Point cloud semantic segmentation method and system based on intra-voxel cohesion guided active learning

Abstract

The application provides a point cloud semantic segmentation method and system based on voxel cohesion guided active learning, which comprises the following steps: obtaining target three-dimensional point cloud data corresponding to a scene to be detected; inputting the target three-dimensional point cloud data into a point cloud semantic segmentation model to obtain a semantic segmentation category result corresponding to the target three-dimensional point cloud data output by the point cloud semantic segmentation model, wherein the point cloud semantic segmentation model is obtained by training sample three-dimensional point cloud data labeled with a semantic segmentation category label, the sample three-dimensional point cloud data is point cloud data in a sample voxel grid determined based on a voxel cohesion degree, and the voxel cohesion degree is determined according to the number of data points corresponding to point cloud data points of different semantic segmentation category labels in the sample voxel grid. The application improves the accuracy and generalization ability of the point cloud semantic segmentation result.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a point cloud semantic segmentation method and system based on voxel cohesion guided active learning. Background Art

[0002] Understanding three-dimensional scenes is a key research area in modern computer vision and intelligent systems, playing a crucial role in accurate recognition and decision-making in complex environments. Comprehensively analyzing the structure, shape, and position of objects in three-dimensional scenes provides fundamental support for research in areas such as autonomous driving, robotic navigation, and virtual reality.

[0003] Point cloud semantic segmentation, one of the core technologies for achieving 3D scene understanding, aims to assign specific semantic labels to each point in point cloud data, thereby achieving comprehensive scene perception and classification. Existing point cloud semantic segmentation methods primarily rely on data-driven deep learning methods, using large labeled datasets to train high-precision segmentation models. However, the sparsity and irregularity of point cloud data, as well as the high cost of labeling, significantly limit the performance of segmentation models, hindering their application in practical applications.

[0004] Therefore, there is an urgent need for a point cloud semantic segmentation method and system based on voxel cohesion-guided active learning to solve the above problems. Summary of the Invention

[0005] In response to the problems existing in the prior art, the present invention provides a point cloud semantic segmentation method and system based on voxel cohesion guided active learning.

[0006] The present invention provides a point cloud semantic segmentation method based on voxel cohesion guided active learning, comprising:

[0007] Obtain the target three-dimensional point cloud data corresponding to the scene to be detected;

[0008] The target three-dimensional point cloud data is input into a point cloud semantic segmentation model to obtain a semantic segmentation category result corresponding to the target three-dimensional point cloud data output by the point cloud semantic segmentation model, wherein the point cloud semantic segmentation model is trained by sample three-dimensional point cloud data marked with semantic segmentation category labels, and the sample three-dimensional point cloud data is point cloud data in a sample voxel grid determined based on voxel cohesion, and the voxel cohesion is determined according to the number of data points corresponding to point cloud data points with different semantic segmentation category labels in the sample voxel grid.

[0009] According to a point cloud semantic segmentation method based on voxel cohesion guided active learning provided by the present invention, the point cloud semantic segmentation model is trained by the following steps:

[0010] Determining, based on a feature map corresponding to the sample lidar point cloud data, a semantic category label for each point cloud data point in the sample lidar point cloud data;

[0011] voxel processing is performed on the sample lidar point cloud data to obtain a plurality of pending voxel grids, wherein the data points in the pending voxel grids are determined based on coordinate information of the point cloud data points in the sample lidar point data and a preset voxel size, and the voxel point category labels of the data points in the pending voxel grids correspond to the semantic category labels of the point cloud data points in the sample lidar point data;

[0012] Calculating the voxel cohesion of the undetermined voxel grid according to the uncertainty score of the undetermined voxel grid, wherein the uncertainty score is calculated based on the number of data points with different voxel point category labels in the voxel grid and the number of all data points in the voxel grid;

[0013] Based on the voxel cohesion of each of the to-be-determined voxel grids, determining the to-be-determined voxel grid with the lowest voxel cohesion as the sample voxel grid;

[0014] Marking the point cloud data points corresponding to each data point in the sample voxel grid in the sample lidar point data to obtain the sample three-dimensional point cloud data marked with the semantic segmentation category label;

[0015] Based on the sample three-dimensional point cloud data, a preset three-dimensional neural network model is trained to obtain the point cloud semantic segmentation model.

[0016] According to a point cloud semantic segmentation method based on voxel cohesion guided active learning provided by the present invention, the voxel cohesion of the undetermined voxel grid is calculated based on the uncertainty score of the undetermined voxel grid, including:

[0017] Based on the voxel cohesion calculation formula, the voxel cohesion of the to-be-determined voxel grid is calculated according to the uncertainty score of the to-be-determined voxel grid. The voxel cohesion calculation formula is:

[0018]

[0019]

[0020] in, Indicates the The first of the undetermined voxel grids The number of data points with voxel class labels, Indicates the A grid of undetermined voxels, represents a collection of voxel grids, Indicates the The first voxel in the undetermined voxel grid The predicted category corresponding to the data point, Represents the total number of voxel point category labels, Indicates the The voxel cohesion corresponding to the undetermined voxel grid, Indicates the The number of all data points in the undetermined voxel grid, Indicates the The uncertainty score corresponding to the undetermined voxel grid, Indicates the total number of data point categories within the undetermined voxel grid.

[0021] According to a point cloud semantic segmentation method based on voxel cohesion guided active learning provided by the present invention, the point cloud data points corresponding to each data point in the sample voxel grid in the sample lidar point data are marked to obtain the sample three-dimensional point cloud data marked with the semantic segmentation category label, including:

[0022] Based on a preset perception distance, each data point within the sample voxel grid is judged. If a key data point exists within the preset perception distance with any data point to be marked as the center, a feature difference value is calculated based on the feature vectors corresponding to the data point to be marked and the key data point, wherein the key data point is a data point within the sample voxel grid that has been determined to be a point cloud data point to be marked;

[0023] When it is determined that the feature difference value is less than or equal to a preset feature difference threshold, the data points to be marked are removed to obtain a sample voxel grid after redundant data points are removed;

[0024] Marking the point cloud data points corresponding to the respective data points of the sample voxel grid after the redundant data points are removed in the sample lidar point data to obtain the sample three-dimensional point cloud data marked with the semantic segmentation category label;

[0025] When it is determined that the feature difference value is greater than a preset feature difference threshold, or when the key data point does not exist within the preset perception distance centered on the data point to be marked, the data point to be marked is retained as the key data point.

[0026] According to a point cloud semantic segmentation method based on voxel cohesion guided active learning provided by the present invention, the method further includes:

[0027] Sort the uncertainty scores in descending order, and sort the top scores in the sorting results. The undetermined voxel grids corresponding to the uncertain scores are used as the category coordination voxel grids;

[0028] Obtaining pseudo-label data of each data point in the category-coordinated voxel grid and a preset voxel point category weight corresponding to the pseudo-label data;

[0029] Obtaining a data point uncertainty score corresponding to each data point in the category-coordinated voxel grid according to a product result between the uncertainty score of the category-coordinated voxel grid and the preset voxel point category weight;

[0030] Sort the uncertainty scores of the data points in descending order, and place the order of the results first. The data points corresponding to the uncertain scores of the data points are used as the data points to be marked;

[0031] Mark the point cloud data points corresponding to the data points to be marked in the sample lidar point data.

[0032] According to a point cloud semantic segmentation method based on voxel cohesion guided active learning provided by the present invention, a preset three-dimensional neural network model is trained based on the sample three-dimensional point cloud data to obtain the point cloud semantic segmentation model, including:

[0033] Based on the sample three-dimensional point cloud data and the cross entropy loss function, the preset three-dimensional neural network model is trained to obtain the point cloud semantic segmentation model.

[0034] The present invention also provides a point cloud semantic segmentation system based on voxel cohesion guided active learning, comprising:

[0035] Point cloud data acquisition module, used to obtain the target three-dimensional point cloud data corresponding to the scene to be detected;

[0036] A point cloud semantic segmentation module is used to input the target three-dimensional point cloud data into a point cloud semantic segmentation model to obtain a semantic segmentation category result corresponding to the target three-dimensional point cloud data output by the point cloud semantic segmentation model, wherein the point cloud semantic segmentation model is trained by sample three-dimensional point cloud data marked with semantic segmentation category labels, and the sample three-dimensional point cloud data is point cloud data in a sample voxel grid determined based on voxel cohesion, and the voxel cohesion is determined according to the number of data points corresponding to point cloud data points with different semantic segmentation category labels in the sample voxel grid.

[0037] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor, wherein when the processor executes the program, it implements any of the above-described point cloud semantic segmentation methods based on voxel cohesion-guided active learning.

[0038] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for point cloud semantic segmentation based on voxel cohesion-guided active learning as described above is implemented.

[0039] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-described point cloud semantic segmentation methods based on voxel cohesion-guided active learning.

[0040] The point cloud semantic segmentation method and system based on voxel cohesion-guided active learning provided by the present invention determine the point cloud data in the sample voxel grid through voxel cohesion, and then perform point cloud semantic segmentation processing based on the point cloud semantic segmentation model trained with these point cloud data, thereby improving the accuracy and generalization ability of the point cloud semantic segmentation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 A schematic diagram of the process of the point cloud semantic segmentation method based on voxel cohesion guided active learning provided by the present invention;

[0043] Figure 2 A schematic diagram of the active learning-based point cloud semantic segmentation framework provided by the present invention;

[0044] Figure 3 A schematic diagram of the structure of the point cloud semantic segmentation system based on voxel cohesion guided active learning provided by the present invention;

[0045] Figure 4 This is a schematic structural diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0046] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0047] Point cloud semantic segmentation technology plays a key role in multiple fields, such as environmental perception in autonomous driving, path planning in robot navigation, scene reconstruction in virtual reality, building information modeling, and material classification. However, the sparsity and irregularity of point cloud data, as well as the high cost of labeling, significantly limit the performance of segmentation models, hindering their application in practical applications.

[0048] With the development of deep learning, many point cloud semantic segmentation methods have been proposed, which can be roughly divided into the following categories:

[0049] 1. 2D projection-based methods: 3D point clouds are projected onto 2D images using multi-view, bird's-eye, and spherical projections, and then analyzed using a designed 2D convolutional neural network. While this method leverages the advantages of 2D convolutional networks, 3D geometric information is inevitably lost during the projection process, limiting segmentation performance.

[0050] Second, voxel-based methods: By voxelizing point clouds, irregular and unevenly dense point clouds are converted into regular voxel grids, enabling the extension of two-dimensional convolutional networks to three-dimensional data analysis. Early voxel-based methods often faced problems with high computational complexity and information loss. The introduction of sparse convolutions significantly reduced computational costs and improved the efficiency of voxel-based methods.

[0051] 3. Point-based methods: These methods directly process raw, non-uniform point cloud data, using flexible methods such as multilayer perceptrons (MLPs) and 3D point convolutions to process the points, preserving the original 3D information of the point cloud to the greatest extent possible. However, while these methods can preserve rich geometric information, their performance relies on large amounts of labeled data, which poses a challenge for practical applications.

[0052] Compared to indoor scenes, outdoor LiDAR point clouds, due to their large size, uneven density, and sparseness, place higher demands on the efficiency of segmentation networks. To this end, many studies have improved processing efficiency by projecting 3D point clouds onto 2D images, such as spherical views (range images) or bird's-eye views, or by fusing different views. Furthermore, there is also research on directly processing point cloud data, aiming to more efficiently build models for irregular data. These methods can serve as backbone networks in active learning frameworks.

[0053] By selectively labeling the most informative samples, active learning can significantly reduce labeling costs while ensuring model performance close to that of fully supervised learning. In particular, in applications with moderate accuracy requirements and limited labeling budgets, active learning strategies can prioritize labeling samples with higher uncertainty, achieving an effective balance between cost and accuracy. Therefore, active learning has demonstrated great potential and practicality in point cloud semantic segmentation tasks for 3D scene understanding, especially in resource-constrained scenarios.

[0054] To address the difficulty of labeling large-scale point cloud data, active learning methods gradually select and request the network to label some samples, such as scans, regions, points, or boxes, during training. Most selection strategies are based on uncertainty or diversity principles. For example, uncertainty sampling selects the most challenging samples by measuring the uncertainty of the model's prediction score (such as the softmax output) for each point, or the gap between the two highest scores.

[0055] Existing active learning methods typically select samples based on uncertainty and diversity. However, these methods struggle to fully capture the geometric and topological features of 3D point cloud data when directly processing it. This is particularly true when processing large-scale outdoor LiDAR point clouds. These methods are often inefficient or require complex preprocessing, making them unsuitable for some scenarios. Furthermore, the strong correlation between point cloud data samples makes existing active learning strategies difficult to achieve, resulting in poor segmentation accuracy.

[0056] In response to the problems existing in the above-mentioned prior art, the present invention proposes a point cloud semantic segmentation method based on voxel cohesion-guided active learning. When training the point cloud semantic segmentation model, this method first combines the local and global geometric features of the point cloud data to voxelize the point cloud, and measures its information complexity by calculating the entropy value of each voxel block, and records it as the voxel cohesion to screen out the voxel set with the highest uncertainty. Secondly, by measuring the similarity between data points in each voxel grid in the set, accurate filtering of redundant information points within the voxel is achieved. Finally, an online category coordination module is designed to optimize the threshold setting according to the data density and labeling requirements of the category, and dynamically adjust the number of candidate points of each category during the labeling process, thereby reducing the negative impact of category imbalance on the model training effect and improving labeling efficiency and data utilization.

[0057] Figure 1 The flowchart of the point cloud semantic segmentation method based on voxel cohesion guided active learning provided by the present invention is as follows: Figure 1 As shown, the present invention provides a point cloud semantic segmentation method based on voxel cohesion guided active learning, comprising:

[0058] Step 101: Obtain target three-dimensional point cloud data corresponding to the scene to be detected.

[0059] In this invention, laser radar scanning is used to calculate distance by emitting a laser beam and measuring the time it takes for it to reflect back. This technology obtains the three-dimensional coordinates of objects in a scene, generating three-dimensional point cloud data for that scene. Three-dimensional point cloud data is a collection of points in space, each typically containing three-dimensional coordinates (X, Y, Z). Together, these points describe the shape, position, and surface characteristics of an object.

[0060] Step 102: input the target three-dimensional point cloud data into a point cloud semantic segmentation model to obtain a semantic segmentation category result corresponding to the target three-dimensional point cloud data output by the point cloud semantic segmentation model, wherein the point cloud semantic segmentation model is trained by sample three-dimensional point cloud data marked with semantic segmentation category labels, and the sample three-dimensional point cloud data is point cloud data in a sample voxel grid determined based on voxel cohesion, and the voxel cohesion is determined according to the number of data points corresponding to point cloud data points with different semantic segmentation category labels in the sample voxel grid.

[0061] Point cloud semantic segmentation involves classifying each point in point cloud data, or the objects to which each point belongs, into specific semantic categories, such as cars, trees, buildings, and pedestrians. In this paper, the model input is the target 3D point cloud data, i.e., the point cloud dataset to be semantically segmented. The output is the semantic segmentation classification results corresponding to this point cloud data, i.e., the semantic labels assigned to each point or point set.

[0062] In the present invention, the point cloud semantic segmentation model is trained by learning sample three-dimensional point cloud data marked with semantic segmentation category labels. These sample data contain rich scene and object information, and each point or point set is correctly labeled with a semantic category.

[0063] During training, the model uses iterative optimization algorithms (such as gradient descent) to learn how to extract features from the input point cloud data and classify points into the correct semantic categories based on these features. Specifically, in order to process large-scale point cloud data, the present invention organizes the point cloud data into a voxel grid, which is the smallest unit in three-dimensional space, similar to the pixel in a two-dimensional image. The voxel grid divides the three-dimensional space into a series of small cubes, each of which can contain a certain number of points.

[0064] Furthermore, in the present invention, voxel cohesion is determined based on the number of data points corresponding to point cloud data points with different semantic segmentation category labels in the sample voxel grid. Specifically, the number (or density) of data points within the voxel grid and the distribution of the semantic categories to which these data points belong reflect the voxel's "cohesion." Voxels with high cohesion are likely to contain more or more concentrated points of a particular category. Therefore, based on voxel cohesion, more representative point cloud data from the sample voxel grid can be selected as training samples, which helps the model learn more accurate features and classification rules.

[0065] After the model is trained, new, unlabeled target 3D point cloud data can be input into the model. The model will process the point cloud data and output the semantic segmentation category results corresponding to each point or point set.

[0066] The point cloud semantic segmentation method based on voxel cohesion-guided active learning provided by the present invention determines the point cloud data in the sample voxel grid through voxel cohesion, and then performs point cloud semantic segmentation processing based on the point cloud semantic segmentation model trained with these point cloud data, thereby improving the accuracy and generalization ability of the point cloud semantic segmentation results.

[0067] Based on the above embodiment, the point cloud semantic segmentation model is trained by the following steps:

[0068] Determining, based on a feature map corresponding to the sample lidar point cloud data, a semantic category label for each point cloud data point in the sample lidar point cloud data;

[0069] voxel processing is performed on the sample lidar point cloud data to obtain a plurality of pending voxel grids, wherein the data points in the pending voxel grids are determined based on coordinate information of the point cloud data points in the sample lidar point data and a preset voxel size, and the voxel point category labels of the data points in the pending voxel grids correspond to the semantic category labels of the point cloud data points in the sample lidar point data;

[0070] Calculating the voxel cohesion of the undetermined voxel grid according to the uncertainty score in the undetermined voxel grid, wherein the uncertainty score is calculated based on the number of data points with different voxel point category labels in the voxel grid and the number of all data points in the voxel grid;

[0071] Based on the voxel cohesion of each of the to-be-determined voxel grids, determining the to-be-determined voxel grid with the lowest voxel cohesion as the sample voxel grid;

[0072] Marking the point cloud data points corresponding to each data point in the sample voxel grid in the sample lidar point data to obtain the sample three-dimensional point cloud data marked with the semantic segmentation category label;

[0073] Based on the sample three-dimensional point cloud data, a preset three-dimensional neural network model is trained to obtain the point cloud semantic segmentation model.

[0074] Figure 2 The schematic diagram of the active learning-based point cloud semantic segmentation framework provided by the present invention can be referred to Figure 2As shown, in the present invention, the sparse point voxel convolutional network (SPVCNN) is trained as the three-dimensional backbone network in the preset three-dimensional neural network. First, the present invention randomly selects a portion of points in the sample lidar point cloud as initial marker points, and these marker points are used to guide subsequent model training. Then, the point cloud data is input into the three-dimensional backbone network SPVCNN for training, and the feature map of each point is extracted. After the feature extraction is completed, the input point cloud is voxelized. After voxelization, the present invention selects the voxel grid based on the voxel cohesion metric. Preferably, the present invention provides a voxel cohesion measurement and redundancy filtering module (VCMRF for short), which evaluates its uncertainty by measuring the cohesion within each voxel grid, and accurately filters out redundant information points based on similarity calculation, retaining representative key points. Optionally, the filtered point cloud data is processed by the Online Adaptive Category Coordination Module (ACCM), which dynamically adjusts the number of points of each category in the candidate pool to balance the representation and distribution of each category in the dataset. Finally, the points in the processed candidate pool are labeled online and added to the training set for the next iteration to further optimize the model's performance and accuracy.

[0075] In this method, sample LiDAR point cloud data is acquired. These data points represent sampling points on the surface of an object or environment in three-dimensional space. Each point cloud data point is then assigned a semantic category label, such as "vehicle," "pedestrian," or "tree," through manual labeling or prediction based on other trained models. In this method, feature maps are extracted from the sample LiDAR point cloud data using a three-dimensional backbone network to obtain semantic category labels for each point cloud data point in the sample LiDAR point cloud data.

[0076] Furthermore, the entire point cloud data space is divided into small three-dimensional cubes, called voxel grids. The size of each voxel grid is preset, and the data points within the voxel grid are assigned to corresponding voxels based on the coordinate information of the point cloud data points in the point cloud data space. In the present invention, the data points within each voxel grid inherit the semantic category label of any point cloud data points it contains. Therefore, a voxel grid may contain points of multiple different categories, or all of them may belong to the same category.

[0077] In the present invention, an uncertainty score is calculated for each voxel grid. This score reflects the degree of inconsistency in the class labels of the point cloud data points within the voxel grid. The uncertainty score can be calculated based on the ratio of the number of data points with different class labels within the voxel grid to the total number of data points in the voxel grid.

[0078] Next, this uncertainty score is used to calculate the cohesion of the voxel grid. A lower cohesion means that the category labels of the point cloud data points within the voxel grid are more diverse, and may contain parts or edges of objects of multiple different categories. Furthermore, the voxel grid with the lowest voxel cohesion is selected from all the pending voxel grids as the sample voxel grid. In the present invention, all point cloud data points belonging to the sample voxel grid are labeled to obtain a sample three-dimensional point cloud dataset with accurate semantic segmentation labels. Then, the preset three-dimensional neural network model is trained using this labeled sample three-dimensional point cloud data. During the training process, the model learns how to predict the semantic category of each point based on the spatial distribution and characteristics of the point cloud data. After sufficient iterations and optimization, a model that can accurately perform point cloud semantic segmentation is finally obtained, namely a point cloud semantic segmentation model. The model can understand and interpret complex three-dimensional environments, thereby supporting more advanced functions such as path planning and obstacle avoidance for self-driving cars.

[0079] Based on the above embodiment, the step of calculating the voxel cohesion of the undetermined voxel grid according to the uncertainty score of the undetermined voxel grid includes:

[0080] Based on the voxel cohesion calculation formula, the voxel cohesion of the to-be-determined voxel grid is calculated according to the uncertainty score of the to-be-determined voxel grid. The voxel cohesion calculation formula is:

[0081]

[0082]

[0083] in, Indicates the The first of the undetermined voxel grids The number of data points with voxel class labels, Indicates the A grid of undetermined voxels, represents a collection of voxel grids, Indicates the The first voxel in the undetermined voxel grid The predicted category corresponding to the data point, Represents the total number of voxel point category labels, Indicates the The voxel cohesion corresponding to the undetermined voxel grid, Indicates the The number of all data points in the undetermined voxel grid, Indicates the The uncertainty score corresponding to the undetermined voxel grid, Indicates the total number of data point categories within the undetermined voxel grid.

[0084] In the present invention, it is necessary to select and label the point cloud data points with the highest value from the sample lidar point cloud data to maximize the model performance, that is, identifying and using the labeled data with the most information for training is the key to obtaining a high-quality model, thereby achieving accurate segmentation prediction.

[0085] Specifically, the present invention first performs voxel processing on the sample lidar point cloud data and converts it into a three-dimensional voxel grid set The process involves sampling the continuous 3D spatial input into a discrete voxel grid, and points falling within the same grid are merged. Each voxel grid is treated as an independent selection unit.

[0086] For sample lidar point cloud data The Point cloud data points , accordingly, the point cloud data points in the voxel grid coordinates are:

[0087] ;

[0088] in, Represents a predefined voxel grid size. Point cloud data points As a point in three-dimensional space, its coordinates are expressed as . is a point cloud data point Coordinates in the voxel grid. In the present invention, the point cloud data points Each coordinate component of , , Divide by the preset voxel size Δ and then round down (i.e., take the largest integer not greater than this number) to divide the continuous three-dimensional space into discrete voxel grids.

[0089] Furthermore, for each voxel grid , evaluate its semantic cohesion, i.e. voxel cohesion. Since point cloud data usually shows local semantic similarity, it is assumed that points in the same voxel grid have similar semantics. Therefore, the present invention evaluates the importance of each voxel grid by voxel cohesion. The number of data points corresponding to the category label of each voxel point in each voxel grid The calculation process is as follows:

[0090]

[0091] The uncertainty score of the undetermined voxel grid is calculated as follows:

[0092]

[0093] in, Indicates the Uncertainty score for the voxel grid.

[0094] The formula for calculating voxel cohesion is:

[0095]

[0096] In this invention, the voxel cohesion of a voxel grid is calculated using its uncertainty score. In each 3D point cloud scan, the voxel grid with the lowest cohesion is selected. Lower cohesion indicates greater class diversity within the voxels, and marking these voxels contributes more to model training. Model training or subsequent model updates are then performed based on these selected voxel grids.

[0097] Based on the above embodiment, the step of marking the point cloud data points corresponding to the sample lidar point data for each data point within the sample voxel grid to obtain the sample three-dimensional point cloud data marked with the semantic segmentation category label includes:

[0098] Based on a preset perception distance, each data point within the sample voxel grid is judged. If a key data point exists within the preset perception distance with any data point to be marked as the center, a feature difference value is calculated based on the feature vectors corresponding to the data point to be marked and the key data point, wherein the key data point is a data point within the sample voxel grid that has been determined to be a point cloud data point to be marked;

[0099] When it is determined that the feature difference value is less than or equal to a preset feature difference threshold, the data points to be marked are removed to obtain a sample voxel grid after redundant data points are removed;

[0100] Marking the point cloud data points corresponding to the respective data points of the sample voxel grid after the redundant data points are removed in the sample lidar point data to obtain the sample three-dimensional point cloud data marked with the semantic segmentation category label;

[0101] When it is determined that the feature difference value is greater than a preset feature difference threshold, or when the key data point does not exist within the preset perception distance centered on the data point to be marked, the data point to be marked is retained as the key data point.

[0102] In the present invention, redundancy filtering can also be performed on voxels to further streamline the data points within the voxels, retaining only key data points that can represent the entire voxel information, thereby avoiding redundant labeling and increasing labeling costs.

[0103] Specifically, the present invention predefines a perception distance and feature difference threshold , for any data point in the voxel grid , that is, The first data points, check at data points When the center is the preset sensing distance, Are there other data points within the range that have been selected as key data points to establish an evaluation set? .

[0104] If the evaluation set If it is empty, then the data point Points that are considered representative can be used as key data points for subsequent labeling. If not empty, the data point is calculated With the evaluation set Each data point in (i.e., key data points) feature difference values , feature difference value is calculated as follows:

[0105]

[0106] in, and Represents data points With data points The feature vector of can be obtained through the trained feature extraction model.

[0107] In the present invention, if the evaluation set There is at least one data point in ,satisfy , that is, the feature difference between the two data points is small, indicating that the data points On the features and data points Too close, so the data points The redundant data points are defined as redundant and removed from the sample voxel grid. The redundant data point removal strategy provided by the present invention ensures the independence of the selected data points in space and features, effectively improving the expression and utilization efficiency of point cloud data.

[0108] Based on the above embodiment, the method further includes:

[0109] Sort the uncertainty scores in descending order, and sort the top scores in the sorting results. The undetermined voxel grids corresponding to the uncertain scores are used as the category coordination voxel grids;

[0110] Obtaining pseudo-label data of each data point in the category-coordinated voxel grid and a preset voxel point category weight corresponding to the pseudo-label data;

[0111] Obtaining a data point uncertainty score corresponding to each data point in the category-coordinated voxel grid according to a product result between the uncertainty score of the category-coordinated voxel grid and the preset voxel point category weight;

[0112] Sort the uncertainty scores of the data points in descending order, and place the order of the results first. The data points corresponding to the uncertain scores of the data points are used as the data points to be marked;

[0113] Mark the point cloud data points corresponding to the data points to be marked in the sample lidar point data.

[0114] Active learning strategies can effectively select information-rich samples, but these samples may not be sufficient to maintain balance between categories, especially in cases of extreme class imbalance. Without class balancing, the model may perform well on the majority class and poorly on the minority class. This deviation may reduce the effectiveness of point cloud semantic segmentation models in practical applications, especially in scenarios where the minority class is crucial. To this end, the present invention introduces an online class balancing module to ensure that all categories have sufficient sample coverage and provide a more balanced class representation, thereby improving the overall performance of the model.

[0115] Specifically, according to the order of uncertainty scores from large to small, select the Undetermined voxel grids are used as category coordination voxel grids. Assume that these voxel grids contain Points. The uncertainty score for a voxel grid is ,in, For voxels For each data point in , its uncertainty score is also recorded as .

[0116] set up is the preset voxel point category weight vector, where is the number of voxel point categories, The weight of the corresponding voxel point category c. The pseudo label of each data point is obtained by predicting the relevant model , i represents the i-th data point, , whose uncertainty score is , the data point i in the category-coordinated voxel grid is updated by the following formula to obtain the data point uncertainty score corresponding to each data point in the category-coordinated voxel grid :

[0117]

[0118] in, Represents pseudo labels The preset voxel point category weight vector corresponding to the data point.

[0119] Finally, the categories are coordinated in the voxel grid in descending order of uncertainty scores. The point cloud data points corresponding to the data points are marked.

[0120] On the basis of the above embodiment, the preset three-dimensional neural network model is trained based on the sample three-dimensional point cloud data to obtain the point cloud semantic segmentation model, including:

[0121] Based on the sample three-dimensional point cloud data and the cross entropy loss function, the preset three-dimensional neural network model is trained to obtain the point cloud semantic segmentation model.

[0122] In this paper, in order to accurately train the point cloud semantic segmentation model and optimize its parameters, the standard cross entropy loss function is used as the overall loss function. The cross entropy loss function is a standard loss function for classification tasks. It calculates the difference between the predicted probability distribution output by the model and the target probability distribution of the true label. The formula of this loss function can be expressed as:

[0123]

[0124] in, represents the number of samples, represents the number of categories, Representation sample For category The true label (usually 0 or 1), Represents the model prediction sample Belong to category probability.

[0125] In the present invention, the cross-entropy loss function is used to optimize the model by minimizing the distance between the predicted probability and the target probability, which can directly provide feedback for the prediction of each category. It not only improves the model's prediction probability for the correct category, but also suppresses the prediction probability for the wrong category, thereby effectively improving the accuracy of classification.

[0126] In one embodiment, the performance of the point cloud semantic segmentation method based on voxel cohesion guided active learning provided by the present invention is experimentally verified, and the selected benchmark datasets include nuScenes and SemanticKITTI.

[0127] The nuScenes dataset includes 1,000 driving scenes from various cities, encompassing a wide range of weather and lighting conditions, reflecting the complexity of real-world driving. Each scene lasts 20 seconds, with a total of 28,130 training frames and 6,019 validation frames. Each frame consists of a point cloud generated by 32 lidar beams, with point-by-point annotations, and RGB images captured by six cameras from different viewpoints, segmented using sixteen categories.

[0128] The SemanticKITTI dataset is a popular segmentation dataset focused on LiDAR point clouds. It provides a total of 43,000 scans with point-by-point semantic annotations. Of these, 23,201 scans from sequences 00 to 10 are used for training and validation, including 19,130 ​​training scans and 4,071 validation scans, supporting deep segmentation research across 19 categories. These features make SemanticKITTI an ideal choice for understanding and evaluating point cloud data processing techniques.

[0129] This paper uses SPVCNN as the backbone, with a hidden dimension of 64 for SemanticKITTI and 128 for NuScenes to speed up the network. Data augmentation is performed on the input point cloud using random perturbations and random flipping. The stochastic gradient descent (SGD) optimizer is employed with an initial learning rate (lr) of 0.24 and a momentum of 0.9. The learning rate is decayed using a cosine annealing strategy with a weight decay coefficient of 1e-4. The total number of training rounds is set to 65, and the batch size is set to 8. All code is implemented in PyTorch, and experiments were conducted on a server equipped with a GeForce RTX 4090, 24GB of video memory, Ubuntu 22.04 operating system, and CUDA version 12.1.2.

[0130] In this paper, the mean intersection over union (mIoU) is mainly used to evaluate the performance of the method. mIoU is a commonly used indicator for evaluating classification accuracy, especially for semantic segmentation tasks, because it comprehensively considers the prediction accuracy and recall rate of each category. The calculation method of mIoU is as follows:

[0131] First, the intersection over union (IoU) is calculated for each category separately, and then the IoU of all categories is averaged. The IoU of a category is calculated by dividing the number of points that the model correctly classifies for that category by the sum of the total number of points in the predicted set and the true set. The specific calculation formula is:

[0132]

[0133]

[0134] in, 、 and denote the number of true positives, false positives, true negatives and false negatives, respectively. is the total number of categories.

[0135] To verify the effectiveness of VCMRF and OCC, we conducted a series of ablation experiments on the SemantiKitti dataset. Table 1 shows the performance comparison results of the method provided by this invention with other common strategies under different benchmark active learning strategies:

[0136] Table 1 Comparison results of different benchmark active learning strategies

[0137]

[0138] Furthermore, to evaluate the efficiency of our proposed method, we replaced several baseline active learning strategies on a unified network architecture (SPVCNN) while strictly controlling the labeling budget (i.e., labeling only 0.1% of the points). Experimental results demonstrate that our method significantly improves the mean Intersection Over Union (MIoU) by 11.5% compared to the random point selection method, reaching 61.7%. Other methods, such as the entropy-based and edge-based strategies, achieve mIoUs of 56.8% and 52.1%, respectively. In comparison, our method more efficiently identifies and labels informative points while effectively addressing class imbalance, significantly improving model performance.

[0139] Furthermore, to verify the efficiency of each module in the present invention's method, several modules (i.e., the backbone network, voxel cohesion and redundancy filtering module, and online adaptive class coordination module) were combined in various experimental combinations to validate their effectiveness. Given a total of 0.1% of points with identical labels, direct training using the backbone network achieved a measurable intersection over union (MIU) of only 54.3% with an initial random 0.1% of the points labeled. Using the voxel cohesion and redundancy filtering module, segmentation accuracy improved to 60.2%. Combining the voxel cohesion and redundancy filtering module with the online adaptive class coordination module increased performance to 61.7%.

[0140] Furthermore, the present invention presents a visual comparison of the ground truth, fully supervised results, the proposed method, and the randomized active learning strategy in five randomly selected typical scenarios. This comparison intuitively demonstrates that the proposed method is very close to the results of fully supervised learning in terms of overall effectiveness. In some scenarios, the proposed method even demonstrates superior performance for rare class recognition. In contrast, the results of the randomized active learning strategy are inferior to those of the proposed method.

[0141] In some scenarios, the random strategy exhibits low semantic segmentation accuracy, particularly for rare categories (such as traffic signs, tree trunks, and other motor vehicles) and scene boundaries (such as the intersection of pedestrians). Randomly selected samples fail to fully capture important information in the data, resulting in high segmentation error rates in these scenarios. This suggests that random active learning strategies struggle to effectively capture complex patterns and detailed information in the data, impacting the overall performance of the segmentation model. In contrast, the proposed method intelligently selects high-information samples for labeling, enabling the model to better learn key features. This reduces labeling costs while still achieving segmentation results comparable to fully supervised methods.

[0142] Table 2 shows the comparison results between the proposed method and existing methods on the SemantiKitti dataset, covering 19 semantic segmentation categories:

[0143] Table 2 Comparison of SemantiKitti dataset

[0144]

[0145] SPVCNN, as a fully supervised method and the three-dimensional network skeleton used in the present invention, achieved the highest mIoU (64.5%), but the labeling cost was extremely high. ReDAL achieved an mIoU close to that of full supervision (59.8%) at a labeling rate of 5%, demonstrating the high efficiency of active learning. SLiDR and LESS also achieved good results using extremely low labeled data, especially in the truck and road categories, but had low accuracy in rare categories such as cyclists. In contrast, the method of the present invention achieved an mIoU of 61.7% using only 0.1% of the labeled points, and even surpassed the fully supervised results in some categories. This shows that the method of the present invention can maintain excellent segmentation performance while reducing the labeling cost.

[0146] In response to the problems of existing three-dimensional point cloud semantic segmentation methods, such as long dataset annotation time, high labor cost, noise sensitivity and insufficient accuracy, the present invention proposes a point cloud semantic segmentation method based on voxel cohesion guided active learning.

[0147] Specifically, the present invention proposes a voxel cohesion metric and redundancy filtering module. This module uses the voxel cohesion within each voxel to assess the uncertainty of individual voxels, thereby obtaining the similarity relationship between points within the voxel and accurately filtering out redundant information points within the voxel. This process not only filters out redundant data points but also retains key information points that represent all points within the voxel. In addition, this module avoids the existing practice of directly labeling entire voxels, reducing the labeling budget wasted due to redundant labeling and effectively solving the problem of redundant labeling within each voxel.

[0148] On this basis, an online adaptive category coordination module is used to dynamically adjust the number of candidate points for each category during the annotation process by setting specific thresholds for different categories. An adaptive threshold adjustment mechanism is employed to automatically optimize the threshold setting based on the category's data density and annotation requirements, thereby effectively improving the representativeness and balance of each category in the dataset and reducing the negative impact of category imbalance on model training. Furthermore, by optimizing the allocation of annotation resources, annotation efficiency and data utilization are improved, addressing the category imbalance problem during the point selection process and preventing the network from expressing overconfidence in the dominant category.

[0149] This paper, through the innovative combination of voxel cohesion and redundancy filtering modules and an online adaptive category coordination module, not only effectively improves labeling efficiency but also ensures category balance in the dataset, thereby significantly enhancing the accuracy and generalization of the model. This has important practical applications in improving point cloud labeling efficiency and data processing accuracy, and is particularly suitable for autonomous driving, robotic navigation, building information modeling and other fields.

[0150] The point cloud semantic segmentation system based on voxel cohesion guided active learning provided by the present invention is described below. The point cloud semantic segmentation system based on voxel cohesion guided active learning described below and the point cloud semantic segmentation method based on voxel cohesion guided active learning described above can be referenced to each other.

[0151] Figure 3 This is a schematic diagram of the structure of the point cloud semantic segmentation system based on voxel cohesion guided active learning provided by the present invention, such as Figure 3As shown, the present invention provides a point cloud semantic segmentation system based on voxel cohesion guided active learning, including a point cloud data acquisition module 301 and a point cloud semantic segmentation module 302, wherein the point cloud data acquisition module 301 is used to obtain target three-dimensional point cloud data corresponding to the scene to be detected; the point cloud semantic segmentation module is used to input the target three-dimensional point cloud data into a point cloud semantic segmentation model to obtain a semantic segmentation category result corresponding to the target three-dimensional point cloud data output by the point cloud semantic segmentation model, wherein the point cloud semantic segmentation model is trained by sample three-dimensional point cloud data marked with semantic segmentation category labels, and the sample three-dimensional point cloud data is point cloud data in a sample voxel grid determined based on voxel cohesion, and the voxel cohesion is determined according to the number of data points corresponding to point cloud data points with different semantic segmentation category labels in the sample voxel grid.

[0152] The point cloud semantic segmentation system based on voxel cohesion-guided active learning provided by the present invention determines the point cloud data in the sample voxel grid through voxel cohesion, and then performs point cloud semantic segmentation processing based on the point cloud semantic segmentation model trained with these point cloud data, thereby improving the accuracy and generalization ability of the point cloud semantic segmentation results.

[0153] The system provided by the present invention is used to execute the above-mentioned method embodiments. Please refer to the above-mentioned embodiments for the specific processes and detailed contents, which will not be repeated here.

[0154] Figure 4 A schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 4 As shown, the electronic device may include: a processor 401, a communications interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communications interface 402, and the memory 403 communicate with each other via the communication bus 404. The processor 401 may call logic instructions in the memory 403 to execute a point cloud semantic segmentation method based on voxel cohesion-guided active learning. The method includes: obtaining target three-dimensional point cloud data corresponding to a scene to be detected; inputting the target three-dimensional point cloud data into a point cloud semantic segmentation model, and obtaining a semantic segmentation category result corresponding to the target three-dimensional point cloud data output by the point cloud semantic segmentation model. The point cloud semantic segmentation model is trained using sample three-dimensional point cloud data labeled with semantic segmentation category labels. The sample three-dimensional point cloud data is point cloud data in a sample voxel grid determined based on voxel cohesion. The voxel cohesion is determined based on the number of data points corresponding to point cloud data points with different semantic segmentation category labels in the sample voxel grid.

[0155] Furthermore, the logic instructions in the aforementioned memory 403 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product, stored in a storage medium, includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0156] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the point cloud semantic segmentation method based on voxel cohesion guided active learning provided by the above methods, the method including: obtaining target three-dimensional point cloud data corresponding to the scene to be detected; inputting the target three-dimensional point cloud data into a point cloud semantic segmentation model, and obtaining a semantic segmentation category result corresponding to the target three-dimensional point cloud data output by the point cloud semantic segmentation model, wherein the point cloud semantic segmentation model is trained by sample three-dimensional point cloud data marked with semantic segmentation category labels, and the sample three-dimensional point cloud data is point cloud data in a sample voxel grid determined based on voxel cohesion, and the voxel cohesion is determined according to the number of data points corresponding to point cloud data points with different semantic segmentation category labels in the sample voxel grid.

[0157] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the point cloud semantic segmentation method based on voxel cohesion-guided active learning provided in the above-mentioned embodiments, the method comprising: obtaining target three-dimensional point cloud data corresponding to the scene to be detected; inputting the target three-dimensional point cloud data into a point cloud semantic segmentation model, and obtaining a semantic segmentation category result corresponding to the target three-dimensional point cloud data output by the point cloud semantic segmentation model, wherein the point cloud semantic segmentation model is trained by sample three-dimensional point cloud data marked with semantic segmentation category labels, and the sample three-dimensional point cloud data is point cloud data in a sample voxel grid determined based on voxel cohesion, and the voxel cohesion is determined according to the number of data points corresponding to point cloud data points with different semantic segmentation category labels in the sample voxel grid.

[0158] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0159] Through the description of the above embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A point cloud semantic segmentation method based on voxel cohesion guided active learning, characterized in that: include: Obtain the target three-dimensional point cloud data corresponding to the scene to be detected; Inputting the target three-dimensional point cloud data into a point cloud semantic segmentation model to obtain a semantic segmentation category result corresponding to the target three-dimensional point cloud data output by the point cloud semantic segmentation model, wherein the point cloud semantic segmentation model is trained by sample three-dimensional point cloud data marked with semantic segmentation category labels, and the sample three-dimensional point cloud data is point cloud data in a sample voxel grid determined based on voxel cohesion, and the voxel cohesion is determined according to the number of data points corresponding to point cloud data points with different semantic segmentation category labels in the sample voxel grid; The point cloud semantic segmentation model is trained by the following steps: Determining, based on a feature map corresponding to the sample lidar point cloud data, a semantic category label for each point cloud data point in the sample lidar point cloud data; voxel processing is performed on the sample lidar point cloud data to obtain a plurality of pending voxel grids, wherein the data points in the pending voxel grids are determined based on coordinate information of the point cloud data points in the sample lidar point data and a preset voxel size, and the voxel point category labels of the data points in the pending voxel grids correspond to the semantic category labels of the point cloud data points in the sample lidar point data; Calculating the voxel cohesion of the undetermined voxel grid according to the uncertainty score of the undetermined voxel grid, wherein the uncertainty score is calculated based on the number of data points with different voxel point category labels in the voxel grid and the number of all data points in the voxel grid; Based on the voxel cohesion of each of the to-be-determined voxel grids, determining the to-be-determined voxel grid with the lowest voxel cohesion as the sample voxel grid; Marking the point cloud data points corresponding to each data point in the sample voxel grid in the sample lidar point data to obtain the sample three-dimensional point cloud data marked with the semantic segmentation category label; Based on the sample three-dimensional point cloud data, a preset three-dimensional neural network model is trained to obtain the point cloud semantic segmentation model; The step of calculating the voxel cohesion of the undetermined voxel grid according to the uncertainty score of the undetermined voxel grid includes: Based on the voxel cohesion calculation formula, the voxel cohesion of the to-be-determined voxel grid is calculated according to the uncertainty score of the to-be-determined voxel grid. The voxel cohesion calculation formula is: Among them, C jk represents the number of data points with the k-th voxel point category label in the j-th undetermined voxel grid, v j represents the jth undetermined voxel grid, V represents the voxel grid set, p(x ji ) represents the predicted category corresponding to the i-th data point in the j-th undetermined voxel grid, N represents the total number of voxel point category labels, S(v j ) represents the voxel cohesion corresponding to the jth undetermined voxel grid, |n j | represents the number of all data points in the jth undetermined voxel grid, represents the uncertainty score corresponding to the j-th undetermined voxel grid, and K represents the total number of data point categories in the undetermined voxel grid; The method further comprises: Sorting the uncertainty scores in descending order, and taking the undetermined voxel grids corresponding to the first M uncertainty scores in the sorting result as the category coordination voxel grids; Obtaining pseudo-label data of each data point in the category-coordinated voxel grid and a preset voxel point category weight corresponding to the pseudo-label data; Obtaining a data point uncertainty score corresponding to each data point in the category-coordinated voxel grid according to a product result between the uncertainty score of the category-coordinated voxel grid and the preset voxel point category weight; Sort the uncertainty scores of the data points in descending order, and use the data points corresponding to the uncertainty scores of the first L data points in the sorting result as the data points to be marked; Mark the point cloud data points corresponding to the data points to be marked in the sample lidar point data.

2. The point cloud semantic segmentation method based on voxel cohesion guided active learning according to claim 1, characterized in that: The step of labeling the point cloud data points corresponding to the sample lidar point data for each data point within the sample voxel grid to obtain the sample three-dimensional point cloud data labeled with the semantic segmentation category label includes: Based on a preset perception distance, each data point within the sample voxel grid is judged. If a key data point exists within the preset perception distance with any data point to be marked as the center, a feature difference value is calculated based on the feature vectors corresponding to the data point to be marked and the key data point, wherein the key data point is a data point within the sample voxel grid that has been determined to be a point cloud data point to be marked; When it is determined that the feature difference value is less than or equal to a preset feature difference threshold, the data points to be marked are removed to obtain a sample voxel grid after redundant data points are removed; Marking the point cloud data points corresponding to the respective data points of the sample voxel grid after the redundant data points are removed in the sample lidar point data to obtain the sample three-dimensional point cloud data marked with the semantic segmentation category label; When it is determined that the feature difference value is greater than a preset feature difference threshold, or when the key data point does not exist within the preset perception distance centered on the data point to be marked, the data point to be marked is retained as the key data point.

3. The point cloud semantic segmentation method based on voxel cohesion guided active learning according to claim 1, characterized in that: The method of training a preset three-dimensional neural network model based on the sample three-dimensional point cloud data to obtain the point cloud semantic segmentation model includes: Based on the sample three-dimensional point cloud data and the cross entropy loss function, the preset three-dimensional neural network model is trained to obtain the point cloud semantic segmentation model.

4. A point cloud semantic segmentation system based on voxel cohesion guided active learning, characterized by: include: Point cloud data acquisition module, used to obtain the target three-dimensional point cloud data corresponding to the scene to be detected; a point cloud semantic segmentation module, configured to input the target three-dimensional point cloud data into a point cloud semantic segmentation model, and obtain a semantic segmentation category result corresponding to the target three-dimensional point cloud data output by the point cloud semantic segmentation model, wherein the point cloud semantic segmentation model is trained using sample three-dimensional point cloud data marked with semantic segmentation category labels, the sample three-dimensional point cloud data being point cloud data in a sample voxel grid determined based on voxel cohesion, wherein the voxel cohesion is determined based on the number of data points corresponding to point cloud data points with different semantic segmentation category labels in the sample voxel grid; The point cloud semantic segmentation model is trained by the following steps: Determining, based on a feature map corresponding to the sample lidar point cloud data, a semantic category label for each point cloud data point in the sample lidar point cloud data; voxel processing is performed on the sample lidar point cloud data to obtain a plurality of pending voxel grids, wherein the data points in the pending voxel grids are determined based on coordinate information of the point cloud data points in the sample lidar point data and a preset voxel size, and the voxel point category labels of the data points in the pending voxel grids correspond to the semantic category labels of the point cloud data points in the sample lidar point data; Calculating the voxel cohesion of the undetermined voxel grid according to the uncertainty score of the undetermined voxel grid, wherein the uncertainty score is calculated based on the number of data points with different voxel point category labels in the voxel grid and the number of all data points in the voxel grid; Based on the voxel cohesion of each of the to-be-determined voxel grids, determining the to-be-determined voxel grid with the lowest voxel cohesion as the sample voxel grid; Marking the point cloud data points corresponding to each data point in the sample voxel grid in the sample lidar point data to obtain the sample three-dimensional point cloud data marked with the semantic segmentation category label; Based on the sample three-dimensional point cloud data, a preset three-dimensional neural network model is trained to obtain the point cloud semantic segmentation model; The step of calculating the voxel cohesion of the undetermined voxel grid according to the uncertainty score of the undetermined voxel grid includes: Based on the voxel cohesion calculation formula, the voxel cohesion of the to-be-determined voxel grid is calculated according to the uncertainty score of the to-be-determined voxel grid. The voxel cohesion calculation formula is: Among them, C jk represents the number of data points with the k-th voxel point category label in the j-th undetermined voxel grid, v j represents the jth undetermined voxel grid, V represents the voxel grid set, p(x ji ) represents the predicted category corresponding to the i-th data point in the j-th undetermined voxel grid, N represents the total number of voxel point category labels, S(v j ) represents the voxel cohesion corresponding to the jth undetermined voxel grid, |n j | represents the number of all data points in the jth undetermined voxel grid, represents the uncertainty score corresponding to the j-th undetermined voxel grid, and K represents the total number of data point categories in the undetermined voxel grid; Also includes: Sorting the uncertainty scores in descending order, and taking the undetermined voxel grids corresponding to the first M uncertainty scores in the sorting result as the category coordination voxel grids; Obtaining pseudo-label data of each data point in the category-coordinated voxel grid and a preset voxel point category weight corresponding to the pseudo-label data; Obtaining a data point uncertainty score corresponding to each data point in the category-coordinated voxel grid according to a product result between the uncertainty score of the category-coordinated voxel grid and the preset voxel point category weight; Sort the uncertainty scores of the data points in descending order, and use the data points corresponding to the uncertainty scores of the first L data points in the sorting result as the data points to be marked; Mark the point cloud data points corresponding to the data points to be marked in the sample lidar point data.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the point cloud semantic segmentation method based on voxel cohesion guided active learning as described in any one of claims 1 to 3 is implemented.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for point cloud semantic segmentation based on voxel cohesion guided active learning as described in any one of claims 1 to 3 is implemented.

7. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method for point cloud semantic segmentation based on voxel cohesion guided active learning as described in any one of claims 1 to 3 is implemented.

Images