Cross-domain three-dimensional point cloud semantic segmentation method and system based on active learning strategy

Abstract

The invention relates to an image processing technology, in particular to a cross-domain three-dimensional point cloud semantic segmentation method and system based on an active learning strategy. The method comprises the steps that a semantic segmentation model is established and initialized, wherein the semantic segmentation model comprises a feature extractor and a classifier; training the semantic segmentation model, and optimizing parameters of the semantic segmentation model; carrying out region division on the point cloud data of the source domain and the target domain, calculating the comprehensive score of each region, and marking the region with the higher score; marking the point cloud data of the target domain region by adopting a similarity punishment mechanism so as to select diversified regions for marking, updating the data set, and retraining the semantic segmentation model; and iteratively updating the semantic segmentation model, and processing the three-dimensional point cloud data by using the final semantic segmentation model. According to the method, the demand of target domain annotation data is reduced, and the segmentation performance of the model in the target domain is improved.

Description

Technical Field

[0001] The present invention relates to image processing technology, and specifically to a cross-domain three-dimensional point cloud semantic segmentation method and system based on active learning strategy. Background Art

[0002] Point cloud semantic segmentation is an important part of analyzing LiDAR 3D point cloud data. In the field of point cloud semantic segmentation, a variety of methods have emerged.

[0003] Before the application of machine learning, clustering algorithms were generally used for semantic segmentation of point clouds, such as K-means, DBSCAN, and hierarchical clustering algorithms. In existing point cloud semantic segmentation methods, clustering algorithms are often used as segmentation tools for point cloud regions and as excellent tools for pre-processing point clouds in conjunction with machine learning. However, clustering algorithms have many defects in processing point cloud data that lack prior knowledge and point cloud data with irregular shapes.

[0004] Machine learning algorithms used for plane or two-dimensional data processing are difficult to be directly applied to point cloud semantic segmentation. Point cloud data itself is highly discrete and does not have the characteristic of known nearest neighbor elements like images; this results in a significant difference in the morphology of point cloud data from different perspectives in the same area. Models trained using simple neural networks are unlikely to have good generalization capabilities and cannot perform reliable semantic segmentation from different perspectives. This problem has become a major obstacle to the large-scale application of machine learning in three-dimensional point cloud semantic segmentation.

[0005] The Pointnet and Pointnet++ algorithms that appeared around 2017 successfully solved this problem through structural optimization. Pointnet extracts global and local location information by linearly combining each point vector of the point cloud with other point vectors. At the same time, the K nearest neighbor method derived from the graph convolutional neural network provides a relatively simple and reliable solution for point cloud semantic segmentation. However, this method of point cloud segmentation is computationally intensive and the parameter learning method is relatively rigid.

[0006] In 2019, the rise of Transformer provided a unique way to perform semantic segmentation. Applying the attention mechanism of Transformer can make semantic segmentation of point clouds more efficient and accurate, and can also collect more location information and point cloud structure information at a point. The method using Transformer is superior to the traditional CNN-based method in terms of computational complexity and segmentation accuracy.

[0007] The deep learning methods mentioned above can achieve acceptable accuracy within a domain with sufficient data. However, the application scenarios of laser point clouds usually involve data from multiple domains, especially models used in autonomous driving. If manually annotated datasets from multiple domains are used, a large amount of expensive manual annotation is required. Choosing to improve the domain adaptation characteristics of the model is obviously a cheaper method. The domain adaptation characteristics of ordinary machine learning methods are not excellent, so it is necessary to optimize the domain adaptation of the point cloud semantic segmentation machine learning method.

[0008] Therefore, methods for domain adaptation optimization have emerged. DANN is a method that improves the domain adaptation characteristics of the model in the form of a generative adversarial network. By adding a domain classification indicator to the model, the difference between the two domains in the model is narrowed. By reducing the domain classification error, the model can obtain good domain adaptation properties. There is also a class of methods that improves the domain adaptation ability of the model by combining data from multiple modalities for comprehensive analysis. Xmuda chooses to combine 3D point cloud and 2D image data, which significantly improves the generalization ability of the model.

[0009] In addition to 3D point clouds, active learning is a method that can reduce the amount of data required for model training when there is insufficient data set or insufficient annotation resources. Active learning uses a specific strategy to select a portion of the data that is most effective in improving the accuracy of the model and then processes it for training. There are many active learning strategies, from the simplest random selection to the ReDAL method that combines color discontinuity, structural complexity, and softmax entropy. Choosing an excellent active learning strategy can significantly reduce the training cost of the model. Summary of the invention

[0010] In order to solve the technical problems existing in the prior art, the present invention provides a cross-domain three-dimensional point cloud semantic segmentation method and system based on an active learning strategy, introduces a region-based segmentation strategy, selects a small number of the most informative regions in the target domain for annotation, significantly reduces the demand for target domain annotation data, and improves the segmentation performance of the model in the target domain.

[0011] In an embodiment of the present invention, the technical solution of the method is: a cross-domain three-dimensional point cloud semantic segmentation method based on an active learning strategy, comprising the following steps:

[0012] S1. Establish a semantic segmentation model, which includes a feature extractor and a classifier; the feature extractor is used to extract descriptive features from point cloud data and generate feature representations of point cloud data; the classifier is used to map the feature representations of point cloud data to the category probability distribution of point cloud data; initialize the feature extractor and the classifier;

[0013] S2. Train the semantic segmentation model and optimize the parameters of the semantic segmentation model;

[0014] S3, divide the point cloud data of the source domain and the target domain into regions, calculate the comprehensive score of each region, and mark the regions with higher scores;

[0015] S4. Use a similarity penalty mechanism to annotate the point cloud data in the target domain area to select diverse areas for annotation; integrate the annotated data into the source domain data, update the data set, and retrain the semantic segmentation model;

[0016] S5. Iteratively update the semantic segmentation model to obtain a final semantic segmentation model; and use the final semantic segmentation model to process the three-dimensional point cloud data.

[0017] In an embodiment of the present invention, the system is implemented based on the above-mentioned cross-domain three-dimensional point cloud semantic segmentation method, including:

[0018] Data acquisition module, used to acquire point cloud data;

[0019] The semantic segmentation model is used to perform semantic segmentation on the acquired point cloud data.

[0020] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0021] 1. Traditional 3D point cloud semantic segmentation methods usually rely on a large amount of labeled data and perform poorly when the source domain and target domain feature differences are large. This paper applies active learning strategies to cross-domain 3D point cloud semantic segmentation tasks for the first time. By selecting a small number of the most informative areas in the target domain for labeling, the demand for labeled data in the target domain is significantly reduced, and the segmentation performance of the model in the target domain is improved. This innovation solves the problem of performance degradation caused by the lack of labeled data in current cross-domain 3D point cloud tasks.

[0022] 2. Traditional point cloud segmentation methods usually perform annotation and selection based on the global scene. This invention introduces a region-based segmentation strategy that divides the point cloud scene into multiple regions and uses these regions as the minimum annotation unit. Each region is independently evaluated to determine its potential contribution to improving model performance. Through this regionalization process, not only the accuracy of annotation is improved, but also unnecessary data redundancy is reduced, thereby significantly reducing the annotation cost.

[0023] 3. In the annotation selection process, the present invention proposes a similarity penalty mechanism to avoid selecting multiple regions with similar features. This mechanism detects regions with similar features through a clustering algorithm and imposes penalties on the scores of these regions, thereby giving priority to more diverse regions for annotation. This similarity penalty strategy ensures that the selected data has a higher amount of information and avoids data redundancy and annotation waste.

[0024] 4. In order to evaluate the annotation value of each area, the present invention adopts a comprehensive multi-dimensional scoring mechanism. The scoring mechanism combines four key indicators: Softmax entropy, Maximum Mean Discrepancy (MMD), color discontinuity and structural complexity. These indicators evaluate the uncertainty, inter-domain differences and information volume of the region from different angles. Through comprehensive scoring, the most valuable areas for model improvement can be selected more accurately. The application of this multi-dimensional scoring mechanism in cross-domain 3D point cloud tasks has greatly improved the annotation efficiency and model accuracy.

[0025] 5. The present invention significantly improves the cross-domain segmentation performance of the model by annotating only a small amount of data in the target domain. In the experiment, using less than 10% of the target domain annotated data, significantly higher segmentation accuracy was achieved than the traditional method. This result shows that the present invention can achieve efficient cross-domain semantic segmentation with extremely low annotation cost, and is suitable for various complex cross-domain scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flow chart of a cross-domain three-dimensional point cloud semantic segmentation method in an embodiment of the present invention;

[0027] Figure 2 It is a schematic diagram of comprehensive scoring and labeling of target domain data in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] This paper proposes a cross-domain 3D point cloud semantic segmentation method and system based on active learning strategy, aiming to improve model performance and reduce annotation cost by intelligently selecting key data areas in the target domain. The semantic segmentation method mainly includes preliminary model training, regional feature extraction, regional scoring mechanism, cluster analysis and similarity penalty mechanism, as well as annotation and model iteration steps.

[0029] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0030] Example

[0031] like Figure 1As shown, this embodiment provides a cross-domain three-dimensional point cloud semantic segmentation method based on an active learning strategy, comprising the following steps:

[0032] S1. Establish a semantic segmentation model, which includes a feature extractor and a classifier; the feature extractor is used to extract descriptive features from point cloud data and generate feature representation of point cloud data; the classifier is used to map the feature representation of point cloud data to the category probability distribution of point cloud data; initialize the feature extractor and the classifier.

[0033] This embodiment uses an open source model to perform semantic segmentation on point cloud data, and the model mainly includes a feature extractor and a classifier. Among them, the classifier is a linear layer.

[0034] In this embodiment, a feature extractor is used to extract descriptive features from point cloud data to provide input for subsequent classifiers. The feature extractor can be abstracted as a function Where x is the input point cloud data, including the geometric coordinates (x, y, z) of each point cloud data. is a parameter that controls the feature extraction process. The point cloud data x is mapped into a high-dimensional feature space to generate feature representation Used to describe the geometric structure and global contextual semantic information of point cloud data.

[0035] To achieve this goal, this embodiment uses the open source architecture SPVCNN as a specific technical means of implementing the feature extractor. SPVCNN is an efficient neural network model for point cloud data, which can efficiently extract local and global features of point cloud data through sparse convolution. Its core mechanism includes: first, dividing the sparse point cloud data into voxel grids to capture the local geometric relationship of the point cloud data; then, extracting multi-scale features on voxel grids of different resolutions through sparse convolution operations; finally, further integrating these multi-scale features to generate a descriptive feature vector for each point cloud data. Mathematically, SPVCNN can be described in the following form:

[0036]

[0037] Among them, SPVCNN represents the operation logic of sparse convolutional network; To control the parameters of the feature extraction process, specifically, parameters such as convolution kernel weight, bias, and sparse matrix configuration may be included; d represents the dimension of the feature vector.

[0038] During the initialization phase, the parameters of the feature extractor It has not been optimized, but initialized by a standardized random method. The weights of the convolution kernel are usually initialized to random values ​​that obey the standard normal distribution to ensure that there is no initial bias in feature extraction in all directions; the bias parameters are initialized to zero to avoid artificial offsets in the feature generation process. At the same time, the sparse matrix configuration of SPVCNN is dynamically generated according to the initial distribution of the point cloud data, and is constructed by calculating the proximity relationship between points and voxels.

[0039] The classifier is used to represent the features output by the feature extractor. Predict the category of point cloud data. The classifier can be abstracted as a function in are the parameters of the classifier; the input of the classifier is the feature representation of the point cloud data obtained from the feature extractor; the output of the classifier is the probability distribution of the point cloud data belonging to each category, that is, Where k is the total number of categories, p i Represents the probability that the point cloud data belongs to the i-th category, satisfying In the initialization phase, the parameters of the classifier The weights are also set randomly, with the weights initialized to a small range of random values, and the biases are initialized to zero to ensure that the classifier has no preference for all categories in the initial state. Initially, the probability distribution of the classifier output is uniformly distributed:

[0040]

[0041] By initializing the feature extractor and the classifier, the system of this embodiment can start processing point cloud data in a neutral state and lay the foundation for subsequent parameter optimization. The feature extractor provides high-dimensional feature representation of point cloud data, while the classifier maps these feature representations into the category probability distribution of point cloud data, thereby completing the preliminary point cloud classification task.

[0042] S2. Train the semantic segmentation model and optimize the parameters of the semantic segmentation model.

[0043] The model training process is done through the source domain and target domain The specific steps are as follows:

[0044] S21. Source domain data and target domain data Input the feature extractor SPVCNN to extract source domain features and target domain features

[0045] The parameters of the feature extractor are gradually adjusted through a subsequent optimization process so that the features it extracts can better describe the geometric structure and global contextual semantic information of the point cloud data.

[0046] S22. Calculate source domain features and target domain features The maximum mean difference (MMD) between them is used as the feature alignment loss.

[0047] The feature alignment loss aims to minimize the difference between the source domain features and the target domain features, thereby improving the adaptability of the semantic segmentation model to the target domain point cloud data. The feature alignment loss can be calculated by the following function Calculate the formula to obtain:

[0048]

[0049] in is the distribution of source domain features, is the distribution of target domain features.

[0050] S23. Source domain features Input classifier to predict the category probability distribution of source domain data Compare the probability of the category to which the point cloud data predicted by the classifier belongs with the true label of the source domain Compare and calculate the cross entropy loss As the classification loss, to measure the accuracy of the classification results:

[0051]

[0052] Where k is the number of categories, Representing point cloud data The true label belongs to category j (if it belongs to Otherwise the true label is 0); Predicting point cloud data for the classifier The probability of belonging to class j.

[0053] S24. Combine feature alignment loss and classification loss to construct an overall loss function, specifically:

[0054]

[0055] Where λ is a trade-off coefficient, which is used to balance the relative importance of classification loss and feature alignment loss.

[0056] S25. By optimizing the overall loss function Parameters for the feature extractor and the parameters of the classifier to update.

[0057] In this embodiment, the parameter update rule is:

[0058]

[0059] Where η is the learning rate, which is used to control the step size of parameter update.

[0060] Through the above training steps, the feature extractor and classifier of the semantic segmentation model can gradually adapt to the characteristics of the source domain data, while improving the generalization ability of the target domain data through feature alignment.

[0061] S3. Divide the point cloud data of the source domain and the target domain into regions, calculate the comprehensive score of each region, and mark the regions with higher scores.

[0062] After the initial training of the model is completed, the point cloud data of the target domain and the source domain are divided into regions, and the comprehensive scores of the regions are calculated, which specifically includes the following steps:

[0063] S31. Use the VCCS (Voxel Cloud Connectivity Segmentation) algorithm to divide the point cloud data of the source domain and the target domain into multiple regions. and in is the nth region of the source domain, is the nth region of the target domain.

[0064] These regions will serve as basic annotation units, and the point cloud data within each region will be analyzed and selected as a whole to calculate a comprehensive score for each region.

[0065] S32. Use the trained semantic segmentation model to predict each point cloud data in the target domain, and convert the prediction output into a probability distribution through the Softmax function. The Softmax entropy value is calculated based on the converted probability distribution to measure the annotation value of the region.

[0066] The Softmax entropy value is calculated by the following formula:

[0067]

[0068] Softmax entropy is used to measure the prediction uncertainty of the semantic segmentation model for each region. The higher the entropy value, the higher the annotation value of the region.

[0069] S33. The maximum mean difference (MMD) is used to calculate the feature difference between the source domain and the target domain. The formula is as follows:

[0070]

[0071] in, Used to calculate the target domain area and source domain area The distance between them is used to measure the importance of the region to the cross-domain model adaptability. The calculation formula is as follows:

[0072]

[0073] in, To transform the target domain features Mapped to high-dimensional Hilbert space, To transform the source domain features Map to the high-dimensional Hilbert space, and then calculate the L2 norm of the target domain features and the source domain features in the mapped Hilbert space. H represents the Hilbert space.

[0074] S34. For each target domain area Compute color discontinuities.

[0075] First, calculate the color difference between each point cloud data in the target domain and its adjacent point cloud data. Regions with large color differences usually contain more semantic information. The specific calculation formula for color difference is:

[0076]

[0077] in Represents the distance target domain data The set of the nearest k points.

[0078] Color discontinuity for each target domain region The average value of the color difference of each point cloud data in the area:

[0079]

[0080] S35. The surface change rate of each region in the target domain is calculated to evaluate the structural complexity of the region.

[0081] The surface change rate is used to measure complex boundary areas and surface irregularities. The specific calculation formula is:

[0082]

[0083] in, Indicate point The local surface change rate.

[0084] S36. Weighted sum of Softmax entropy value, feature difference, color discontinuity, and surface change rate to obtain a comprehensive score for each area:

[0085]

[0086] Among them, α, β, γ, and δ are all weight coefficients.

[0087] This comprehensive score is used to rank the importance of regions, and regions with higher scores are selected for annotation first.

[0088] S4. Use a similarity penalty mechanism to annotate the point cloud data in the target domain area to select diverse areas for annotation; integrate the annotated data into the source domain data, update the dataset, and retrain the semantic segmentation model.

[0089] In the annotation selection process, this paper proposes a similarity penalty mechanism to avoid selecting multiple regions with similar features. This mechanism detects regions with similar features through a clustering algorithm and imposes a penalty in the scores of these regions, thereby giving priority to selecting more diverse regions for annotation.

[0090] The basic unit of annotation in this embodiment is "region" rather than a single point cloud data. By dividing the point cloud data into multiple regions, the workload of annotating data is significantly reduced and the annotation efficiency is improved. In order to avoid repeated selection of similar regions in the same annotation selection, this embodiment designs a similarity penalty mechanism. The specific steps are as follows:

[0091] S41. Extract features from each target domain region, and use a clustering algorithm (such as K-means) to cluster the regions, and classify regions with similar features into the same category.

[0092] S42: For regions belonging to the same cluster, select them according to the score ranking, but for regions with lower rankings, apply a decay factor d k-1 Reduce its score, where k means that the region is the kth region of the cluster to which it belongs sorted from high to low according to the score. In this way, it is ensured that the final selected regions are diverse and avoid selecting similar regions.

[0093] S43. Manually annotate the top K regions with the highest scores, integrate the annotated data into the source domain data, retrain the semantic segmentation model using the updated data set, and update the parameters of the semantic segmentation model.

[0094] S5. Iteratively update the semantic segmentation model to obtain a final semantic segmentation model; and use the final semantic segmentation model to process the three-dimensional point cloud data.

[0095] Repeat steps S2-S4 until the annotation cost cap is reached or the model performance reaches the target requirement.

[0096] This embodiment also provides a cross-domain 3D point cloud semantic segmentation system based on an active learning strategy, characterized in that it is implemented based on the above-mentioned cross-domain 3D point cloud semantic segmentation method and includes the following modules:

[0097] Data acquisition module, used to acquire point cloud data;

[0098] The semantic segmentation model is used to perform semantic segmentation on the acquired point cloud data.

[0099] Experimental results: In the S3DIS and SemanticKITTI datasets, the segmentation accuracy mIoU of the target domain point cloud at different annotation scales is shown in Table 1.

[0100] Table 1

[0101]

[0102] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A cross-domain 3D point cloud semantic segmentation method based on active learning strategy, characterized in that: The following steps are involved: S1. Establish a semantic segmentation model, which includes a feature extractor and a classifier; the feature extractor is used to extract descriptive features from point cloud data and generate feature representations of point cloud data; the classifier is used to map the feature representations of point cloud data to the category probability distribution of point cloud data; initialize the feature extractor and the classifier; S2. Train the semantic segmentation model and optimize the parameters of the semantic segmentation model; S3, divide the point cloud data of the source domain and the target domain into regions, calculate the comprehensive score of each region, and mark the regions with higher scores; S4. Use a similarity penalty mechanism to annotate the point cloud data in the target domain area to select diverse areas for annotation; Integrate the labeled data into the source domain data, update the dataset, and retrain the semantic segmentation model; S5. Iteratively update the semantic segmentation model to obtain a final semantic segmentation model; and use the final semantic segmentation model to process the three-dimensional point cloud data.

2. The cross-domain 3D point cloud semantic segmentation method according to claim 1, characterized in that: The feature extractor is represented as a function Where x is the input point cloud data, including the geometric coordinates (x, y, z) of each point cloud data. is a parameter that controls the feature extraction process; through the function The point cloud data x is mapped into a high-dimensional feature space to generate feature representation Used to describe the geometric structure and global contextual semantic information of point cloud data.

3. The cross-domain 3D point cloud semantic segmentation method according to claim 2, characterized in that: The feature extractor is implemented using the open source architecture SPVCNN, which extracts local and global features of point cloud data through sparse convolution; First, the sparse point cloud data is divided into voxel grids to capture the local geometric relationships of the point cloud data. Then, multi-scale features are extracted on voxel grids of different resolutions through sparse convolution operations. Finally, the multi-scale features are further integrated to generate a descriptive feature vector for each point cloud data.

4. The cross-domain 3D point cloud semantic segmentation method according to claim 3, characterized in that: Including the configuration of convolution kernel weights, biases, and sparse matrices.

5. The cross-domain 3D point cloud semantic segmentation method according to claim 1, characterized in that: The classifier is represented as a function in are the parameters of the classifier; the input of the classifier is the feature representation of the point cloud data obtained from the feature extractor; the output of the classifier is the probability distribution of the point cloud data belonging to each category, that is, Where k is the total number of categories, p i Represents the probability that the point cloud data belongs to the i-th category.

6. The cross-domain 3D point cloud semantic segmentation method according to claim 1, characterized in that: Step S2 includes: S21, inputting the source domain data and the target domain data into the feature extractor respectively to extract source domain features and target domain features; S22, calculate the maximum mean difference between the source domain features and the target domain features as the feature alignment loss; S23, input the source domain features into the classifier to predict the category probability distribution of the source domain data; compare the probability of the category to which the point cloud data predicted by the classifier belongs with the true label of the source domain, and calculate the cross entropy loss as the classification loss; S24, integrating feature alignment loss and classification loss to construct an overall loss function; S25. Update the parameters of the feature extractor and the classifier by optimizing the overall loss function.

7. The cross-domain 3D point cloud semantic segmentation method according to claim 1, characterized in that: Step S3 includes: S31. Use the VCCS algorithm to divide the point cloud data of the source domain and the target domain into multiple regions. and in is the nth region of the source domain, is the nth region of the target domain; S32. Use the trained semantic segmentation model to predict each point cloud data in the target domain, and convert the prediction output into a probability distribution through the Softmax function. Calculate the Softmax entropy value based on the converted probability distribution; S33, using the maximum mean difference to calculate the feature difference between the source domain and the target domain; S34. For each target domain area Calculate color discontinuities; S35, evaluating the structural complexity of each region in the target domain by calculating the surface change rate of the region; S36. The Softmax entropy value, feature difference, color discontinuity, and surface change rate are weighted and summed to obtain a comprehensive score for each area.

8. The cross-domain 3D point cloud semantic segmentation method according to claim 7, characterized in that: Feature Difference in Step S33 The calculation formula is: in, Used to calculate the target domain area and source domain area The distance between in, To transform the target domain features Mapped to high-dimensional Hilbert space, To transform the source domain features Map to high-dimensional Hilbert space, and then calculate the L2 norm of the target domain features and the source domain features in the mapped Hilbert space, where H represents the Hilbert space; The calculation process of color discontinuity in step S34 is: First, the color difference between each point cloud data in the target domain and its adjacent point cloud data is calculated; the color discontinuity of each target domain area is the average value of the color difference of each point cloud data in the area.

9. The cross-domain 3D point cloud semantic segmentation method according to claim 1, characterized in that: Step S4 includes: S41, extracting features from each target domain region, and clustering the regions using a clustering algorithm, classifying regions with similar features into the same category; S42: For regions belonging to the same cluster, select them according to the score ranking, but for regions with lower rankings, apply a decay factor d k-1 Reduce its score, where k means that the region is the kth region of the cluster to which it belongs sorted from high to low in terms of score; S43. Manually annotate the top K regions with the highest scores, integrate the annotated data into the source domain data, retrain the semantic segmentation model using the updated data set, and update the parameters of the semantic segmentation model.

10. A cross-domain 3D point cloud semantic segmentation system based on active learning strategy, characterized in that: The method according to any one of claims 1 to 9 is implemented, comprising the following modules: Data acquisition module, used to acquire point cloud data; The semantic segmentation model is used to perform semantic segmentation on the acquired point cloud data.

Images