Point cloud semantic segmentation method based on dynamic uncertainty active learning strategy

Abstract

The invention provides a point cloud semantic segmentation method and system based on a dynamic uncertainty active learning strategy, and belongs to the technical field of pavement monitoring, and the method comprises the steps: obtaining a plurality of labeled first super-point samples and a plurality of unlabeled second super-point samples based on a three-dimensional point cloud data sample; training the initial model based on all the first super-point samples to obtain a first training model; predicting each second super-point sample by using the first training model to obtain the actual uncertainty of each second super-point sample; screening all the second super-point samples based on the actual uncertainty to obtain a plurality of candidate super-point samples; constructing a training data set based on all candidate super-point samples; and performing updating training on the first training model based on the training data set to obtain a point cloud semantic segmentation model. Through intelligent screening and optimization of annotation sample distribution, the category diversity is improved, and the annotation cost is reduced, so that the segmentation precision and generalization ability of the model are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of road surface monitoring, and in particular to a point cloud semantic segmentation method based on a dynamic uncertainty active learning strategy. Background Art

[0002] With the rapid development of fields such as autonomous driving, urban planning, and remote sensing modeling, 3D LiDAR point cloud semantic segmentation technology has become particularly important. LiDAR can obtain high-precision 3D point cloud data by emitting laser pulses and measuring their reflection time. These data contain rich spatial and color information. However, the high dimensionality and sparsity of point cloud data make the semantic segmentation task complex and demanding on computing resources, and usually rely on deep learning technology to achieve accurate classification. At present, although the fully supervised learning-based method can provide high-precision segmentation results, its performance is highly dependent on large-scale, high-quality annotated datasets. Since the annotation process of 3D point cloud data is cumbersome and expensive, the high annotation cost and limited annotation data have become the main bottleneck restricting the improvement of model performance.

[0003] In order to solve the problem of scarce and high cost of labeled data, active learning methods have been introduced into the field of point cloud semantic segmentation, which intelligently selects the most informative samples for labeling to reduce the number of required annotations. However, existing active learning methods still have shortcomings when processing large-scale point cloud data. These methods often ignore the spatial distribution of samples and the diversity of categories, resulting in the concentration of labeled samples in certain high-uncertainty areas while ignoring other important areas, limiting the generalization ability and segmentation accuracy of the model. In addition, existing methods lack efficient and accurate strategies in the process of merging and dividing super points, making it difficult to significantly reduce the labeling cost while ensuring labeling accuracy.

[0004] Therefore, there is an urgent need for an innovative active learning strategy that can optimize the distribution of labeled samples, improve category diversity, and effectively reduce the labeling cost, so as to promote the widespread application and development of 3D point cloud semantic segmentation technology in various application fields. Summary of the invention

[0005] The present invention provides a point cloud semantic segmentation method, system, electronic device and storage medium based on a dynamic uncertainty active learning strategy, which are used to solve the defects in the prior art and achieve.

[0006] The present invention provides a point cloud semantic segmentation method based on a dynamic uncertainty active learning strategy, comprising the following steps: Input the three-dimensional point cloud data into the point cloud semantic segmentation model to obtain the semantic category of each point in the three-dimensional point cloud data; wherein the point cloud semantic segmentation model is constructed based on the following method: Based on the three-dimensional point cloud data samples, a plurality of first superpoint samples and a plurality of second superpoint samples are obtained, wherein the first superpoint samples are labeled superpoint samples, and the second superpoint samples are unlabeled superpoint samples; Training an initial model based on all of the first superpoint samples to obtain a first training model; Use the first training model to predict each of the second super-point samples to obtain an actual uncertainty of each of the second super-point samples; Screening all the second super-point samples based on the actual uncertainty to obtain a plurality of candidate super-point samples; Based on all the candidate superpoint samples, construct a training data set; The first training model is updated and trained based on the training data set to obtain the point cloud semantic segmentation model.

[0007] According to a point cloud semantic segmentation method based on a dynamic uncertainty active learning strategy provided by the present invention, the second super point sample includes a plurality of three-dimensional points; the using the first training model to predict each of the second super point samples to obtain the actual uncertainty of each of the second super point samples specifically includes: Dividing all three-dimensional points included in each of the second super-point samples to obtain a first point set and a second point set, wherein the first point set is a set of three-dimensional points with the highest category proportion in the second super-point sample; Calculate the information entropy of each of the three-dimensional points using the first training model to obtain a first uncertainty of each three-dimensional point in the first point set and a second uncertainty of each three-dimensional point in the second point set; Calculate an initial uncertainty of each of the second super-point samples based on all of the first uncertainties and all of the second uncertainties; Multiply each of the initial uncertainties by the category balance weight coefficient of each of the second super-point samples to obtain the actual uncertainty of each of the second super-point samples, where the category balance weight coefficient of each of the second super-point samples is determined based on the proportion of the second super-point samples of each category in the super-point samples of all categories.

[0008] According to a point cloud semantic segmentation method based on a dynamic uncertainty active learning strategy provided by the present invention, the screening of all the second super point samples based on the actual uncertainty to obtain a plurality of candidate super point samples specifically includes: In each round of screening, the second super-point sample with the highest actual uncertainty is selected as the candidate super-point sample, and based on the spatial distance and feature similarity between the candidate super-point sample and the remaining second super-point samples, the actual uncertainty of the remaining second super-point samples is dynamically attenuated until a preset number of candidate super-point samples are screened out.

[0009] According to a point cloud semantic segmentation method based on a dynamic uncertainty active learning strategy provided by the present invention, the training data set is constructed based on all the candidate superpoint samples, specifically including: Establish a multi-branch tree for preserving the neighbor relationship of super points; Based on the multi-branch tree, selecting a first adjacent super-point sample corresponding to each of the candidate super-point samples from all super-point samples; Determining whether each of the first adjacent super-point samples is similar to a corresponding candidate super-point sample; Determining second adjacent superpoint samples from all of the first adjacent superpoint samples, and calculating a purity value of each of the second adjacent superpoint samples, wherein the second adjacent superpoint samples are similar to the corresponding candidate superpoint samples; Determine, from all of the second adjacent super-point samples, a third adjacent super-point sample whose purity value is higher than a preset purity threshold, and a fourth adjacent super-point sample whose purity value is not higher than a preset purity threshold; Dividing the fourth adjacent superpoint sample into a plurality of first sub-superpoint samples, and determining a second sub-superpoint sample having a purity value higher than a preset purity threshold from all the first sub-superpoint samples; The training data set is constructed based on all the candidate superpoint samples, all the third adjacent superpoint samples and all the second sub-superpoint samples.

[0010] According to a point cloud semantic segmentation method based on a dynamic uncertainty active learning strategy provided by the present invention, the step of determining whether each of the first adjacent superpoint samples is similar to a corresponding candidate superpoint sample specifically includes: Solve the first normal vector of the current first adjacent superpoint sample and the second normal vector of the corresponding candidate superpoint sample; Calculating the cosine value of the angle between the first normal vector and the second normal vector; Determine whether the cosine value of the angle is less than a preset normal vector similarity threshold; When the cosine value of the angle is less than a preset normal vector similarity threshold, calculating the color similarity between the current first adjacent super point sample and the corresponding candidate super point sample; Determining whether the color similarity is less than a preset color similarity threshold; When the color similarity is less than a preset color similarity threshold, it is determined that the current first adjacent super point sample is similar to the corresponding candidate super point sample.

[0011] According to a point cloud semantic segmentation method based on a dynamic uncertainty active learning strategy provided by the present invention, the updating and training of the first training model based on the training data set to obtain the point cloud semantic segmentation model specifically includes: Performing update training on the first training model based on the training data set to obtain a second training model; The second training model is used as the first training model, and the first training model is continued to be used to predict each of the second superpoint samples to obtain the actual uncertainty of each of the second superpoint samples, all of the second superpoint samples are screened based on the actual uncertainty to obtain multiple candidate superpoint samples, a training data set is constructed based on all of the candidate superpoint samples, and the first training model is updated based on the training data set until a preset stop condition is reached to obtain the point cloud semantic segmentation model.

[0012] The present invention also provides a point cloud semantic segmentation system based on a dynamic uncertainty active learning strategy, comprising the following modules: A first processing module, used for inputting the three-dimensional point cloud data into the point cloud semantic segmentation model to obtain the semantic category of each point in the three-dimensional point cloud data; The second processing module is used to obtain a plurality of first superpoint samples and a plurality of second superpoint samples based on the three-dimensional point cloud data samples, wherein the first superpoint samples are labeled superpoint samples and the second superpoint samples are unlabeled superpoint samples.

[0013] A training module, used for training an initial model based on all the first superpoint samples to obtain a first training model; The second processing module is further used to use the first training model to predict each of the second super-point samples to obtain an actual uncertainty of each of the second super-point samples; The second processing module is further used to screen all the second super-point samples based on the actual uncertainty to obtain a plurality of candidate super-point samples; The second processing module is further used to construct a training data set based on all the candidate superpoint samples; The training module is also used to update the first training model based on the training data set to obtain the point cloud semantic segmentation model.

[0014] The present invention also provides 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 program, a point cloud semantic segmentation method based on a dynamic uncertainty active learning strategy as described above is implemented.

[0015] 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 a dynamic uncertainty active learning strategy as described in any one of the above is implemented.

[0016] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-mentioned point cloud semantic segmentation methods based on dynamic uncertainty active learning strategies.

[0017] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: By dividing the three-dimensional point cloud data samples into multiple labeled first superpoint samples and unlabeled second superpoint samples, the orderly classification and efficient management of data are achieved, ensuring the effective distinction between labeled and unlabeled data. By training the initial model based on all labeled first superpoint samples, a preliminary model with basic semantic recognition ability is constructed, which can quickly learn the basic semantic features of point cloud data. By using the first training model to predict each unlabeled second superpoint sample and calculate its actual uncertainty, the annotation value of each unlabeled superpoint is accurately evaluated to ensure the efficient use of annotation resources. By screening all second superpoint samples based on actual uncertainty, multiple candidate superpoint samples are obtained, thereby optimizing the allocation of annotation resources and giving priority to the samples with the most value in improving model performance. By constructing a training data set based on all candidate superpoint samples, high-value and diverse annotation samples are integrated, the quality and representativeness of the training data are improved, and the learning effect of the model is enhanced. By updating the first training model based on the training data set, a point cloud semantic segmentation model with better performance and stronger generalization ability is obtained, which significantly improves the segmentation accuracy and robustness of the model in complex urban outdoor scenes. In summary, the present invention optimizes the distribution of labeled samples, improves the diversity of categories, and effectively reduces the labeling cost through a series of systematic steps, thereby promoting the widespread application and development of 3D point cloud semantic segmentation technology in various application fields. This active learning strategy not only improves the segmentation accuracy and generalization ability of the model, but also significantly reduces the overall labeling investment through efficient labeling resource utilization, ensuring the efficiency and economy of the system in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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 creative work.

[0019] Figure 1 It is one of the flow charts of the point cloud semantic segmentation method based on the dynamic uncertainty active learning strategy provided by the present invention.

[0020] Figure 2 This is the second flow chart of the point cloud semantic segmentation method based on the dynamic uncertainty active learning strategy provided by the present invention.

[0021] Figure 3 This is the third flow chart of the point cloud semantic segmentation method based on the dynamic uncertainty active learning strategy provided by the present invention.

[0022] Figure 4 It is a schematic diagram of the ablation experiment results of the SPDU and SPMD modules provided by the present invention.

[0023] Figure 5 It is a schematic diagram of the performance comparison between the present invention and other active learning methods on the Semantic3D dataset.

[0024] Figure 6 It is a schematic diagram of comparison results of different active learning strategies provided by the present invention.

[0025] Figure 7 It is a structural schematic diagram of the point cloud semantic segmentation system based on the dynamic uncertainty active learning strategy provided by the present invention.

[0026] Figure 8 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] It should be noted that, in the description of the present invention, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "include one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. The orientation or positional relationship indicated by the terms "upper", "lower", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the system or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0029] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0030] Combine the following Figure 1-Figure 8 The present invention describes a point cloud semantic segmentation method, system, electronic device and storage medium based on a dynamic uncertainty active learning strategy.

[0031] Figure 1 This is one of the flow charts of the point cloud semantic segmentation method based on the dynamic uncertainty active learning strategy provided by the present invention, such as Figure 1 As shown, including but not limited to the following steps: Step 1: Input the 3D point cloud data into the point cloud semantic segmentation model to obtain the semantic category of each point in the 3D point cloud data.

[0032] In this embodiment, the three-dimensional point cloud data of an urban outdoor scene is first obtained, which is collected by a high-precision three-dimensional laser radar (LiDAR) system. In order to ensure the quality of the point cloud data, the original data is first preprocessed, including steps such as denoising, downsampling, and coordinate normalization to eliminate measurement errors and redundant points and improve the efficiency and accuracy of subsequent processing. The preprocessed point cloud data is input into a pre-trained point cloud semantic segmentation model, which has been trained with a large-scale and diverse data set and has good generalization ability and high-precision semantic classification performance.

[0033] By inputting the preprocessed 3D point cloud data into the point cloud semantic segmentation model, the model semantically classifies each point in the point cloud, identifies and labels its semantic category, such as roads, buildings, vehicles, pedestrians and vegetation.

[0034] The specific training method of the point cloud semantic segmentation model will be described in detail in steps 2 to 7 of the subsequent embodiments.

[0035] Step 2: Based on the three-dimensional point cloud data samples, a plurality of first super point samples and a plurality of second super point samples are obtained, where the first super point samples are labeled super point samples and the second super point samples are unlabeled super point samples.

[0036] Specifically, the GlobalEnergy algorithm is first used to perform super-point segmentation on the preprocessed point cloud data. The GlobalEnergy algorithm clusters adjacent points with similar features into super-points by evaluating the spatial distribution and feature similarity of points in the point cloud. After the segmentation is completed, some super-point samples are labeled as first super-point samples based on the existing labeled data. These labeled first super-point samples contain rich semantic information and can provide high-quality training data for the training of the initial model. At the same time, the unlabeled second super-point samples are retained for the subsequent active learning process. By intelligently selecting and labeling these unlabeled samples, the performance of the model is gradually improved.

[0037] In this way, the 3D point cloud data is systematically divided into labeled and unlabeled super-point samples, which not only achieves orderly management and efficient use of data, but also lays a solid foundation for the implementation of active learning strategies. The labeled first super-point samples are used as the initial training set to ensure that the initial model has basic semantic recognition capabilities; while the unlabeled second super-point samples are gradually introduced into high-value labeled samples through uncertainty evaluation and screening in subsequent steps to optimize the model's learning process.

[0038] Step 3: Train the initial model based on all first superpoint samples to obtain the first training model.

[0039] Specifically, all the labeled first superpoint samples are first input into the deep learning framework as training data, such as using the RandLA-Net model architecture. In order to ensure the effectiveness of the training process and the generalization ability of the model, the training data is first preprocessed, including feature normalization and data enhancement, such as random rotation, scaling, and adding noise, to simulate different environmental conditions and enhance the robustness of the model.

[0040] Then, the initial model is trained using these labeled first superpoint samples. During the training process, the model continuously optimizes its weight parameters through the back-propagation algorithm to minimize the loss function between the predicted category and the actual category. Since the first superpoint samples already contain rich semantic information and high-quality labeled data, the initial model can quickly learn the basic semantic features and spatial structure of the point cloud data, thus having preliminary semantic segmentation capabilities.

[0041] Step 4: Use the first training model to predict each second super-point sample to obtain the actual uncertainty of each second super-point sample.

[0042] In the early stages of active learning, the model may have low prediction accuracy for certain types or features of point cloud data due to the limitation of insufficient labeled data. In particular, when the model is not fully exposed to data of a certain semantic category, its prediction accuracy is low. Generally, due to the semantic coherence of point cloud data in space, such inaccurate predictions may lead to generally high uncertainty in connected areas. In order to avoid continuously selecting super-points for labeling in these high-uncertainty areas, the present invention specifically introduces the consideration of spatial dispersion when measuring super-point uncertainty. This measure aims to optimize the distribution of labeled samples and ensure that the selected super-points are better representative and diverse in space, thereby improving the learning efficiency and generalization ability of the model.

[0043] In one possible implementation, refer to Figure 2 , Figure 2 This is the second flow chart of the point cloud semantic segmentation method based on the dynamic uncertainty active learning strategy provided by the present invention. Step 4 specifically includes steps 41-44: Step 41: Divide all three-dimensional points included in each second super point sample to obtain a first point set and a second point set, wherein the first point set is a set consisting of three-dimensional points with the highest category proportion in the second super point sample.

[0044] In the specific operation process, all the three-dimensional points inside each second superpoint sample are first counted according to their predicted semantic categories to determine the category with the highest proportion in the superpoint. Then, all points belonging to the dominant category are classified into the first point set, and points of other categories are classified into the second point set. This division method ensures that in the subsequent uncertainty calculation, the points of the dominant category can be focused on, thereby reducing the uncertainty deviation caused by category diversity. In addition, by clarifying the division of dominant and secondary categories, superpoints with lower purity can be identified more effectively, so that appropriate merging or division strategies can be adopted in subsequent steps to further optimize the selection of labeled samples.

[0045] Step 42: Use the first training model to calculate the information entropy of each three-dimensional point to obtain a first uncertainty for each three-dimensional point in the first point set and a second uncertainty for each three-dimensional point in the second point set.

[0046] Specifically, the first training model trained in step 3 is used to predict the semantic category of each 3D point in the second superpoint sample. The model outputs the probability distribution of each point belonging to each category. , where c represents the category and pv represents the point v. Based on these probability distributions, the information entropy Uv of each point is calculated as its uncertainty, and the formula is as follows: ; Wherein, C is the total number of categories. The higher the information entropy Uv, the greater the prediction uncertainty of the model for the point. Subsequently, for the first point set and the second point set obtained in step 41, the information entropy of each point is calculated respectively to obtain the first uncertainty Uvi of each point in the first point set and the second uncertainty Uvj of each point in the second point set.

[0047] Step 43: Calculate the initial uncertainty of each second super-point sample based on all the first uncertainties and all the second uncertainties.

[0048] Specifically, first, the uncertainties Uvi of all points in the first point set obtained in step 42 are accumulated to obtain the total uncertainty of the superpoint sample on the dominant category. At the same time, the uncertainties Uvj of all points in the second point set are accumulated to obtain the total uncertainty of the superpoint sample on the non-dominant category. Then, the two total uncertainties are combined using the following formula to calculate the initial uncertainty of the superpoint sample: : ; Among them, ψ represents the majority point set inside the super point, and ϕ represents the other point set inside the super point. In this way, the initial uncertainty not only reflects the uncertainty of the dominant category points, but also reduces the probability of super point samples with lower purity being misselected by subtracting the uncertainty of non-dominant category points, thereby reducing the negative impact of noise data on model training.

[0049] Step 44: Multiply each initial uncertainty by the category balance weight coefficient of each second super-point sample to obtain the actual uncertainty of each second super-point sample. The category balance weight coefficient of each second super-point sample is determined based on the proportion of the second super-point samples of each category in the super-point samples of all categories.

[0050] Specifically, the category balance weight coefficient ω(c) is calculated by the following formula: ; Wherein, c is one of the preset categories, sp is the second super point sample, d(sp) is the category of the second super point sample, U is the set of the second super point samples, and L is the set of the first super point samples.

[0051] Actual uncertainty of the second superpoint sample Calculated by the following formula: ; In this way, the actual uncertainty It not only reflects the prediction uncertainty of super-point samples in semantic categories, but also takes into account the balance of category distribution. Specifically, when the super-point samples of a certain category account for a high proportion in the whole, its corresponding weight coefficient ω(c) will be lower, thereby reducing the actual uncertainty of the super-point samples of this category and avoiding over-labeling of samples of this category during active learning. On the contrary, for those categories that account for a low proportion in the whole, their weight coefficient ω(c) is higher, thereby increasing their actual uncertainty and increasing the probability of them being selected for labeling.

[0052] Step 5: Screen all second super-point samples based on the actual uncertainty to obtain multiple candidate super-point samples.

[0053] In a possible implementation, step 5 specifically includes the following steps: In each round of screening, the second super-point sample with the highest actual uncertainty is selected as the candidate super-point sample, and based on the spatial distance and feature similarity between the candidate super-point sample and the remaining second super-point samples, the actual uncertainty of the remaining second super-point samples is dynamically attenuated until a preset number of candidate super-point samples are screened out.

[0054] Specifically, after obtaining the actual uncertainty of all second superpoint samples in step 4, the second superpoint sample with the highest uncertainty is selected as the starting point and determined as the first candidate superpoint sample to be labeled. Subsequently, in order to avoid redundant labeling in the local area, the uncertainties of other second superpoint samples near the second superpoint sample as the starting point are dynamically attenuated.

[0055] Based on the feature similarity and spatial dispersion of the remaining second superpoint samples and the candidate superpoint samples, the actual uncertainty of the remaining second superpoint samples is attenuated, thereby reducing the possibility of these remaining second superpoint samples as the next candidate superpoint samples to be annotated. This dynamic attenuation strategy ensures the efficiency and wide coverage of the annotation process. The dynamic attenuation method is as follows: ; in, is the actual uncertainty of the attenuated second superpoint sample, is the actual uncertainty of the second superpoint sample before decay, is the attenuation coefficient, and dist is the Euclidean distance between the center of the second superpoint sample undergoing the attenuation operation and the center of the candidate superpoint sample.

[0056] According to this dynamic attenuation rule, each time a candidate superpoint sample is selected, the actual uncertainty of the remaining second superpoint samples is dynamically updated until the K annotation budgets are exhausted, and finally K candidate superpoint samples are obtained.

[0057] Step 6: Construct a training dataset based on all candidate superpoint samples.

[0058] In this embodiment, after multiple candidate superpoint samples are screened out based on the actual uncertainty in step 5, step 6 is performed next, that is, a training data set is constructed based on all candidate superpoint samples. The reason for this operation is to integrate high-value candidate superpoint samples and their adjacent superpoint samples to form a diverse and representative training data set, thereby optimizing the training effect of the model and improving the accuracy and generalization ability of point cloud semantic segmentation.

[0059] In one possible implementation, refer to Figure 3 , Figure 3 This is the third flow chart of the point cloud semantic segmentation method based on the dynamic uncertainty active learning strategy provided by the present invention. Step 6 specifically includes steps 61-67: Step 61: Establish a multi-branch tree for storing the neighbor relationship of super points.

[0060] Specifically, the spatial position and feature information of all superpoint samples in the constructed training data set are first input into the multitree data structure. A multitree is a hierarchical data structure that can effectively organize and store the proximity relationships between a large number of superpoints. By dividing the superpoints into layers according to their spatial coordinates, the multitree can significantly reduce the query time and computational complexity while keeping the data organized.

[0061] In the actual construction process, the superpoint samples generated by the GlobalEnergy algorithm are used as nodes of the multitree, and each node contains the spatial coordinates, color features, and normal vector information of the superpoint. The system first inserts the superpoint into the corresponding node of the multitree according to its spatial position, ensuring that adjacent or close superpoints are organized at similar levels of the tree. In this way, the neighboring superpoints of any superpoint can be quickly located and accessed by traversing the local area of ​​the multitree, greatly improving the efficiency of neighbor relationship queries.

[0062] Step 62: Based on the multi-branch tree, select the first adjacent super point sample corresponding to each candidate super point sample from all super point samples.

[0063] First, the multi-branch tree T constructed in step 61 is used to query the first adjacent superpoint sample of each candidate superpoint sample. For each candidate superpoint sample, a neighbor search operation is performed in the multi-branch tree to quickly locate the superpoint samples with similar spatial positions and features to the candidate superpoint sample, forming a first adjacent superpoint sample set. In this way, the first adjacent superpoint sample corresponding to each candidate superpoint sample can be efficiently extracted from all superpoint samples to form an accurate and relevant superpoint adjacency relationship network.

[0064] Step 63: Determine whether each first adjacent superpoint sample is similar to the corresponding candidate superpoint sample.

[0065] In this embodiment, after selecting the first adjacent superpoint sample corresponding to each candidate superpoint sample from all superpoint samples based on the multi-branch tree in step 62, step 63 is then performed, i.e., determining whether each first adjacent superpoint sample has similarity with the corresponding candidate superpoint sample in normal vector. The reason for this determination is to accurately identify the superpoint samples located on the same plane, thereby ensuring the effectiveness and accuracy of the merging operation and further improving the segmentation accuracy and generalization ability of the model.

[0066] In a possible implementation, step 63 specifically includes the following steps: Solve the first normal vector of the current first adjacent superpoint sample and the second normal vector of the corresponding candidate superpoint sample; Calculate the cosine of the angle between the first normal vector and the second normal vector; Determine whether the cosine value of the angle is less than a preset normal vector similarity threshold; When the angle cosine value is less than the preset normal vector similarity threshold, the color similarity between the current first adjacent super point sample and the corresponding candidate super point sample is calculated; Determine whether the color similarity is less than a preset color similarity threshold; When the color similarity is less than a preset color similarity threshold, it is determined that the current first adjacent super point sample is similar to the corresponding candidate super point sample.

[0067] Specifically, the Singular Value Decomposition (SVD) method is first used to calculate the normal vectors Vj and Vi of the current first adjacent superpoint sample Spj and the corresponding candidate superpoint sample Spi. The normal vector is an important geometric feature that describes the surface orientation of the superpoint sample. The SVD method can accurately extract the main direction information of the superpoint sample.

[0068] Then, the cosine value of the angle αj between the two normal vectors is calculated as follows: ; Among them, Vi and Vj are the normal vectors of the candidate superpoint sample and the first adjacent superpoint sample respectively, ⋅ represents the vector dot product, and are the modulus of the normal vectors. The closer the angle cosine value αj is to 1, the more similar the directions of the two normal vectors are, and the more consistent the planes where the superpoint samples are located are.

[0069] Next, the calculated angle cosine value αj is compared with the preset normal vector similarity threshold α. If αj is less than the threshold α, it is determined that the normal vector similarity of the two superpoint samples is low, indicating that they are not located on the same plane. At this time, the system skips the subsequent color similarity calculation and continues to process the next superpoint sample pair.

[0070] When the angle cosine value αj is greater than or equal to the threshold α, the color similarity βj of the two superpoint samples is further calculated. The color similarity is calculated by converting the RGB color information into the LAB color space, and setting weights of 0.5, 1.0 and 1.0 for the brightness (L) and chromaticity (A, B) components respectively to reduce the impact of changes in lighting conditions on color similarity. Specifically, the RGB color values ​​of the two superpoint samples are first converted into LAB color values, and then their color distances in the LAB space are calculated and compared according to the preset color similarity threshold β. If βj is less than the threshold β, it is determined that the two superpoint samples are not similar enough in color features, and the system continues to skip subsequent similarity confirmations and process the next superpoint sample pair. When the color similarity βj is greater than or equal to the threshold β, it is determined that the current first adjacent superpoint sample Spj and the corresponding candidate superpoint sample Spi have similar features. This judgment shows that the two superpoint samples are not only located on the same plane in terms of geometric structure, but also highly consistent in color features, and have the potential to be in the same semantic category. Therefore, the two superpoint samples are marked as similar and prepared for the merge operation.

[0071] Step 64: Determine second neighboring superpoint samples from all first neighboring superpoint samples, and calculate the purity value of each second neighboring superpoint sample, the second neighboring superpoint sample being similar to the corresponding candidate superpoint sample.

[0072] Specifically, first, based on the first adjacent superpoint samples that have been confirmed to be similar in step 63, the neighboring superpoints of these first adjacent superpoint samples are extracted from the multi-branch tree structure to form a set of second adjacent superpoint samples. This step is intended to expand the coverage of the area around the candidate superpoint samples to ensure that the purity and semantic consistency of the superpoint samples are evaluated over a larger range. Next, the purity is calculated for each second adjacent superpoint sample. The purity value is defined as the proportion of the dominant category in the superpoint sample. The specific calculation method is to count the number of points of each semantic category in the superpoint sample, determine the category with the highest proportion, and calculate its proportion. The formula is as follows: φj=the number of points with the highest category proportion / the total number of points in the superpoint sample. Among them, φj represents the purity value of the second adjacent superpoint sample Spj.

[0073] Step 65: Determine, from all second-adjacent super-point samples, a third-adjacent super-point sample whose purity value is higher than a preset purity threshold, and a fourth-adjacent super-point sample whose purity value is not higher than the preset purity threshold.

[0074] Specifically, first traverse all the second-neighboring superpoint samples obtained in step 64, and evaluate the purity value φj of each sample. The purity value is defined as the proportion of the dominant category in the superpoint sample. Compare the purity value of each second-neighboring superpoint sample with the preset purity threshold φ. If the purity value φj of a superpoint sample is greater than or equal to the threshold φ, the superpoint sample is classified as the third-neighboring superpoint sample, indicating that it has a high semantic consistency and is suitable for direct inclusion in the training data set; conversely, if the purity value φj of a superpoint sample is less than the threshold φ, the superpoint sample is classified as the fourth-neighboring superpoint sample, indicating that it contains multiple semantic categories and may introduce noise.

[0075] Step 66: Divide the fourth adjacent super point sample into a plurality of first sub-super point samples, and determine a second sub-super point sample having a purity value higher than a preset purity threshold from all the first sub-super point samples.

[0076] In the specific operation process, each superpoint sample classified as the fourth adjacent superpoint sample in step 65 is first further subdivided. A density-based clustering algorithm, such as DBSCAN (Density-Based Spatial Clustering of Applications with Noise), is used to divide the three-dimensional points in each fourth adjacent superpoint sample to generate multiple first sub-superpoint samples. This clustering method can automatically identify and segment point groups with obvious spatial aggregation according to the spatial density and distribution characteristics of point cloud data, thereby effectively dividing complex and diverse low-purity superpoint samples into multiple sub-superpoint samples with higher semantic consistency.

[0077] After the division is completed, the purity value φk of each generated first sub-superpoint sample is calculated, and the formula is as follows: φk = the number of points with the highest category proportion / the total number of points in the first sub-superpoint sample. Among them, φk represents the purity value of the first sub-superpoint sample Spk. Through this calculation, the semantic consistency of each sub-superpoint sample can be evaluated to ensure that only those second sub-superpoint samples Spk′ with purity values ​​higher than the preset purity threshold φ are selected to be included in the training data set, while sub-superpoint samples with low purity values ​​are further screened out or processed in other ways to avoid introducing low-purity samples containing multiple semantic categories into the training data.

[0078] Step 67: construct a training data set based on all candidate superpoint samples, all third adjacent superpoint samples and all second sub-superpoint samples.

[0079] In the specific operation process, the multiple candidate superpoint samples screened in step 5, all the third adjacent superpoint samples determined in step 65, and all the second sub-superpoint samples screened in step 66 are first aggregated to form a comprehensive superpoint set. In order to ensure the diversity and representativeness of the training data set, this comprehensive superpoint set is further enhanced and balanced. For example, to address the problem of uneven category distribution, oversampling or undersampling methods can be used to ensure that each semantic category has a sufficient number of samples in the training data set. In addition, data enhancement techniques such as random rotation, scaling, and adding noise can be used to increase the diversity of training data and improve the model's adaptability to different environments and conditions.

[0080] Next, the processed comprehensive superpoint set is organized into a training data set and input into the deep learning framework for model training. During the training process, the optimized training data set is used to continuously adjust the model parameters through the back propagation algorithm to minimize the loss function between the predicted results and the actual labels. Since the training data set contains high-quality superpoint samples from different spatial regions and diverse semantic categories, the model can learn more comprehensive and detailed semantic features, improving its segmentation accuracy and robustness in complex urban outdoor scenes.

[0081] Step 7: Update the first training model based on the training data set to obtain a point cloud semantic segmentation model.

[0082] In this embodiment, after constructing a training data set based on all candidate superpoint samples, the third adjacent superpoint samples, and the second sub-superpoint samples in step 6, step 7 is then performed, that is, updating and training the first training model based on the training data set to obtain the final point cloud semantic segmentation model. The reason for this operation is to use the optimized high-quality training data set to further improve the semantic recognition ability and generalization performance of the model, thereby achieving more accurate and more robust point cloud semantic segmentation.

[0083] In a possible implementation, step 7 specifically includes the following steps: Performing update training on the first training model based on the training data set to obtain a second training model; The second training model is used as the first training model, and the first training model is continued to be used to predict each second super point sample to obtain the actual uncertainty of each second super point sample, all the second super point samples are screened based on the actual uncertainty to obtain multiple candidate super point samples, a training data set is constructed based on all the candidate super point samples, and the first training model is updated based on the training data set until the preset stopping condition is reached to obtain a point cloud semantic segmentation model.

[0084] Specifically, the first training model is updated and trained using the optimized training data set to generate a second training model with better performance. During the training process, the weight parameters of the model are continuously adjusted through the back-propagation algorithm to minimize the loss function between the predicted results and the actual labels. Since the training data set has been strictly screened and optimized, and contains high-uncertainty and high-purity candidate super-point samples, third adjacent super-point samples, and second sub-super-point samples, the second training model can learn more comprehensive and detailed semantic features and spatial structures during the training process, significantly improving its segmentation accuracy and robustness in complex urban outdoor scenes.

[0085] Subsequently, the second training model is used as the new first training model to enter the next round of active learning cycle. The updated first training model is continued to be used to predict all unlabeled second superpoint samples, calculate their actual uncertainty, and screen the unlabeled samples based on the uncertainty to select new candidate superpoint samples. Subsequently, the training data set is rebuilt based on these new candidate superpoint samples, and the first training model is further updated and trained. Through this iterative training and optimization process, high-value labeled samples can be continuously introduced to gradually improve the performance and generalization ability of the model.

[0086] Furthermore, in order to verify the actual effect of the point cloud semantic segmentation model of the present invention, the present invention conducted experiments on the Semantic3D three-dimensional lidar point cloud dataset of large-scale urban outdoor scenes. The Semantic3D dataset includes large-scale three-dimensional point cloud data from different regions in Europe, and is mainly used to study the semantic segmentation technology of three-dimensional point clouds. The dataset is acquired by ground laser scanning equipment and covers a wide range of urban outdoor scenes, including churches, streets, railroad tracks, squares, castles, etc., with a total of more than 4 billion points, divided into 8 specific semantic categories. The Semantic3D dataset was selected as the experimental object to verify the effectiveness and robustness of the present invention in processing large-scale complex urban point cloud data, and to ensure that the proposed active learning strategy can maintain excellent performance in a diverse and highly complex data environment.

[0087] In order to efficiently process the Semantic3D dataset, the present invention first uses the grid downsampling technology to reduce the amount of original point cloud data to the tens of millions level. This preprocessing step not only significantly reduces the consumption of computing resources, but also retains the key geometric and semantic features of the point cloud data. Subsequently, the GlobalEnergyModel method is used to perform super-point segmentation on the downsampled points, and the point cloud data is further organized into multiple super-point samples with spatial coherence and feature similarity. In addition, in order to efficiently manage the neighbor relationship between superpoints, a multi-branch tree structure is constructed. This data structure optimizes the storage and query efficiency of the neighbor relationship of superpoints. In order to improve the generalization performance of the model, the input point cloud data is randomly rotated, scaled, and random noise is added to simulate different environmental conditions and enhance the robustness of the model. In the framework of active learning, RandLA-Net is used as the segmenter, and the Adam optimizer is used, where β1 is set to 0.9 and β2 is set to 0.99. The initial learning rate of the learning process is set to 0.01. After each training round, the learning rate decreases by 8%. The total number of training rounds is 30, and the batch size is set to 6. In each active learning cycle, the attenuation coefficient of the uncertainty dynamic measurement method is set to the maximum distance of the superpoints in the scene, and 3000 spatially dispersed and highly uncertain candidate superpoints with category diversity are screened from the unlabeled set. For the merging operation of superpoints, the normal vector similarity threshold is set to 0.97. When calculating the similarity of color features, the weights of the LAB components are set to 0.5, 1.0, and 1.0 respectively, and the similarity threshold is set to 0.95. In addition, the identified noise superpoints are further divided and labeled according to the purity threshold of 0.9. Finally, these labeled superpoints are added to the labeled set. All codes are implemented based on the Tensorflow framework and experiments are carried out on a server equipped with an NVIDIA A100 GPU, with 80GB of device memory, Ubuntu 22.04 operating system, and CUDA version 12.2. Through this series of implementation steps, the present invention can efficiently process large-scale point cloud data, optimize the selection of labeled samples and the training process of the model, and ensure that the model has excellent segmentation performance in diverse and complex environments.

[0088] The present invention uses Mean Intersection over Union (MIoU) and OverallAccuracy (OA) to evaluate the performance of the model trained using the active learning strategy of the present invention on the validation set. MIoU measures the average overlap between the model-predicted categories and the true categories, reflecting the segmentation accuracy of the model in each category; OA measures the proportion of points predicted correctly by the model to the total number of points, reflecting the overall classification accuracy of the model. The specific definitions are as follows: ; ; ; Among them, Tp, Fp, Tn and Fn represent the number of true positives, false positives, true negatives and false negatives respectively, and C is the predicted label category. Through these two indicators, the segmentation performance of the model in different categories and overall can be comprehensively evaluated to ensure the effectiveness and superiority of the proposed active learning strategy in practical applications.

[0089] In order to verify the effectiveness of the uncertainty dynamic measurement method SPDU and the super point merging and division module SPMD, the present invention conducted a series of ablation experiments on the Semantic3D dataset. Figure 4 , Figure 4 It is a schematic diagram of the ablation experiment results of the SPDU and SPMD modules provided by the present invention. In order to verify the effectiveness of each module, experiments with different combinations were conducted, including using only SPDU, only SPMD, and a combination of SPDU and SPMD. The benchmark experiment uses randomly sampled superpoints for annotation and training. The annotation budget for each round of active learning cycle is set to 3000 superpoints, and a total of 3 rounds are performed. The final mIoU of the model on the test set is used as an indicator to measure the segmentation effect.

[0090] Depend on Figure 4 It can be seen that when using the superpoint random sampling strategy, the model predicts mIoU of only 54.7%. When the SPDU module is applied, the model segmentation accuracy is significantly improved to 63.2%; when only the SPMD module is used, the model performance is slightly improved to 58.3%; and when the two modules are mixed, the model performance can reach 67.8% accuracy when a total of 9000 superpoints are annotated. This shows the effectiveness of the SPDU and SPMD modules and verifies the significant advantages of the active learning strategy proposed in this invention in improving the model segmentation accuracy.

[0091] In order to further verify the superiority of the active learning framework proposed in this paper, it is compared with five other advanced active learning methods on the Semantic3D dataset. Figure 5 , Figure 5 is a schematic diagram of the performance comparison between the present invention and other active learning methods on the Semantic3D dataset, such as Figure 5 As shown in the figure, the horizontal axis represents the consumed annotation budget, that is, the number of annotated superpoints, which increases with the iteration of the active learning cycle; the vertical axis represents the performance of the model on the test set after each round of active learning cycle, measured by the mIoU indicator. Figure 5The active learning method of the present invention is significantly superior to other methods. Specifically, the segmentation performance of the model is higher when the same annotation budget is consumed. Especially in the early active learning cycle, when the annotation budget is less, the segmentation accuracy of the model trained by this method is much higher than that of other methods. In addition, under the same accuracy requirement, this method consumes less annotation budget. Specifically, refer to Figure 6 , Figure 6 is a schematic diagram of the comparison results of different active learning strategies provided by the present invention, Figure 6 The annotation budget required for different active learning methods to achieve 90% full-supervision performance (mIoU) on the Semantic3D test set is shown.

[0092] Depend on Figure 6 It can be seen that compared with the random selection of superpoint sample strategy, the active learning framework proposed in this paper reduces the annotation budget by up to 56.1% while achieving the same performance; even compared with the current advanced active learning method SSDR, the annotation budget required by this invention is reduced by 21.8%. This result shows that this method has a significant advantage in reducing the annotation budget, further verifying its high efficiency and economy in practical applications.

[0093] In summary, through experiments and analysis on the Semantic3D dataset, the effectiveness and superiority of the active learning strategy proposed in the present invention in the task of semantic segmentation of point clouds have been verified. Experimental results show that the use of grid downsampling and GlobalEnergyModel methods for super-point segmentation, combined with the active learning framework of SPDU and SPMD modules, can significantly improve the segmentation accuracy and generalization ability of the model, while significantly reducing the annotation budget. Ablation experiments and comparative tests further prove the effectiveness of each module and the superiority of the overall method, ensuring the efficiency, economy and reliability of the present invention in practical applications, and providing solid technical support and guarantee for the development of three-dimensional point cloud semantic segmentation technology.

[0094] Reference Figure 7 , Figure 7 : is a structural diagram of a point cloud semantic segmentation system based on a dynamic uncertainty active learning strategy provided by the present invention, the system comprising: The first processing module is used to input the three-dimensional point cloud data into the point cloud semantic segmentation model to obtain the semantic category of each point in the three-dimensional point cloud data; The second processing module is used to obtain a plurality of first superpoint samples and a plurality of second superpoint samples based on the three-dimensional point cloud data samples, wherein the first superpoint samples are labeled superpoint samples and the second superpoint samples are unlabeled superpoint samples.

[0095] A training module, used for training an initial model based on all first superpoint samples to obtain a first training model; The second processing module is further used to predict each second super point sample using the first training model to obtain an actual uncertainty of each second super point sample; The second processing module is further used to screen all second super-point samples based on the actual uncertainty to obtain a plurality of candidate super-point samples; The second processing module is also used to construct a training data set based on all candidate superpoint samples; The training module is also used to update the first training model based on the training data set to obtain a point cloud semantic segmentation model.

[0096] In a possible implementation manner, the second processing module is further configured to: Divide all the three-dimensional points included in each second super point sample to obtain a first point set and a second point set, where the first point set is a set of three-dimensional points with the highest category proportion in the second super point sample; Calculate the information entropy of each three-dimensional point using the first training model to obtain a first uncertainty of each three-dimensional point in the first point set and a second uncertainty of each three-dimensional point in the second point set; Calculating an initial uncertainty of each second super-point sample based on all first uncertainties and all second uncertainties; Each initial uncertainty is multiplied by the category balance weight coefficient of each second super-point sample to obtain the actual uncertainty of each second super-point sample. The category balance weight coefficient of each second super-point sample is determined based on the proportion of the second super-point samples of each category in the super-point samples of all categories.

[0097] In a possible implementation, the second processing module is further used to: in each round of screening, select the second super-point sample with the highest actual uncertainty as the candidate super-point sample, and based on the spatial distance and feature similarity between the candidate super-point sample and the remaining second super-point samples, dynamically attenuate the actual uncertainty of the remaining second super-point samples until a preset number of candidate super-point samples are screened out.

[0098] In a possible implementation manner, the second processing module is further configured to: Establish a multi-branch tree for preserving the neighbor relationship of super points; Based on the multi-branch tree, the first adjacent super-point sample corresponding to each candidate super-point sample is selected from all super-point samples; Determine whether each first adjacent superpoint sample is similar to the corresponding candidate superpoint sample; Determining second neighboring superpoint samples from all first neighboring superpoint samples, and calculating the purity value of each second neighboring superpoint sample, the second neighboring superpoint sample being similar to the corresponding candidate superpoint sample; Determine, from all the second adjacent super-point samples, a third adjacent super-point sample whose purity value is higher than a preset purity threshold, and a fourth adjacent super-point sample whose purity value is not higher than the preset purity threshold; Dividing the fourth adjacent superpoint sample into a plurality of first sub-superpoint samples, and determining a second sub-superpoint sample having a purity value higher than a preset purity threshold from all the first sub-superpoint samples; A training data set is constructed based on all candidate superpoint samples, all third adjacent superpoint samples, and all second sub-superpoint samples.

[0099] In a possible implementation manner, the second processing module is further configured to: Solve the first normal vector of the current first adjacent superpoint sample and the second normal vector of the corresponding candidate superpoint sample; Calculate the cosine of the angle between the first normal vector and the second normal vector; Determine whether the cosine value of the angle is less than a preset normal vector similarity threshold; When the angle cosine value is less than the preset normal vector similarity threshold, the color similarity between the current first adjacent super point sample and the corresponding candidate super point sample is calculated; Determine whether the color similarity is less than a preset color similarity threshold; When the color similarity is less than a preset color similarity threshold, it is determined that the current first adjacent super point sample is similar to the corresponding candidate super point sample.

[0100] In a possible implementation, the training module is further used to: Performing update training on the first training model based on the training data set to obtain a second training model; The second training model is used as the first training model, and the first training model is continued to be used to predict each second super point sample to obtain the actual uncertainty of each second super point sample, all the second super point samples are screened based on the actual uncertainty to obtain multiple candidate super point samples, a training data set is constructed based on all the candidate super point samples, and the first training model is updated based on the training data set until the preset stopping condition is reached to obtain a point cloud semantic segmentation model.

[0101] It should be noted that the point cloud semantic segmentation system based on dynamic uncertainty active learning strategy provided by the present invention can execute the point cloud semantic segmentation method based on dynamic uncertainty active learning strategy of any of the above-mentioned embodiments during specific operation, which will not be elaborated in this embodiment.

[0102] Figure 8 is a schematic diagram of the structure of the electronic device provided by the present invention, such as Figure 8As shown, the electronic device may include: a processor 810 (processor), a communication interface 820 (CommunicationsInterface), a memory 830 (memory) and a communication bus 840 , wherein the processor 810 , the communication interface 820 , and the memory 830 communicate with each other through the communication bus 840 . The processor 810 can call the logic instructions in the memory 830 to execute a point cloud semantic segmentation method based on a dynamic uncertainty active learning strategy, the method comprising: inputting three-dimensional point cloud data into a point cloud semantic segmentation model to obtain a semantic category of each point in the three-dimensional point cloud data; wherein the point cloud semantic segmentation model is constructed in the following manner: based on three-dimensional point cloud data samples, a plurality of first super point samples and a plurality of second super point samples are obtained, the first super point samples are labeled super point samples, and the second super point samples are unlabeled super point samples; an initial model is trained based on all the first super point samples to obtain a first training model; each second super point sample is predicted using the first training model to obtain an actual uncertainty of each second super point sample; all the second super point samples are screened based on the actual uncertainty to obtain a plurality of candidate super point samples; a training data set is constructed based on all the candidate super point samples; and the first training model is updated and trained based on the training data set to obtain a point cloud semantic segmentation model.

[0103] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory 830 (ROM, Read-Only Memory), random access memory 830 (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0104] 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 the dynamic uncertainty active learning strategy provided by the above-mentioned embodiments, and the method includes: inputting three-dimensional point cloud data into a point cloud semantic segmentation model to obtain the semantic category of each point in the three-dimensional point cloud data; wherein the point cloud semantic segmentation model is constructed based on the following method: based on three-dimensional point cloud data samples, a plurality of first super point samples and a plurality of second super point samples are obtained, the first super point samples are labeled super point samples, and the second super point samples are unlabeled super point samples; based on all the first super point samples, an initial model is trained to obtain a first training model; each second super point sample is predicted using the first training model to obtain the actual uncertainty of each second super point sample; based on the actual uncertainty, all the second super point samples are screened to obtain a plurality of candidate super point samples; based on all the candidate super point samples, a training data set is constructed; based on the training data set, the first training model is updated and trained to obtain a point cloud semantic segmentation model.

[0105] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the processor 810 executes the point cloud semantic segmentation method based on the dynamic uncertainty active learning strategy provided in the above-mentioned embodiments, the method comprising: inputting three-dimensional point cloud data into a point cloud semantic segmentation model to obtain a semantic category of each point in the three-dimensional point cloud data; wherein the point cloud semantic segmentation model is constructed based on the following method: based on three-dimensional point cloud data samples, a plurality of first super-point samples and a plurality of second super-point samples are obtained, the first super-point samples are labeled super-point samples, and the second super-point samples are unlabeled super-point samples; an initial model is trained based on all the first super-point samples to obtain a first training model; each second super-point sample is predicted using the first training model to obtain the actual uncertainty of each second super-point sample; all the second super-point samples are screened based on the actual uncertainty to obtain a plurality of candidate super-point samples; a training data set is constructed based on all the candidate super-point samples; and the first training model is updated and trained based on the training data set to obtain a point cloud semantic segmentation model.

[0106] The system embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, i.e., they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0107] Through the description of the above implementation modes, those skilled in the art can clearly understand that each implementation mode can be implemented by means of software plus a necessary general hardware platform, or of course by hardware. Based on such an understanding, the above technical solution can essentially or in other words be embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0108] 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 embodiments of the present invention.

Claims

1. A point cloud semantic segmentation method based on dynamic uncertainty active learning strategy, characterized in that: include: Input the three-dimensional point cloud data into the point cloud semantic segmentation model to obtain the semantic category of each point in the three-dimensional point cloud data; wherein the point cloud semantic segmentation model is constructed based on the following method: Based on the three-dimensional point cloud data samples, a plurality of first superpoint samples and a plurality of second superpoint samples are obtained, wherein the first superpoint samples are labeled superpoint samples, and the second superpoint samples are unlabeled superpoint samples; Training an initial model based on all of the first superpoint samples to obtain a first training model; Use the first training model to predict each of the second super-point samples to obtain an actual uncertainty of each of the second super-point samples; Screening all the second super-point samples based on the actual uncertainty to obtain a plurality of candidate super-point samples; Based on all the candidate superpoint samples, construct a training data set; The first training model is updated and trained based on the training data set to obtain the point cloud semantic segmentation model.

2. The point cloud semantic segmentation method based on dynamic uncertainty active learning strategy according to claim 1 is characterized in that: The second super point samples include a plurality of three-dimensional points; and using the first training model to predict each of the second super point samples to obtain an actual uncertainty of each of the second super point samples specifically includes: Dividing all three-dimensional points included in each of the second super-point samples to obtain a first point set and a second point set, wherein the first point set is a set of three-dimensional points with the highest category proportion in the second super-point sample; Calculate the information entropy of each of the three-dimensional points using the first training model to obtain a first uncertainty of each three-dimensional point in the first point set and a second uncertainty of each three-dimensional point in the second point set; Calculate an initial uncertainty of each of the second super-point samples based on all of the first uncertainties and all of the second uncertainties; Multiply each of the initial uncertainties by the category balance weight coefficient of each of the second super-point samples to obtain the actual uncertainty of each of the second super-point samples, where the category balance weight coefficient of each of the second super-point samples is determined based on the proportion of the second super-point samples of each category in the super-point samples of all categories.

3. The point cloud semantic segmentation method based on dynamic uncertainty active learning strategy according to claim 1 is characterized in that: The screening of all the second super-point samples based on the actual uncertainty to obtain a plurality of candidate super-point samples specifically includes: In each round of screening, the second super-point sample with the highest actual uncertainty is selected as the candidate super-point sample, and based on the spatial distance and feature similarity between the candidate super-point sample and the remaining second super-point samples, the actual uncertainty of the remaining second super-point samples is dynamically attenuated until a preset number of candidate super-point samples are screened out.

4. The point cloud semantic segmentation method based on dynamic uncertainty active learning strategy according to claim 1, characterized in that: The step of constructing a training data set based on all the candidate superpoint samples specifically includes: Establish a multi-branch tree for preserving the neighbor relationship of super points; Based on the multi-branch tree, selecting a first adjacent super-point sample corresponding to each of the candidate super-point samples from all super-point samples; Determining whether each of the first adjacent super-point samples is similar to a corresponding candidate super-point sample; Determining second adjacent superpoint samples from all of the first adjacent superpoint samples, and calculating a purity value of each of the second adjacent superpoint samples, wherein the second adjacent superpoint samples are similar to the corresponding candidate superpoint samples; Determine, from all of the second adjacent super-point samples, a third adjacent super-point sample whose purity value is higher than a preset purity threshold, and a fourth adjacent super-point sample whose purity value is not higher than a preset purity threshold; Dividing the fourth adjacent superpoint sample into a plurality of first sub-superpoint samples, and determining a second sub-superpoint sample having a purity value higher than a preset purity threshold from all the first sub-superpoint samples; The training data set is constructed based on all the candidate superpoint samples, all the third adjacent superpoint samples and all the second sub-superpoint samples.

5. The point cloud semantic segmentation method based on dynamic uncertainty active learning strategy according to claim 4 is characterized in that: The determining whether each of the first adjacent super-point samples is similar to the corresponding candidate super-point sample specifically includes: Solve the first normal vector of the current first adjacent superpoint sample and the second normal vector of the corresponding candidate superpoint sample; Calculating the cosine value of the angle between the first normal vector and the second normal vector; Determine whether the cosine value of the angle is less than a preset normal vector similarity threshold; When the cosine value of the angle is less than a preset normal vector similarity threshold, calculating the color similarity between the current first adjacent super point sample and the corresponding candidate super point sample; Determining whether the color similarity is less than a preset color similarity threshold; When the color similarity is less than a preset color similarity threshold, it is determined that the current first adjacent super point sample is similar to the corresponding candidate super point sample.

6. The point cloud semantic segmentation method based on dynamic uncertainty active learning strategy according to claim 1, characterized in that: The updating and training of the first training model based on the training data set to obtain the point cloud semantic segmentation model specifically includes: Performing update training on the first training model based on the training data set to obtain a second training model; The second training model is used as the first training model, and the first training model is continued to be used to predict each of the second superpoint samples to obtain the actual uncertainty of each of the second superpoint samples, all of the second superpoint samples are screened based on the actual uncertainty to obtain multiple candidate superpoint samples, a training data set is constructed based on all of the candidate superpoint samples, and the first training model is updated based on the training data set until a preset stop condition is reached to obtain the point cloud semantic segmentation model.

7. A point cloud semantic segmentation system based on dynamic uncertainty active learning strategy, characterized in that: include: A first processing module, used for inputting the three-dimensional point cloud data into the point cloud semantic segmentation model to obtain the semantic category of each point in the three-dimensional point cloud data; The second processing module is used to obtain a plurality of first superpoint samples and a plurality of second superpoint samples based on the three-dimensional point cloud data samples, wherein the first superpoint samples are labeled superpoint samples and the second superpoint samples are unlabeled superpoint samples. A training module, used for training an initial model based on all the first superpoint samples to obtain a first training model; The second processing module is further used to use the first training model to predict each of the second super-point samples to obtain an actual uncertainty of each of the second super-point samples; The second processing module is further used to screen all the second super-point samples based on the actual uncertainty to obtain a plurality of candidate super-point samples; The second processing module is further used to construct a training data set based on all the candidate superpoint samples; The training module is also used to update the first training model based on the training data set to obtain the point cloud semantic segmentation model.

8. 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 the dynamic uncertainty active learning strategy as described in any one of claims 1 to 6 is implemented.

9. 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 point cloud semantic segmentation method based on a dynamic uncertainty active learning strategy as described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the point cloud semantic segmentation method based on a dynamic uncertainty active learning strategy as described in any one of claims 1 to 6 is implemented.

Images