantages in handling regular structures and computational efficiency, making it suitable for tasks requiring regular representations. However, this type of method has a high memory footprint, especially when processing high-resolution voxel meshes, requiring significant computational resources. Point-based methods, on the other hand, directly process raw point cloud data without voxelization, preserving the original geometric information of the point cloud data. This makes them more suitable for scenarios with high requirements for original geometric information and irregular structures. PointNet and PointNet++, proposed by Qi et al., are two classic point-based point cloud data processing networks used for point cloud data classification and segmentation tasks. This irregular structure makes point-based methods more suitable for indoor scene point cloud data of varying densities and complexities, and makes it easier to capture subtle geometric and local features. However, due to the lack of explicit regular grids, point-based methods struggle to model the global structure.

[0004] On the other hand, existing semantic segmentation methods for indoor scenes in service robots mainly construct an end-to-end network structure similar to U-Net through a hierarchical downsampling strategy. Downsampling is primarily achieved through farthest-point sampling, which exhibits exponential time complexity, resulting in high latency when dealing with large-scale scene semantic segmentation. Furthermore, existing semantic segmentation methods for indoor scenes in service robots typically involve designing very complex modules to extract geometric information, which slows down the network's response speed. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method, apparatus, device and storage medium for semantic segmentation of indoor scenes of service robots, which aims to effectively extract local geometric information and reduce the impact of downsampling operations on speed, so as to ultimately improve the speed and accuracy of semantic segmentation of indoor scenes of service robots.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] Firstly, a semantic segmentation method for indoor scenes of service robots is provided, including:

[0008] RGB-D point cloud data is collected, mapped to a high-dimensional space using a multilayer perceptron, and the number of points in the point cloud is reduced by downsampling the point cloud layer by layer based on a grid sampling strategy.

[0009] For the downsampled point cloud in each layer, a local region is constructed around each point using a KD tree, and geometrically separable convolution is used to learn local features from this local region to increase the feature dimension of the point cloud.

[0010] By upsampling the point cloud features, point cloud features with the same number of input points are obtained, and then semantic segmentation of the indoor scene is performed.

[0011] Secondly, a semantic segmentation device for indoor scenes of service robots is provided, comprising:

[0012] An RGB-D point cloud data acquisition unit, mounted on the service robot, is used to acquire RGB-D point cloud data of the indoor scene;

[0013] A semantic segmentation network for RGB-D point clouds in indoor scenes of service robots includes an encoder and a decoder. The encoder is composed of a multilayer perceptron and L cascaded encoder modules. Each encoder module consists of a downsampling network and a geometrically separated convolutional network. The decoder is composed of L cascaded upsampling networks. The output of the l-th geometrically separated convolutional network is connected to the input of the (L+1)-l-th upsampling network.

[0014] The multilayer perceptron takes the collected RGB-D point cloud data as input and maps it to a high-dimensional space. The L encoder modules downsample the point cloud layer by layer based on a grid sampling strategy to reduce the number of points in the point cloud. For the downsampled point cloud in each layer, a local region is constructed around each point using a KD tree. The geometrically separable convolution learns local features from this local region to increase the feature dimension of the point cloud. The L upsampling networks interpolate the point cloud features to obtain point cloud features with the same number of input points for indoor scene semantic segmentation.

[0015] Thirdly, a semantic segmentation device for indoor scenes of service robots is provided, comprising:

[0016] A processor and a memory, the memory for storing computer programs, and the processor for calling and running the computer programs stored in the memory to perform the methods as described in the first aspect or other embodiments thereof.

[0017] Fourthly, a computer-readable storage medium is provided for storing a computer program that causes a computer to perform the methods as described in the first aspect or in various embodiments thereof.

[0018] Fifthly, a computer program product comprising instructions is provided, characterized in that, when the instructions are executed on a computer, the computer causes the computer to perform the method as described in the first aspect or in various embodiments thereof.

[0019] Compared with existing technologies, this invention uses grid sampling and geometrically separable convolution, thereby improving the accuracy and efficiency of semantic segmentation of RGB-D point cloud data in indoor service robot scenarios. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the process of this invention.

[0021] Figure 2 This is a schematic diagram of the network structure of the present invention.

[0022] Figure 3 This is a schematic diagram of the geometrically separated convolutional network structure of the present invention.

[0023] Figure 4 This is a segmentation example applying the method of the present invention. Detailed Implementation

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings and examples.

[0025] To address the semantic segmentation problem in indoor scenes for service robots, this invention proposes a semantic segmentation method for indoor scenes of service robots. This method employs raster sampling and geometrically separable convolution to effectively improve the accuracy and efficiency of semantic segmentation of RGB-D point cloud data. The principle and process are described below:

[0026] First, the collected RGB-D point cloud data is mapped to a high-dimensional space using a multilayer perceptron, and a grid sampling strategy is used to downsample the point cloud layer by layer to reduce the number of points in the point cloud.

[0027] This step uses a multilayer perceptron to map RGB-D point cloud data onto a regular grid structure to achieve point cloud feature upscaling. This can improve the overall semantic consistency while preserving the original geometric details, thereby enabling a better understanding of the global information in the RGB-D point cloud data.

[0028] Secondly, for the downsampled point cloud in each layer, a local region is constructed around each point using a KD tree, and geometrically separable convolution is used to learn local features from this local region, thereby increasing the feature dimension of the point cloud.

[0029] This step, by constructing local regions and geometrically separable convolutions, can accurately capture the local structure and features in RGB-D point cloud data, thereby improving the ability to model complex geometric relationships in indoor service robot scenarios.

[0030] Finally, the obtained point cloud features are interpolated using upsampling to restore the number of points in the point cloud, resulting in point cloud features with the same number of input points, which are then used for semantic segmentation of the indoor scene.

[0031] This invention utilizes raster sampling to improve the downsampling speed of the model and uses geometrically separable convolution to improve the model's ability to understand geometric and semantic information, thereby improving the accuracy and efficiency of semantic segmentation of RGB-D point clouds in indoor scenes.

[0032] This invention is based on an improved PointNet++ service robot indoor scene RGB-D point cloud semantic segmentation network, which can provide a more accurate and reliable RGB-D point cloud data semantic segmentation solution for applications such as autonomous driving, intelligent navigation and environmental perception.

[0033] Figure 1 The complete process of semantic segmentation of indoor scenes for service robots using the method of the present invention is shown, and the process is described as follows:

[0034] Step 1: Construct the above-mentioned semantic segmentation network for indoor scenes of service robots based on the improved PointNet++ RGB-D point cloud.

[0035] Specifically, the RGB-D point cloud semantic segmentation network for indoor service robot scenes involved in this invention mainly consists of an end-to-end improved PointNet++ deep learning network. The network structure is as follows: Figure 2 As shown. For input RGB-D point cloud data containing N points... The i-th point Includes its three-dimensional coordinates and RGB color information (r i ,g i ,b i ) T The network first uses a multilayer perceptron to map the RGB-D point cloud data P to a high-dimensional space, thereby increasing the dimensionality of the point cloud features. The features of the mapped RGB-D point cloud data are then... Where, p′ i p i The mapping.

[0036] Subsequently, multiple cascaded encoder modules are used to progressively downsample the data to reduce the number of points and increase their dimensionality. Correspondingly, the subsequent decoder consists of cascaded upsampled networks. The final output of the RGB-D point cloud semantic segmentation network for the service robot's indoor scene is... Where q represents the number of categories in the semantic segmentation of the RGB-D point cloud data. For each point in the output point cloud data... The category to which a feature belongs can be determined by finding the maximum value among the q features.

[0037] The encoder module of this invention consists of a downsampling network and a geometrically separable convolutional neural network (GSCNN). Unlike the commonly used farthest-point sampling, which has exponential time complexity, this invention proposes a raster downsampling method with linear time complexity, and uses this method to achieve the aforementioned downsampling. For the l-th layer encoder module, its input is the point cloud data output from the (l-1)-th layer encoder module. Where c represents the number of feature channels. This raster downsampling method converts point cloud data from n... l-1 The number of points is downsampled to n l (n l <n l-1 New point cloud data is obtained from ) points. Meanwhile, this invention designs a lightweight, geometrically separable convolution module, which enhances the network's ability to extract local information and accelerates the network's operation speed.

[0038] This invention utilizes voxelization technology to divide a large-scale point cloud into several smaller sub-point clouds, thereby reducing the GPU memory required for subsequent computations. Point cloud voxelization is the process of converting point cloud data into a discrete voxel mesh representation, which can transform disordered point cloud data into a structured and regular voxel mesh. This invention creates a voxel mesh within the three-dimensional space of the point cloud data. This mesh covers the entire point cloud region and ensures that each voxel overlaps with the point cloud data. Specifically, this method uses the FNV-1a hash algorithm to compress the discrete mesh into a unique identifier sequence.

[0039] The FNV-1a hash algorithm is a simple yet effective hash algorithm frequently used for fast hashing of data. First, an initial hash value is chosen, typically a small prime number; for example, a 32-bit prime number on a 32-bit system and a 64-bit prime number on a 64-bit system. This initial hash value is called `hash`. Next, each byte of the input data is iterated over. For each byte, the hash value is updated using the following formula:

[0040]

[0041] in, The XOR operation is represented by FNV_PRIME_64, which is a pre-selected constant of 1099511628211. After traversing all bytes, the final hash value is the obtained hash. The key to this algorithm is the use of a combination of XOR and multiplication. This allows each byte to influence the final hash value through a series of bitwise and multiplication operations, thus ensuring good distribution and hashing properties. After discretization, the point cloud data is randomly sampled. This method uses a loop to traverse the points in the voxels, then selects and randomly shuffles the indices according to certain rules, and finally adds the processed indices to a list, which is the downsampled point cloud. Specifically, the coordinates of the point cloud in continuous space are converted into discrete voxels. The FNV-1a algorithm is used to compress the voxel coordinates into unique identifier data. The identifier data is then traversed sequentially, and the point set corresponding to the first occurrence of each identifier data is taken as the downsampled point cloud data. Then, geometrically separated convolutional networks are used to expand the dimensionality of its features.

[0042] This invention employs the k-nearest neighbors (KNN) algorithm for local feature extraction. Specifically, for the downsampled point cloud data obtained from the l-th layer encoder module... Construct a local region around each point using a KD-tree, and from... The algorithm quickly finds the k nearest neighbors for each point. For the k nearest neighbor nodes found, geometrically separable convolution is used to extract local features, as shown in the following formula:

[0043]

[0044] in, This indicates the max pooling operation performed to maintain the rotation invariance of the point cloud. It is the j-th neighbor node corresponding to the i-th sampling point, w ij and b ij All of these are geometric information extracted from local coordinates, represented as:

[0045]

[0046]

[0047] MLP is a multilayer perceptron containing a linear layer, a batch normalized layer, and a ReLU nonlinear activation function. || represents the concatenation operation. These are the relative coordinates of the j-th neighbor node relative to the i-th sampling point. This operation is as follows: Figure 3 As shown.

[0048] This invention uses a decoder for interpolation and semantic segmentation. The decoder is composed of cascaded upsampling networks, which interpolate the point cloud features based on distance in the point cloud.

[0049] In this invention, the upsampling network and the encoder module have the same number of levels, both being L layers. Therefore: for the output... The l-th layer encoder module has nl points, each point having 2c features; the l-th layer encoder module and the input features are... The (L+1)-lth layer upsampling network corresponds to this. For example, in this embodiment of the invention, if L=4, then l=1, 2, 3 or 4.

[0050] a. utilizing and P l This corresponds to calculating the distance between any two points in three-dimensional space coordinates, resulting in a dimension n. l-1 ×n l Distance matrix It represents n l-1 Each point in (low dimension) is related to n l The distance to each point in (high-dimensional) space, represented by the element in the third row and fourth column of the matrix. The third point is far from P l The distance to the fourth point in the middle.

[0051] b. Sort the distance matrix Dist by its second dimension (i.e., the rows of the matrix), and retain the three smallest values ​​in each dimension (e.g., each row) to obtain the matrix. That is Neutral P l The distance of each point to its three nearest points is represented by the element in the third row and second column of the dist array. Neutral P l The distance between the third point and the second nearest point is recorded using idx. The corresponding index.

[0052] c. Use IDX to extract point cloud features A query is performed to obtain a size of n. l-1 The eigenmatrix F of ×3×4c is used to... Deeper neighborhood features are aggregated into P l middle.

[0053] d. To use the distance as a weight to weight the feature matrix F, the dist matrix is ​​copied 4c times along its columns. Using matrices The features are obtained by weighting the feature matrix F.

[0054]

[0055] Where ⊙ represents Hadamard convolution.

[0056] e. Using a multilayer perceptron to... After dimensionality reduction and P l Add them together to obtain the output of the decoder at layer (L+1)-l. For the output of the last upsampling layer Dimensionality reduction using multilayer perceptron

[0057] Step 2: Train an RGB-D point cloud semantic segmentation network using the RGB-D point cloud dataset of the service robot's indoor scene.

[0058] Step 21: Preprocessing of RGB-D point cloud data.

[0059] For the RGB-D point cloud dataset D = {P1, P2, ..., P} of an indoor service robot scene, m},in Each RGB-D point cloud P i various points in the middle The spatial coordinates are converted to positive values, that is...

[0060]

[0061]

[0062]

[0063] Where, x i,min ,y i,min ,z i,min These are the point cloud P. i The minimum value in the three coordinate directions.

[0064] Thus, the preprocessed RGB-D point cloud dataset D′={P′1,P′2,...,P′ is obtained. m},in

[0065] Step 22: Train the RGB-D point cloud semantic segmentation network.

[0066] The preprocessed RGB-D point cloud dataset D′ is divided into a training set and a validation set in a 7:3 ratio. The model is trained using the training set, and the weights of the best-performing model on the validation set are saved to complete the training of the RGB-D point cloud semantic segmentation network.

[0067] Step 3: Perform semantic segmentation on the RGB-D point cloud data of the indoor scene of the service robot to be segmented.

[0068] For the RGB-D point cloud of the indoor scene of the service robot to be segmented Preprocessing is performed on the point cloud E′ following a similar procedure to step 21 to obtain the preprocessed RGB-D point cloud. The RGB-D point cloud semantic segmentation network trained in step 2 is then used to perform semantic segmentation on E′, yielding the network output. For each point in the point cloud The maximum value of each of the q elements is obtained by indexing its category, thus achieving semantic segmentation of RGB-D point clouds. Figure 4 For the segmentation case, the left side represents the input point cloud data, and the right side represents the instance segmentation result.

[0069] This invention also provides a semantic segmentation device for indoor scenes of a service robot. The device may include: an RGB-D point cloud data acquisition unit and a service robot indoor scene RGB-D point cloud semantic segmentation network. The RGB-D point cloud data acquisition unit, mounted on the service robot, is used to acquire RGB-D point cloud data of the indoor scene, and this acquisition can be real-time. The service robot indoor scene RGB-D point cloud semantic segmentation network takes the acquired RGB-D point cloud data as input and executes the semantic segmentation method of this invention. Another structure is also as described. Figure 2As shown, it mainly includes an encoder and a decoder. The encoder consists of a multilayer perceptron and four cascaded encoder modules. Each encoder module is further composed of a downsampling network and a geometrically separated convolutional network. The decoder consists of four cascaded upsampling networks; the output of the l-th geometrically separated convolutional network is connected to the input of the (5-l)-th upsampling network, so that the decoder can interpolate the point cloud features based on distance in the point cloud.

[0070] In one embodiment, the RGB-D point cloud semantic segmentation network for indoor scenes of the service robot of the present invention can realize the aforementioned semantic segmentation method. The multilayer perceptron implements high-dimensional mapping of the RGB-D point cloud data. The encoder module, on the one hand, downsamples layer by layer based on a raster sampling strategy, and on the other hand, learns local features from the local regions constructed around each point, increasing the feature dimension of the point cloud. The decoder then interpolates the point cloud features through upsampling to obtain point cloud features with the same number of input points, thus performing semantic segmentation of the indoor scene. It should be understood that the device embodiment and the method embodiment can correspond to each other, and similar descriptions can be referred to in the method embodiment. To avoid repetition, further details are omitted here.

[0071] The apparatus described above in the embodiments of this application can be implemented in hardware, in software instructions, or in a combination of hardware and software modules. Specifically, the steps of the method embodiments in this application can be completed by integrated logic circuits in the processor's hardware and / or by software instructions. The steps of the method disclosed in the embodiments of this application can be directly embodied as being executed by a hardware encoding processor, or executed by a combination of hardware and software modules in the encoding processor. Optionally, the software module can be located in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. This storage medium is located in memory, and the processor reads information in the memory and completes the steps in the above method embodiments in conjunction with its hardware.

[0072] The semantic segmentation device for indoor scenes of service robots provided by the present invention may include: a processor and a memory, the memory being used to store computer programs, and the processor being used to call and run the computer programs stored in the memory to execute the semantic segmentation method of the present invention.

[0073] The present invention provides a computer-readable storage medium capable of storing a computer program that causes a computer to execute the semantic segmentation method of the present invention.

[0074] The present invention provides a computer program product containing instructions that, when executed on a computer, enable the computer to perform the semantic segmentation method of the present invention.

[0075] The above are merely specific embodiments of this application, but the protection scope of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of this application should be included within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A semantic segmentation method for indoor scenes of service robots, characterized in that, include: RGB-D point cloud data is collected, mapped to a high-dimensional space using a multilayer perceptron, and the number of points in the point cloud is reduced by downsampling the point cloud layer by layer based on a grid sampling strategy. For the downsampled point cloud in each layer, a local region is constructed around each point using a KD tree, and a geometrically separable convolutional network is used to learn local features from this local region to increase the feature dimension of the point cloud. Upsampling is used to interpolate the point cloud features to obtain point cloud features with the same number of input points, and then semantic segmentation of the indoor scene is performed. The downsampling is implemented using a grid downsampling method with linear time complexity. For the th l The layer encoder module, whose input is the first layer encoder module. l -1 layer encoder module output feature channels number of c Point cloud data Raster downsampling from Downsampling at each point New point cloud data is obtained from each point. , ; For the l Downsampled point cloud data obtained by the layer encoder module Construct a local region around each point using a KD-tree, and search for each point within it. k For the nearest neighbor points found, k For each neighboring node, local features are extracted using the geometrically separable convolutional network, as shown in the following formula: in, This indicates a max pooling operation. It is the first i The corresponding sampling point of the th sampling point j One neighboring node, and All of these are geometric information extracted from local coordinates, represented as: Here, MLP is a multilayer perceptron containing a linear layer, a batch normalization layer, and a ReLU nonlinear activation function, and || denotes the concatenation operation. They are the first j The neighbor node relative to the first i The relative coordinates of each sampling point.

2. The semantic segmentation method for indoor scenes of service robots according to claim 1, characterized in that, For containing N RGB-D point cloud data of each point First, a multilayer perceptron is used to map it to a high-dimensional space. The features after mapping are: ,in, express The first in i One point, express The mapping, , express The three-dimensional coordinates express RGB color information; Multiple cascaded encoder modules are used to progressively downsample to reduce the number of points and increase the dimension of the points; each encoder module consists of a downsampling network and a geometrically separated convolutional network.

3. The semantic segmentation method for indoor scenes of service robots according to claim 2, characterized in that, The method of raster downsampling is as follows: The coordinates of the point cloud in continuous space are converted into discrete voxels. The FNV-1a algorithm is used to compress the voxel coordinates into unique identifier data. The identifier data is then iterated through, and the point set corresponding to the first occurrence of each identifier data is taken as the downsampled point cloud data. Then, geometrically separated convolutional networks are used to expand the dimensionality of their features into... .

4. The semantic segmentation method for indoor scenes of service robots according to claim 1 or 3, characterized in that, Interpolation and semantic segmentation are performed through a decoder, which is composed of cascaded upsampling networks and interpolates the point cloud features based on distance in the point cloud. Both the upsampling network and the encoder module are L For layers, then: for output as The l Layer encoder module, which has There are 2 points, each point has 2 c One feature; No. l The layer encoder modulus and input features are The ( L +1)- l Corresponding to the layer sampling network; a. Utilize and This corresponds to calculating the distance between any two points in three-dimensional space coordinates, resulting in a distance matrix. , which means Each point in (low dimension) and The distance to each point in (high dimension); b. Distance matrix Sort by the second dimension, and retain the three smallest values ​​in each dimension to obtain the matrix. That is Neutral The distance of each point to its three nearest points, and using Record its in The corresponding index; c. Use go A query was performed to obtain a size of The feature matrix; d. Using matrices The features are obtained by weighting the feature matrix. ; e. Using multilayer perceptrons to... After dimensionality reduction and Add them together to get the decoder's ( ) L +1)- l Layer output .

5. The semantic segmentation method for indoor scenes of service robots according to claim 4, characterized in that, For the output of the last upsampling layer Dimensionality reduction using multilayer perceptrons ,in q This represents the number of categories for semantic segmentation of RGB-D point cloud data, for each point in the point cloud data. of q The maximum value of each feature is used to determine its category.

6. A semantic segmentation device for indoor scenes of a service robot, used to implement the semantic segmentation method for indoor scenes of a service robot as described in claim 1, characterized in that, include: An RGB-D point cloud data acquisition unit, mounted on the service robot, is used to acquire RGB-D point cloud data of the indoor scene; A semantic segmentation network for RGB-D point clouds in indoor scenes for service robots includes an encoder and a decoder. The encoder consists of a multilayer perceptron and L cascaded encoder modules, each encoder module consisting of a downsampling network and a geometrically separated convolutional network. The decoder consists of L cascaded upsampling networks. l The output of the geometrically separated convolutional network and the L+1th... l The inputs of the upsampling network are connected; The multilayer perceptron takes the collected RGB-D point cloud data as input and maps it to a high-dimensional space. The L encoder modules downsample the point cloud layer by layer based on a grid sampling strategy to reduce the number of points in the point cloud. For the downsampled point cloud in each layer, a local region is constructed around each point using a KD tree. The geometrically separable convolutional network learns local features from this local region to increase the feature dimension of the point cloud. The L upsampling networks interpolate the point cloud features to obtain point cloud features with the same number of input points for indoor scene semantic segmentation.

7. A semantic segmentation device for indoor scenes of service robots, characterized in that, include: A processor and a memory, the memory being used to store a computer program, the processor being used to invoke and run the computer program stored in the memory to perform the method of any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, Used to store a computer program that causes a computer to perform the method as described in any one of claims 1-5.

9. A computer program product containing instructions, characterized in that, When the instructions are executed on a computer, the computer causes the computer to perform the method of any one of claims 1-5.

Images