Weak supervision laser radar point cloud semantic segmentation method based on cross-platform domain generalization

Abstract

The invention relates to the technical field of computer vision and image data processing, and provides a weak supervision laser radar point cloud semantic segmentation method based on cross-platform domain generalization in order to solve the problems of poor model generalization ability and high labeling cost caused by view angle and density differences of ground and aviation platform point cloud data in the prior art. Comprising the following steps: in a training stage, adopting weak supervised learning, simulating point cloud characteristics of a target domain through a progressive domain perception enhancement strategy, applying cross-scale semantic alignment constraint, and forcing feature representation of a model learning domain to be unchanged; and in a target domain application stage, directly inputting target domain point cloud data which is from a second platform and is free from any label into the cross-platform domain invariant three-dimensional semantic segmentation network, and outputting a semantic segmentation result. According to the method disclosed by the invention, the robust semantic segmentation on the unknown target domain can be realized without accessing the target domain data and labels thereof, so that the annotation dependence and the annotation cost are remarkably reduced, and the practical application of multi-platform three-dimensional data fusion is promoted.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image data processing technology, specifically relating to a weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization. Background Technology

[0002] 3D LiDAR point cloud semantic segmentation is a core task in fields such as autonomous driving and remote sensing mapping. It is a crucial task involving assigning semantic labels to each point in the point cloud to achieve comprehensive 3D scene understanding. Based on the degree of dependence of the training process on labeled data, existing methods are mainly divided into three categories: (1) Fully supervised methods rely on a large amount of precisely labeled point cloud data. The main technical paths are: projection-based methods, voxel-based methods and point-based methods. They can usually achieve high segmentation accuracy, but their performance is heavily dependent on expensive manual labeling.

[0003] (2) Unsupervised methods aim to completely eliminate the dependence on labeled data. The mainstream ideas include using cross-modal knowledge distillation, combining superpoint or supervoxel clustering to discover semantic structure in point clouds, and using visual base models to provide supervision signals. These methods can significantly reduce labeling costs, but their segmentation accuracy is still difficult to meet the needs of most practical applications.

[0004] (3) Weakly supervised methods train models under limited or alternative supervisory signals to achieve a balance between labeling costs and segmentation accuracy. Based on the form of supervisory information used, they can be categorized as follows: methods based on two-dimensional image labels, cross-modal methods, and methods based on limited three-dimensional labels. Weakly supervised methods are currently an active research direction, aiming to achieve practical segmentation performance at an acceptable cost.

[0005] With the development of sensor technology, high-quality ground point cloud datasets such as SemanticKITTI and SynLiDAR have emerged, driving progress in fully supervised segmentation algorithms. However, these datasets primarily originate from near-horizontal viewpoints of vehicle-mounted platforms, which differ fundamentally from the overhead viewpoints, point densities, and scene scales of aerial platforms (such as airborne laser scanners (ALS)). This leads to a significant performance degradation when models trained on ground data (source domain) are directly generalized to the aerial domain (target domain) for aerial applications (such as land cover classification and urban modeling). To address this issue, two main technological paradigms have emerged: Unsupervised Domain Adaptation (UDA) aims to improve model segmentation performance in the target domain by training with labeled source domain point clouds and unlabeled target domain point clouds, thereby reducing inter-domain differences. Its main technical approaches include converting point clouds into 2D representations to utilize image domain adaptation methods, designing domain-specific data augmentation, and using temporal information or generating pseudo-labels for self-training. However, UDA methods still require access to target domain data during training, limiting their generalization ability to the specific target domain.

[0006] Domain generalization (DG): As a superior paradigm, it pursues higher generalization ability. Its goal is to use only source domain data during the training phase, enabling the model to directly generalize to the unknown target domain. Mainstream methods include: a) Data augmentation and randomization: Simulating domain differences by simulating different sensor characteristics (such as point cloud density, noise) or environmental conditions; b) Domain-invariant representation learning: Designing network architectures or loss functions to learn features that are insensitive to domain changes; c) Utilizing multimodal or multi-view information: For example, combining bird's-eye view features or image prior knowledge to enhance the robustness of model understanding.

[0007] Both UDA and DG methods are largely based on the assumption of a fully supervised source domain, ignoring the extremely high cost of point cloud annotation. Under weak supervision, how to learn a model with strong generalization ability using a small amount of labeled data and a large amount of unlabeled data remains a challenge that has not yet been effectively solved.

[0008] Therefore, in response to the cross-platform domain generalization problem from autonomous driving to airborne point clouds, and the following shortcomings in existing technical methods: poor cross-platform generalization ability of models, excessive dependence on data annotation, and inability to effectively learn domain-invariant representations under weak supervision, there is an urgent need to provide a new solution that can overcome the above limitations. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention proposes a semantic segmentation method for weakly supervised LiDAR point clouds based on cross-platform domain generalization. The method achieves progressive enhancement and cross-scale alignment by designing a weakly supervised domain generalization framework, and utilizes a small number of labeled autonomous driving point clouds to train a model, enabling zero-label semantic segmentation of unknown airborne point clouds.

[0010] This invention provides a semantic segmentation method for weakly supervised lidar point clouds based on cross-platform domain generalization, the method comprising: A semantic segmentation method for weakly supervised lidar point clouds based on cross-platform domain generalization, the method comprising: Step 110: Obtain the ground lidar point cloud of the first platform as the original point cloud, wherein the original point cloud is sparsely labeled; Step 120: Perform progressive domain perception enhancement on the original point cloud to generate an enhanced point cloud that simulates the point cloud characteristics of an airborne lidar on a second platform; the second platform is an airborne platform different from the first platform. Step 130: Input the original point cloud and the corresponding enhanced point cloud into the 3D semantic segmentation network to extract multi-scale features; Step 140, perform cross-scale semantic alignment, including: enforcing alignment between the predicted distributions of the original point cloud and the augmented point cloud through semantic consistency constraints; and minimizing the semantic distribution differences between the original point cloud and the augmented point cloud in the feature space through contrastive regularization. Step 150: Combining the weakly supervised segmentation loss with labeled data in the original point cloud with the constraint of cross-scale semantic alignment, construct the overall loss function, train the three-dimensional semantic segmentation network, and learn a cross-platform domain-invariant three-dimensional semantic segmentation network. Step 160: Input the unlabeled target domain point cloud data from the second platform directly into the cross-platform domain-invariant 3D semantic segmentation network and output the semantic segmentation result.

[0011] On the other hand, the present invention also provides a weakly supervised lidar point cloud semantic segmentation device based on cross-platform domain generalization, comprising: The first module is used to acquire the ground lidar point cloud of the first platform as the original point cloud, wherein the original point cloud is sparsely labeled. The second module is used to progressively enhance the original point cloud with domain perception to generate an enhanced point cloud that simulates the point cloud characteristics of an airborne lidar on a second platform; the second platform is an airborne platform different from the first platform. The third module is used to input the original point cloud and the corresponding enhanced point cloud into the 3D semantic segmentation network to extract multi-scale features; The fourth module is used for cross-scale semantic alignment, including: enforcing alignment between the predicted distributions of the original point cloud and the augmented point cloud through semantic consistency constraints; and minimizing the semantic distribution differences between the original point cloud and the augmented point cloud in the feature space through contrastive regularization. The fifth module is used to combine the weakly supervised segmentation loss with labeled data in the original point cloud with the constraint of cross-scale semantic alignment to construct an overall loss function, train the three-dimensional semantic segmentation network, and learn a cross-platform domain-invariant three-dimensional semantic segmentation network. The sixth module is used to directly input unlabeled target domain point cloud data from the second platform into a cross-platform domain-invariant 3D semantic segmentation network and output semantic segmentation results.

[0012] Compared with the prior art, the beneficial effects of the present invention include: (1) Based on the domain generalization paradigm, the model can achieve robust semantic segmentation on an unknown airborne point cloud platform without any target domain data or labels during the training phase, thus realizing zero-label target domain deployment.

[0013] (2) Through the weakly supervised learning framework, only a small number of labels are needed to achieve performance comparable to the fully supervised method, which greatly reduces the burden of data labeling and significantly reduces the labeling cost.

[0014] (3) Through progressive domain-aware enhancement and cross-scale semantic alignment, the model learns domain-invariant features that are insensitive to platform differences. It exhibits stable performance when facing point clouds with different acquisition configurations and perspectives, thereby improving the model's generalization ability and robustness.

[0015] The technical solution proposed in this invention is a key step in realizing the collaborative perception and fusion of low-altitude remote sensing and ground mobile mapping data, providing a feasible technical path for macro-applications such as smart cities and resource surveys. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram comparing datasets captured from different perspectives in one embodiment of the present invention, wherein, Figure 1 (a) is the DALES dataset from the airborne dataset. Figure 1 (b) is from the SemanticKITTI dataset for autonomous driving; Figure 2 This is a flowchart of the steps of a weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization in one embodiment of the present invention; Figure 3 This is a diagram of the GTA architecture in one embodiment of the present invention; Figure 4 This is a schematic diagram of the cross-domain generalization results from SemanticKITTI to DALES in the experiment of this invention; Figure 5 This is a schematic diagram of the cross-domain generalization visualization results from SynLiDAR to DALES in the experiment of this invention; Figure 6This is a schematic diagram illustrating the visualization results of generalizing the SemanticKITTI training set to the DALES test set in the experiments of this invention. Figure 6 (a) is a visual diagram of the segmentation corresponding to vegetation. Figure 6 (b) is a visual diagram of the segmentation corresponding to the rod-shaped object. Figure 6 (c) is a segmented visualization diagram corresponding to the car. Figure 6 (d) is a visual diagram of the segmentation corresponding to the building. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Typically, unsupervised domain adaptation (UDA) methods rely on accessing target domain data during the training phase, which limits their generalization ability to the specific domain seen during training. In contrast, domain generalization (DG) aims to improve the robustness of models in unseen domains without requiring any target domain data. To achieve stable performance across heterogeneous platforms, this invention adopts a domain generalization-based research paradigm, aiming to learn domain-invariant point cloud representations from data collected from different sensor platforms. Existing domain generalization research has largely focused on adaptation between different datasets within autonomous driving scenarios (such as different weather conditions or cities), failing to adequately address the more challenging generalization scenario across heterogeneous acquisition platforms (such as from ground-based autonomous vehicles to airborne platforms).

[0020] Even for the same semantic category, point clouds captured from different platforms can exhibit significant differences in distribution and structural features. For example... Figure 1 As shown, a comparison of datasets captured from different perspectives is presented. The point cloud density from the airborne dataset DALES exhibits uniform characteristics, such as... Figure 1 As shown in (a), it typically preserves the upper-level geometry and overall infrastructure. In contrast, the point cloud density from the autonomous driving dataset SemanticKITTI decreases with increasing distance, as... Figure 1 As shown in (b), there is a tendency to emphasize the side-view structure while failing to adequately represent the upper-level geometry. It is evident that airborne LiDAR typically preserves the upper-level geometry and global infrastructure but lacks internal and lateral representations, while ground-based autonomous driving datasets emphasize side-view structural details. These inherent differences pose significant challenges to constructing truly domain-invariant representations.

[0021] To address this issue, this invention proposes a cross-platform, weakly supervised 3D point cloud semantic segmentation framework based on domain generalization. From a data-centric perspective, to reduce the model's dependence on target domain data, Progressive Domain-Aware Augmentation (PDA) is employed. This strategy applies random transformations to the source domain point cloud based on a basic augmentation strategy. This encourages the model to learn more generalizable representations across different samples, without relying on domain-specific details, thereby enhancing generalization capabilities even in the absence of target domain data. Furthermore, this invention designs platform-specific augmentation strategies to address the differences between vehicle-mounted and airborne LiDAR data, improving the model's adaptability to different platforms. From a label-centric perspective, weakly supervised methods provide limited annotations. To further enhance semantic refinement, the framework of this invention utilizes a large number of unlabeled points without interfering with the learning of features from labeled points. Specifically, this invention introduces Cross-Scale Semantic Alignment (CSA) to enforce semantic consistency between the original and augmented point clouds across multiple feature scales.

[0022] In one embodiment, such as Figure 2 As shown, this invention provides a semantic segmentation method for weakly supervised lidar point clouds based on cross-platform domain generalization, comprising: Step 110: Obtain the ground lidar point cloud of the first platform as the original point cloud, wherein the original point cloud is sparsely labeled; Step 120: Perform progressive domain perception enhancement on the original point cloud to generate an enhanced point cloud that simulates the point cloud characteristics of an airborne lidar on a second platform; the second platform is an airborne platform different from the first platform. Step 130: Input the original point cloud and the corresponding enhanced point cloud into the 3D semantic segmentation network to extract multi-scale features; Step 140, perform cross-scale semantic alignment, including: enforcing alignment between the predicted distributions of the original point cloud and the augmented point cloud through semantic consistency constraints; and minimizing the semantic distribution differences between the original point cloud and the augmented point cloud in the feature space through contrastive regularization. Step 150: Combine the weakly supervised segmentation loss with labeled data in the original point cloud with the cross-domain semantic alignment loss to construct an overall loss function, train the three-dimensional semantic segmentation network, and learn a cross-platform domain-invariant three-dimensional semantic segmentation network. Step 160: Input the unlabeled target domain point cloud data from the second platform directly into the cross-platform domain-invariant 3D semantic segmentation network and output the semantic segmentation result.

[0023] Existing unsupervised domain adaptation (UDA) and domain generalization (DG) methods focus only on adaptation across different autonomous driving datasets, neglecting the critical need for generalization across LiDAR equipment platforms in real-world scenarios. This invention addresses the cross-platform differences between autonomous driving and airborne point clouds, as well as the annotation challenges of existing datasets. It proposes a novel weakly supervised domain generalization framework (GTA) for LiDAR semantic segmentation, which utilizes a small number of autonomous driving point clouds to learn generalizable representations, thereby achieving robust segmentation of airborne point clouds under zero-annotation conditions.

[0024] To mitigate the challenges posed by domain offset, the proposed framework aims to enhance semantic segmentation of weakly supervised 3D point clouds. It integrates two complementary core functional components: Progressive Domain-Aware Augmentation (PDA) and Cross-Scale Semantic Alignment (CSA), as well as a core functional network: a 3D semantic segmentation network. These components work in conjunction with the network to improve robustness to distributional variations and enforce semantic consistency across different feature scales and domains.

[0025] The specific architecture (GTA framework) of the weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization is as follows: Figure 3 As shown, the original point cloud First, an enhanced point cloud is generated through progressive domain-aware enhancement processing. Then and Inputting into a 3D semantic segmentation network to extract multi-scale features, including extracting... and Corresponding multi-scale features and These features are used to generate initial semantic predictions through a series of multilayer perceptrons, which are then used to compute cross-scale semantic alignment loss. This loss integrates contrastive regularization and consistency supervision. It enforces semantic consistency across scales and domains. Simultaneously, only a subset of labeled points in the original point cloud contributes to the cross-entropy segmentation loss. This guides the network to learn accurate predictions from sparse annotations. The sparse annotations refer to point-level semantic annotations applied to a portion of the labeled dataset.

[0026] Specifically, in step 110, the ground-based lidar point cloud of the first platform is acquired as the original point cloud. The original point cloud It is sparsely labeled, meaning that only some points have point-level semantic annotations, which forms the basis of weakly supervised learning.

[0027] Furthermore, in step 120, the original point cloud is... A progressive domain-aware enhancement process is performed to generate an enhanced point cloud that simulates the point cloud characteristics of a second platform (aviation platform). This enhanced point cloud is denoted as... Specifically, this includes: Ground-based lidar point clouds and airborne lidar point clouds exhibit significant differences in feature distribution due to variations in acquisition perspective, sensor configuration, and scanning trajectory. Traditional data augmentation techniques cannot effectively simulate naturally occurring structured geometric changes across domains. Therefore, this invention proposes a progressive domain-aware augmentation (PDA) strategy. The PDA builds upon traditional geometric perturbations, gradually increasing the complexity and diversity of the perturbations to generate multiple semantically consistent point cloud variants. This progressive scheduling enables the 3D semantic segmentation network to smoothly transition from learning local perturbation invariance to global domain offset robustness, thereby effectively narrowing the cross-domain gap.

[0028] The PDA employs a phased enhancement strategy to construct a continuously segmented enhancement formula, which is used to generate an enhanced point cloud that simulates the point cloud characteristics of the second platform. .

[0029] The training process comprises two phases: a basic enhancement phase and a domain-aware enhancement phase. In the early stages of training, lightweight geometric perturbations, such as mirroring, random rotation, and jitter, are applied to improve robustness to local spatial noise while maintaining the overall topology. As training progresses, the PDA gradually activates domain-aware enhancements, simulating features common in terrestrial lidar, including uniform density distribution, vertical viewing angle bias, and geometric discontinuities caused by occlusion.

[0030] Inspired by the observation that ground-based lidar point clouds become sparser with increasing distance, while aerial point clouds typically maintain a uniform density, this invention introduces a distance-aware density compensation strategy.

[0031] For ground-based lidar point cloud assemblies Each point in the original point cloud The ground projection distance and relative compensation weight are calculated as follows: ; ; in, Point exist The Euclidean distance from the projection point on the plane to the origin. ; Represents the original point cloud The total number of midpoints This represents the relative compensation weight applied to enhance density in distant regions. Furthermore, this invention defines a vertical structural saliency term based on local height differences: ; in, It is a point The height coordinates; It is a point The neighborhood points are clustered together. express Neighborhood point cluster The number of points in the middle; the vertical structure salience term reflects the points. The height difference between the points and their neighbors is used to capture the geometric discontinuities on the roofs and facades of ground-level buildings, thereby reinforcing the occlusion patterns present from an aerial perspective. To avoid semantic distortion, the ground surface is preserved while the facade structure is sparsely retained.

[0032] The phased augmentation strategy ultimately unifies density compensation, structural saliency, and training scheduling into a single continuous segmented augmentation formula: ; in, Indicated by replication factor Copy point To achieve local density enhancement, This indicates rounding down to the nearest integer. Here, It is a monotonically increasing teaching coefficient. Increase the domain offset intensity during the later stages of training. The contributions of density and significance are controlled separately. Here, It is a facade salience mask. This indicates an indicator function that introduces sparse discontinuities in structurally significant regions.

[0033] This unified formula enables joint geometric modeling of domain differences: Simulates uniform density over long distances. Capture vertical structural deviations. Introducing facade sparsity. Exponential scheduling term. Ensure a smooth transition from mild to strong domain perturbations to prevent early training instability.

[0034] In step 130, the original point cloud Its corresponding augmented point cloud The inputs are fed into a 3D semantic segmentation network to extract their corresponding multi-scale features. and .

[0035] The 3D semantic segmentation network employs the existing RandLA-Net network to extract multi-scale features. and Specifically, the local feature aggregation module in the RandLA-Net network is used to perform local spatial encoding and attention aggregation on the point cloud (original point cloud and enhanced point cloud), and a random downsampling strategy is used to generate multi-level point cloud features to form multi-scale features.

[0036] The RandLA-Net network is a neural network specifically designed for semantic segmentation of large-scale 3D point clouds. Its core objective is to efficiently and effectively process massive point cloud data (such as hundreds of thousands or even millions of points) while maintaining good segmentation accuracy. This is achieved through random sampling (downsampling / upsampling) and a powerful local feature aggregation module to compensate for information loss that may be caused by random sampling. When downsampling is used, fast feature extraction efficiency can be obtained.

[0037] Obtained multi-scale features Corresponding to network inputs with consistent structure but diverse geometry after PDA processing, spatial robustness in cross-domain scenarios is significantly improved. By progressively exposing the model to increasingly challenging geometric variations, PDA enhances generalization ability while maintaining semantic stability, laying a solid foundation for subsequent cross-domain feature alignment.

[0038] While PDAs effectively increase cross-domain geometric diversity, they inevitably introduce feature mismatch between the original and augmented point clouds. Under weak supervision, point-level sparse annotations are limited, and local geometric perturbations easily cause semantic drift and accumulate in the feature space, ultimately reducing class discriminability. Furthermore, features at different scales exhibit varying sensitivities to local structure and global semantics, which may further exacerbate potential alignment biases.

[0039] To address these challenges, this invention proposes a cross-scale semantic alignment (CSA) functional component that enforces semantic consistency and contrastive regularization across multiple feature scales to achieve robust alignment under diverse geometric perturbations. Specifically, lower-level features capture fine-grained local geometry, while higher-level features focus on global semantic stability. Furthermore, contrastive regularization explicitly brings semantically consistent samples closer together and pushes away potentially misaligned instances, effectively mitigating the distribution bias introduced by PDA.

[0040] Through CSA, the network learns semantically invariant representations in geometrically diverse domains, thereby significantly improving the generalization and discriminative capabilities of the feature space and providing robust support for weakly supervised scenarios lacking dense annotations.

[0041] Step 140, cross-scale semantic alignment, includes: Step 141: By using semantic consistency constraints, the alignment between the predicted distributions of the original point cloud and the enhanced point cloud is enforced.

[0042] Directly applying supervision to the augmented point cloud can lead to overfitting perturbations in the 3D semantic segmentation network, resulting in biases and decreased segmentation performance. To address this issue, this invention introduces semantic consistency constraints, enforcing alignment between the predicted distributions of the original point cloud and the augmented point cloud. Specifically, it mandates that the semantic predictions of the original point cloud for the same scene and the augmented point cloud of its simulated aviation platform maintain consistency, thereby improving the network's robustness to cross-platform structural changes.

[0043] Cross-cloud consistency loss based on Jensen-Shannon divergence, given the original prediction and enhanced forecasting The softmax probability distribution is defined as: ; in, It is the total number of points in the original point cloud. The Jensen-Shannon divergence is expressed as follows: ; ; in, The Kullback-Leibler divergence is defined by the following formula: ; in, Indicates original prediction / enhanced prediction. It is the summation and average of the original forecast and the enhanced forecast; yes The Middle The point belongs to the th point The predicted probability of each semantic category. semantic category , It is the total number of points in the original point cloud. It represents the total number of semantic categories; yes The Middle The point belongs to the th point The predicted probability of each semantic category.

[0044] Given that 3D semantic segmentation networks are highly sensitive to noise disturbances in the early stages of training, this invention employs a parameterized multi-stage scheduling strategy to gradually activate the consistency objective constraint:

[0045] in, Indicates the current period. It is the maximum number of cycles. and It is an adjustable hyperparameter. The above consistency target constraint strategy is consistent with the aforementioned progressive domain awareness enhancement process. In the early stage, this invention allocates a weaker consistency strength, which is gradually increased as training stabilizes, effectively suppressing the distribution drift introduced by radical geometric perturbations.

[0046] Using the aforementioned consistency objective constraint, the final semantic consistency loss is defined as follows: ; Step 142: By using contrastive regularization, the semantic distribution difference between the original point cloud and the enhanced point cloud in the feature space is minimized.

[0047] Relying solely on consistency constraints in the prediction space can lead to the collapse of feature representations at structurally different points over time, resulting in inter-class confusion under aggressive geometric perturbations. To enhance inter-class separability in the embedding space, this invention further introduces point-level contrastive regularization loss.

[0048] Given the original embedding features and enhanced embedding features This invention applies both along the feature dimension. Normalization. For the feature embedding of each anchor point. The corresponding positive sample is the enhanced counterpart. All other embeddings are considered negative samples. This is based on the temperature parameter. The point-level contrastive regularization loss is defined by the following formula: ; in, Represents cosine similarity. It's a temperature over-parameter. It is the number of points after sampling.

[0049] This regularization term encourages the feature embeddings of anchors to be brought closer to their corresponding positive samples, while suppressing similarity with negative samples in the embedding space, thus complementing the semantic consistency constraints imposed in the prediction space.

[0050] Based on semantic consistency loss and point-level contrastive regularization loss, a cross-domain semantic alignment loss is jointly defined: ; By simultaneously optimizing output space consistency and embedding space discriminability, the 3D semantic segmentation network effectively mitigates feature collapse, enhances structural robustness, and generates domain-invariant 3D semantic representations.

[0051] In step 150, the loss function for training the 3D semantic segmentation network is constructed, including: Based on the above design, this invention proposes GTA, a generalizable LiDAR semantic segmentation framework for weakly supervised domains. This framework jointly optimizes the weakly supervised segmentation loss and semantic alignment regularization introduced by PDA and CSA, thereby improving classification accuracy and cross-domain robustness. The overall loss function is defined as: ; in, Controlling the contribution of semantic alignment regularization terms, The weakly supervised segmentation loss is calculated by applying the cross-entropy between the predicted probability and the weak label, and is given by the following formula: ; in, Indicates a weak supervision signal. Indicates network prediction, This represents a set of weakly labeled points.

[0052] During training, the 3D semantic segmentation network further optimizes semantic consistency in the prediction space and semantic alignment in the embedding space, effectively mitigating distribution differences caused by domain offset. Therefore, the 3D semantic segmentation network achieves enhanced generalization ability and more stable semantic discrimination ability in different scenarios.

[0053] The cross-platform domain-invariant 3D semantic segmentation network obtained through the aforementioned steps has powerful cross-domain generalization capabilities.

[0054] Furthermore, in step 160, the target domain point cloud data without any annotations from the second platform is directly input into the cross-platform domain-invariant 3D semantic segmentation network, and the semantic segmentation result is output.

[0055] The trained 3D semantic segmentation network is applied to real-world cross-platform scenario reasoning. Specifically, target domain LiDAR point cloud data from a second platform (aviation platform), completely unlabeled at the point or scene level, is directly input into the trained network. The network automatically extracts multi-scale depth features of the point cloud via forward propagation and, based on learned domain-invariant semantic knowledge, directly outputs the semantic category label for each 3D point, thus completing fully automatic semantic segmentation of point clouds from unknown platforms. This process requires no model fine-tuning or parameter adjustments for the target platform, fully demonstrating the practicality and generalization ability of the proposed method in real-world cross-platform scenarios.

[0056] In one embodiment, the first platform is an autonomous driving platform.

[0057] In summary, the weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization provided in the above embodiments of the present invention first utilizes a progressive perceptual enhancement strategy to apply staged domain-aware transformations to the input original point cloud, including basic enhancements—mirroring, random rotation, and jitter—as well as perceptual enhancements such as distance-guided upsampling and geometric perceptual selection. Unlike traditional random enhancement strategies that may distort the inherent spatial structure, PDA dynamically adjusts the enhancement intensity according to the training stage. In the early stages of training, the model focuses on learning basic geometric patterns under slight perturbations; as training progresses, it is gradually exposed to more complex geometric changes and density shifts. Distance-guided upsampling and geometric perceptual selection operations increase point density while preserving top points, facade points, and parts of the original point cloud, thereby enriching training diversity without compromising structural integrity. This progressive strategy helps learn feature representations robust to domain changes while maintaining structural and semantic consistency.

[0058] Next, multi-scale perceptrons (MLPs) are applied to generate initial semantic predictions using multi-scale feature extraction and semantic alignment. These predictions are further refined by a cross-scale semantic alignment (CSA) module, which enforces semantic consistency across multiple spatial scales. Specifically, CSA constrains both coarse-grained and fine-grained semantic responses to maintain local coherence while preserving high-frequency structural details. This strategy effectively mitigates semantic drift caused by significant geometric perturbations introduced by PDA, enabling the network to utilize complementary contextual cues and ensuring reliable cross-scale semantic alignment.

[0059] Finally, the initial predictions at weakly labeled points are used to compute the segmentation loss, which is jointly optimized with the cross-scale alignment loss. In summary, these objectives constitute a unified reinforcement-alignment learning paradigm that progressively improves domain generalization and semantic stability under weak supervision.

[0060] This invention also systematically evaluates and analyzes the proposed weakly supervised lidar point cloud semantic segmentation method and strategy based on cross-platform domain generalization through experiments. Experimental results on the SemanticKITTI, SynLiDAR, and DALES datasets show that the proposed method consistently outperforms all baseline methods, even when trained under weak supervision, thus validating the effectiveness of the proposed framework.

[0061] To evaluate the semantic segmentation performance of the model in the Domain Generalization (DG) setting, this invention restricts the training process to a single labeled autonomous driving dataset as the source domain, and then deploys the trained model on previously unseen airborne LiDAR data for target domain testing. Using the DALES dataset, this invention constructs two DG benchmark sets to rigorously evaluate how the proposed GTA framework facilitates cross-platform generalization. Furthermore, this invention compares its method with several state-of-the-art point cloud segmentation generalization networks to provide a comprehensive performance analysis.

[0062] Specifically, this invention conducted extensive experiments on three point cloud datasets acquired using different lidar configurations.

[0063] The first benchmark is SemanticKITTI → DALES. SemanticKITTI was acquired by a mobile LiDAR platform equipped with a Velodyne HDL-64E sensor, with each frame in its training sequence densely annotated at the point level. In contrast, DALES contains airborne LiDAR data captured by a Piper PA-31 Panther Navajo aircraft equipped with a Riegl Q1560 dual-channel system. This benchmark provides a robust testbed for evaluating a model's domain generalization ability when migrating from autopilot LiDAR to airborne LiDAR data.

[0064] The second benchmark is SynLiDAR→DALES. SynLiDAR includes virtual scenes created by professional 3D artists using the Unreal Engine 4 platform, providing diverse synthetic LiDAR scans. This benchmark is able to evaluate whether a model can learn platform-invariant representations from artificially generated point clouds. Because its domain offset comes from both synthetic-to-realistic differences and cross-platform variations, this setting presents a more challenging and comprehensive evaluation scenario.

[0065] This invention selects seven shared evaluation categories for two domain generalization benchmarks based on the semantic classification of the DALES dataset. Since SemanticKITTI and SynLiDAR contain more fine-grained category definitions, several categories are merged accordingly, and all dataset annotations are remapped to a unified label space. For SemanticKITTI and SynLiDAR, this invention employs point-level weak supervision by randomly selecting a fixed proportion of points as annotation samples. All experiments were conducted on a workstation equipped with an Intel Core™ i9-12900KF CPU and an NVIDIA RTX 3090 Ti GPU (24GB RAM).

[0066] (1) Design evaluation indicators To quantitatively evaluate and analyze the model's performance, this invention performs the evaluation at all points on the target domain test set DALES. This invention uses the Intersection over Union (IoU) ratio for each segmentation category and the average IoU ratio (mIoU) for all categories as evaluation metrics. Let... Indicates the total number of categories. Represents semantic categories The actual number of cases, Represents semantic categories The number of false positives, Represents semantic categories The number of false counterexamples. Semantic category. The two corresponding evaluation indicators are calculated as follows: ; ; (2) Comparison with the SOTA method To realistically evaluate the model performance of this invention, several closely related methods were selected as baselines. First, this invention considers data augmentation-based methods, including Pointcutmix—which generates augmented point clouds by blending two point clouds selected from the training set—and Mix3D, which constructs mixed scenes by concatenating points and labels from different scenes. Second, this invention includes domain adaptation methods such as RayCast and fully supervised domain generalization methods for all-weather scenes, such as PointDR, which employs a domain randomization strategy during segmentation training. Finally, this invention also evaluates the generalization performance of SQN, a weakly supervised method designed specifically for large-scale 3D point cloud semantic segmentation. It is worth noting that the implementations of the baseline methods are based on relevant code from the LiDOC project.

[0067] SemanticKITTI→DALES Assessment: SemanticKITTI is a densely annotated autonomous driving dataset collected by a mobile laser scanning (MLS) platform. It contains 45.49 million points and 28 semantic categories, which are further merged into 7 categories for the purposes of this invention.

[0068] This invention trains the model on the SemanticKITTI dataset and evaluates its generalization performance on the DALES dataset. The results are summarized in Table 1. Under fully supervised training, the data augmentation method Mix3D achieves the best generalization performance, reaching 22.61% mIoU. The domain generalization method PointDR for all-weather scenes performs poorly, reaching only 14.71% mIoU, and exhibits the lowest performance in the ground category. The weakly supervised method SQN achieves 20.54% mIoU, surpassing Pointcutmix, RayCast, and PointDR.

[0069] Table 1: Performance Comparison of SemanticKITTI→DALES

[0070] Qualitative results from the migration from SemanticKITTI to DALES, annotated with only 0.1%, are as follows: Figure 4 As shown. The first column displays the original input point cloud. Since the DALES dataset does not provide RGB information, scalar field visualization is used for visualization. The second column presents the ground truth annotations, and the third column displays the predictions generated by the method of this invention. The last column displays the predictions of SQN, which serves as the backbone network of the framework of this invention. Regions where the method of this invention obtains better predictions are marked with circles. It can be observed that the method of this invention has significant improvements over SQN on the DALES dataset, especially in more accurately identifying buildings and cars, thus demonstrating the effectiveness of the proposed method.

[0071] SynLiDAR→DALES Assessment: SynLiDAR is a large-scale synthetic dataset collected and annotated from diverse virtual environments with rich scene content and layouts. It contains over 19 billion points across 32 semantic categories, which are further consolidated into 7 test categories for evaluation. Given the relatively small number of points per frame, this invention employs a 1% weak annotation setting to ensure the model can effectively learn from a sufficient number of points.

[0072] This invention trains the model only on the SynLiDAR dataset and evaluates its target domain performance on the DALES dataset. The results are summarized in Table 2. Under the fully supervised training mechanism, the data augmentation method Pointcutmix achieves the best generalization performance, reaching 22.19% mIoU, surpassing Mix3D, RayCast, and PointDR. The weakly supervised 3D point cloud segmentation method SQN performs poorly, achieving only 14.34% mIoU, and also struggles in distinguishing... vegetation The category aspect is particularly difficult.

[0073] The proposed method, GTA, achieves the highest segmentation performance, reaching 24.02% mIoU using only 1% of the training data. This represents a 1.83% and 8.61% improvement over the data augmentation methods Pointcutmix and Mix3D, respectively, and a 7.51% improvement over the domain adaptation method RayCast. Furthermore, GTA achieves a 9.68% mIoU improvement compared to SQN, which also uses 1% of labeled data. These results demonstrate GTA's ability to learn more robust representations for cross-platform tasks and generalization from synthetic data to real-world data. Notably, the performance gap of SQN between synthetic and autonomous driving datasets highlights the inherent difficulty of generalizing from synthetic to real-world data, while GTA maintains strong performance, confirming its effectiveness.

[0074] Qualitative results from migrating from SynLiDAR to DALES using only 1% annotation are as follows: Figure 5 As shown. The first column shows the original input point cloud, visualized using scalar field visualization techniques on the DALES dataset. The second column depicts the ground truth annotations, and the third column shows the predictions generated by the method of this invention. The last column shows the predictions of SQN

[63] , which serves as the backbone network of the framework of this invention. Regions where the method of this invention obtains better predictions are marked with circles. It can be observed that the method of this invention has significant improvements over SQN on the DALES dataset, especially in more accurately identifying buildings and vegetation, demonstrating the effectiveness of the proposed method.

[0075] Table 2: Performance Comparison of SynLiDAR→DALES

[0076] (3) Visualization of segmentation results and ablation experiment Due to the distinct nature of different object categories within the same scene, the generalization strategies applied during model training may not be equally effective for each category. However, a well-designed network should exhibit a degree of universality and robustness. Based on the observation that GTA achieves optimal generalization overall, this invention further visualizes its performance when generalizing from the SemanticKITTI dataset to the DALES dataset using only 0.1% of the labeled points, and analyzes the inter-class differences in the model's generalization ability. Figure 6 This is a visualization of the generalization results from the SemanticKITTI training set to the DALES test set. Each set of results includes the original point cloud, the ground truth annotations within the rectangular bounding box, the predictions from the baseline method, and the results from the proposed method. Using only 0.1% of the labeled points, the baseline method shows a significant difference from the ground truth annotations, while the segmentation results produced by the proposed GTA method are closer to the ground truth annotations. Figure 6The visualization results show that, with the same amount of labeled data, the GTA of this invention outperforms the baseline method SQN in visualization across several categories, including vegetation, poles, vehicles, and buildings. Figure 6 (a) Figure 6 (b) Figure 6 (c) and Figure 6 (d) Segmentation visualizations corresponding to vegetation, stalks, vehicles, and buildings, respectively. Each visualization includes the original point cloud, segmentation results within rectangular bounding boxes, and the results from ground truth annotations, baseline method predictions, and the proposed method. It can be observed that GTA achieves superior generalization in recognizing category-specific features. For example, in recognizing buildings, the model considers point cloud differences caused by viewpoints, thus learning more robust features compared to SQN, demonstrating the effectiveness of the proposed method.

[0077] To verify the effectiveness of each component in the framework of this invention, an ablation experiment (removal of functional modules) was conducted in the SemanticKITTI to DALES scenario, under weak supervision settings using only 0.1% of the labeled data. The overall results are summarized in Table 3.

[0078] Table 3: SemanticKITTI→DALES Ablation Study

[0079] Where Δ represents the change in performance (mIoU) relative to the removal of the corresponding module, in units of %.

[0080] PDA introduces data perturbations in a phased manner, gradually aligning the geometric and semantic features of the source domain samples with the target domain distribution, effectively mitigating the distribution offset between domains. Furthermore, the local structural changes introduced by phased enhancement significantly expand the diversity of training samples, enabling the model to learn more discriminative structured semantic features and suppressing overfitting tendencies under weak label constraints. After removing PDA, the model's mIoU performance decreased by 7.76%, indicating that the progressive enhancement strategy plays an irreplaceable role in cross-domain scenarios. This result is consistent with existing theoretical analyses of domain generalization and phased adaptation mechanisms, further validating the positive role of PDA in robust feature extraction.

[0081] The CSA mechanism establishes consistency constraints across feature embedding spaces at different scales, ensuring coordination between local and global semantic distributions and thus stabilizing the point cloud semantic structure. Under extremely sparse annotation conditions, CSA effectively compensates for semantic uncertainty caused by the lack of supervision signals and enhances the separability of different categories, enabling the model to maintain clearer semantic boundaries under cross-domain conditions. When CSA is removed, model performance decreases by 8.82%, highlighting the decisive role of cross-scale alignment in weakly supervised semantic structure learning.

[0082] The complete GTA model achieves optimal performance when both PDA and CSA are incorporated, indicating a complementary relationship between the two components at different spatial levels. PDA primarily operates at the level of input sample distribution, enhancing feature diversity by expanding the geometric-semantic sample space of the source domain. CSA, on the other hand, operates on the embedding representation space, improving the stability of the decision boundary through enhanced semantic aggregation. Together, they form a cross-domain robust closed loop from the data space to the representation space, enabling the model to maintain consistent semantic discriminative ability under severe domain shifts. In contrast, retaining only one component leads to a significant performance degradation, suggesting that the sample diversity introduced by PDA and the semantic structure formed by CSA mutually reinforce each other: PDA provides a more challenging training distribution, while CSA ensures stable class separability under such perturbations.

[0083] Ablation experiments clearly demonstrate the crucial role of PDA in enhancing cross-domain sample adaptability, and the significant contribution of CSA in constructing discriminative semantic embedding structures. When used together, these components enable the model to maintain stable and generalizable semantic prediction performance under the dual challenges of extremely low annotation density and significant domain shifts. This highlights the potential engineering value and deployment prospects of GTA in real-world collaborative perception tasks for autonomous systems.

[0084] In summary, this invention designs a weakly supervised domain generalization framework for LiDAR semantic segmentation, bridging the gap between autonomous driving and airborne point clouds, and achieving zero-label deployment in the target domain. By integrating progressive domain augmentation and semantic alignment constraints, GTA effectively mitigates cross-domain differences while fully utilizing limited labeled data and a large amount of unlabeled data. To adapt to the distribution differences between heterogeneous LiDAR platforms, this invention proposes a distance-guided dynamic upsampling strategy and a cross-view augmentation scheme, supplemented by a progressive training process to ensure stable optimization. Furthermore, cross-domain feature consistency and contrastive regularization enhance robustness to perturbations and improve the discriminativeness of the learned representation. Extensive experiments demonstrate that GTA achieves state-of-the-art performance among weakly supervised methods for cross-platform 3D semantic segmentation, significantly reducing labeling costs and enabling reliable deployment on airborne point clouds. Specific beneficial effects include: This invention proposes the first cross-platform weakly supervised 3D point cloud semantic segmentation framework based on domain-invariant (DG), which can effectively learn domain-invariant features from multi-platform data.

[0085] This invention designs a progressive data augmentation strategy to simulate airborne lidar data to enhance cross-platform generalization capabilities.

[0086] This invention introduces cross-scale semantic alignment to make full use of unlabeled points, achieve richer contextual feature learning, and further improve generalizable 3D semantic scene understanding.

[0087] In one embodiment, the present invention also provides a semantic segmentation device for weakly supervised lidar point clouds based on cross-platform domain generalization, comprising: On the other hand, the present invention also provides a weakly supervised lidar point cloud semantic segmentation device based on cross-platform domain generalization, comprising: The first module is used to acquire the ground lidar point cloud of the first platform as the original point cloud, wherein the original point cloud is sparsely labeled. The second module is used to progressively enhance the original point cloud with domain perception to generate an enhanced point cloud that simulates the point cloud characteristics of an airborne lidar on a second platform; the second platform is an airborne platform different from the first platform. The third module is used to input the original point cloud and the corresponding enhanced point cloud into the 3D semantic segmentation network to extract multi-scale features; The fourth module is used for cross-scale semantic alignment, including: enforcing alignment between the predicted distributions of the original point cloud and the augmented point cloud through semantic consistency constraints; and minimizing the semantic distribution differences between the original point cloud and the augmented point cloud in the feature space through contrastive regularization. The fifth module is used to combine the weakly supervised segmentation loss with labeled data in the original point cloud with the constraint of cross-scale semantic alignment to construct an overall loss function, train the three-dimensional semantic segmentation network, and learn a cross-platform domain-invariant three-dimensional semantic segmentation network. The sixth module is used to directly input unlabeled target domain point cloud data from the second platform into a cross-platform domain-invariant 3D semantic segmentation network and output semantic segmentation results.

[0088] On the other hand, the present invention provides a computer device including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization provided in any of the above embodiments. The computer device may be a server. The computer device includes a processor, a memory, a network interface, and a database connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device stores sample data. The network interface of the computer device is used for communication with external terminals via a network connection.

[0089] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization provided in any of the above embodiments.

[0090] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0091] Matters not covered in this invention are common knowledge.

[0092] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0093] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

[0094] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A semantic segmentation method for weakly supervised lidar point clouds based on cross-platform domain generalization, characterized in that, The method includes: Step 110: Obtain the ground lidar point cloud of the first platform as the original point cloud, wherein the original point cloud is sparsely labeled; Step 120: Perform progressive domain perception enhancement on the original point cloud to generate an enhanced point cloud that simulates the point cloud characteristics of an airborne lidar on a second platform; the second platform is an airborne platform different from the first platform. Step 130: Input the original point cloud and the corresponding enhanced point cloud into the 3D semantic segmentation network to extract multi-scale features; Step 140, perform cross-scale semantic alignment, including: enforcing alignment between the predicted distributions of the original point cloud and the augmented point cloud through semantic consistency constraints; and minimizing the semantic distribution differences between the original point cloud and the augmented point cloud in the feature space through contrastive regularization. Step 150: Combining the weakly supervised segmentation loss with labeled data in the original point cloud with the constraint of cross-scale semantic alignment, construct the overall loss function, train the three-dimensional semantic segmentation network, and learn a cross-platform domain-invariant three-dimensional semantic segmentation network. Step 160: Input the unlabeled target domain point cloud data from the second platform directly into the cross-platform domain-invariant 3D semantic segmentation network and output the semantic segmentation result.

2. The weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization according to claim 1, characterized in that, In step 120, the progressive domain-aware enhancement employs a staged enhancement strategy to construct a continuously segmented enhancement formula, which is used to generate an enhanced point cloud that simulates the point cloud characteristics of the second platform. ; The phased enhancement strategy includes at least the following two phases: Basic Augmentation Phase: In the early stages of training, basic geometric augmentations, including mirroring, random rotation, and jitter, are applied. Domain Awareness Enhancement Phase: As training progresses, domain awareness enhancement based on the characteristics of the simulated aviation platform is gradually introduced, including distance-guided upsampling and geometry-aware selection.

3. The weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization according to claim 2, characterized in that, The continuous piecewise enhancement formula is given by the following equation: ； in, Indicated by replication factor Copy point To achieve local density enhancement, Indicates rounding down; It is a monotonically increasing teaching coefficient. Used to increase the domain offset strength in the later stages of training. Control the contributions of density and significance separately; The relative compensation weights applied to enhance density in distant regions are given by the following formula: ； In the above formula, It is a point Ground projection distance, Point exist The Euclidean distance from the projection point on the plane to the origin. ; Represents the original point cloud Total number of midpoint clouds; It is a facade salience mask. This indicates an indicator function used to introduce sparse discontinuities in structurally significant regions; The vertical structure saliency term is given by the following formula: ； in, It is a point The height coordinates; It is a point The neighborhood points are clustered together. express Neighborhood point cluster The number of points in the middle.

4. The weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization according to claim 1, characterized in that, In step 130, the three-dimensional semantic segmentation network adopts the RandLA-Net network; The local feature aggregation module in the RandLA-Net network is used to perform local spatial encoding and attention aggregation on the original point cloud and the enhanced point cloud. A random downsampling strategy is then used to generate multi-level point cloud features to obtain the original point cloud. With augmented point cloud Corresponding multi-scale features and .

5. The weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization according to claim 1, characterized in that, In step 140, the semantic consistency constraint is achieved by calculating the cross-cloud consistency loss based on Jensen-Shannon divergence between the predicted distributions of the original point cloud and the enhanced point cloud, and is scheduled using dynamic weight coefficients related to the training progress.

6. The weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization according to claim 5, characterized in that, The cross-cloud consistency loss based on Jensen-Shannon divergence is defined by the following formula: ； in, It is the total number of points in the original point cloud. Indicates the Jensen-Shannon divergence; It is a pre-given original prediction. It is a pre-given enhanced prediction; The scheduling strategy for the dynamic weight coefficients is as follows: ； in, It is a consistency objective constraint for dynamic weight coefficients. Indicates the current period. It is the maximum number of cycles. and It is an adjustable hyperparameter.

7. The weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization according to claim 6, characterized in that, In step 140, the comparison regularization includes: The feature embeddings of the original point cloud and the enhanced point cloud are used as input. The multi-scale features in the original point cloud are used as anchors, the multi-scale features of the enhanced point cloud are used as positive samples, and all point cloud features other than the enhanced point cloud in the batch are used as negative samples. A point-level contrast regularization loss is constructed to bring positive sample pairs closer together and push negative sample pairs further apart.

8. The weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization according to claim 7, characterized in that, In step 150, the cross-domain semantic alignment loss is given by the following formula: ； in, It is a semantic consistency loss. The point-level contrastive regularization loss is given by the following formula: ； In the above formula, Represents cosine similarity. It's a temperature over-parameter. It is the number of points after sampling; It is the feature embedding of the anchor point. , It is a positive sample.

9. The weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization according to claim 8, characterized in that, The total loss function is given by the following equation: ； in, The weakly supervised segmentation loss is calculated by applying the cross-entropy between the predicted probability and the weak label, and is given by the following formula: ； In the above formula, Indicates a weak supervision signal. Indicates network prediction, This represents a set of weakly labeled points.

10. The weakly supervised lidar point cloud semantic segmentation method based on cross-platform domain generalization according to claim 1, characterized in that, The first platform is an autonomous driving platform.

Images