Bridge type fast classification system and method for autonomous path planning of unmanned aerial vehicles

Abstract

This invention provides a rapid bridge type classification system and method for autonomous path planning of unmanned aerial vehicles (UAVs). The system includes a data acquisition module for acquiring three-dimensional spatial point cloud data of bridges; a preprocessing module for cleaning and standardizing the raw point cloud data; a bridge classification module for obtaining bridge classification results based on the preprocessed raw point cloud data and a bridge type classification model; wherein the bridge type classification model determines an attention score for dynamic anti-occlusion importance sampling based on a symmetry-aware enhanced semantic importance scoring mechanism; and a path planning module for determining the planned inspection path for the UAV based on the bridge classification results. This invention overcomes the problems of poor generalization and weak anti-occlusion in traditional bridge type classification techniques by embedding symmetry-aware feature enhancement and a dynamic anti-occlusion importance sampling mechanism, enabling real-time inspection of UAVs in complex environments.

Description

Technical Field

[0001] This invention relates to the field of bridge classification technology, and in particular to a rapid bridge type classification system and method for autonomous path planning of unmanned aerial vehicles (UAVs). Background Technology

[0002] With the increasing prominence of aging transportation infrastructure, regular health monitoring of large structures such as bridges has become crucial. In recent years, drones equipped with sensors such as LiDAR have become the mainstream method for bridge data collection due to their flexibility, efficiency, and safety. A core prerequisite for achieving automated drone inspection is that the system can automatically identify the type of bridge to be inspected (such as beam bridges, arch bridges, cable-stayed bridges, suspension bridges, etc.). This is because different types of bridges have significantly different core load-bearing components, vulnerable parts, and optimal inspection paths. Only after accurately identifying the bridge type can the system automatically call upon or generate a refined inspection plan that matches it, thereby achieving true "adaptive" and "intelligent" operation.

[0003] Current technologies primarily rely on lidar or other sensors mounted on drones to acquire bridge point cloud data, and then use algorithms to identify the bridge type. However, this method has extremely high requirements for data quality. If occlusion, noise, or uneven point cloud density occurs, key geometric features cannot be effectively extracted, leading to classification failure. Summary of the Invention

[0004] This invention provides a bridge type rapid classification system and method for UAV autonomous path planning, which solves the problem that existing bridge type recognition methods fail to effectively extract key geometric features when there is occlusion, noise, or uneven point cloud density, leading to classification failure.

[0005] A first aspect of this invention provides a bridge-type rapid classification system for autonomous path planning of unmanned aerial vehicles, comprising:

[0006] The data acquisition module is used to acquire the three-dimensional spatial point cloud data of the bridge.

[0007] The preprocessing module is used to clean and standardize the raw point cloud data;

[0008] The bridge classification module is used to obtain bridge classification results based on the preprocessed raw point cloud data and the bridge type classification model. The bridge type classification model determines the attention score of dynamic anti-occlusion importance sampling based on the semantic importance scoring mechanism based on symmetry perception enhancement.

[0009] The path planning module is used to determine the planned inspection path for the drone based on the bridge classification results.

[0010] In one possible implementation, the bridge classification model includes a feature extraction module, a symmetry-aware feature enhancement module, a semantic importance scoring module, a dynamic anti-occlusion importance sampling module, and a bridge classification module; the bridge classification module is used for:

[0011] The preprocessed raw point cloud data is input into the feature extraction module to obtain the raw feature matrix;

[0012] The original feature matrix is ​​input into the symmetry-aware feature enhancement module to obtain the enhanced semantic feature matrix;

[0013] The enhanced semantic feature matrix is ​​input into the semantic importance scoring module to obtain the semantic importance score;

[0014] The semantic importance score and the local geometric features based on the preprocessed raw point cloud computing are input into the dynamic anti-occlusion importance sampling module to obtain the filtered key feature points;

[0015] The selected key feature points are input into the bridge type classification module to obtain the bridge classification results.

[0016] In one possible implementation, the bridge classification module is used for:

[0017] Principal component analysis is performed on the original feature matrix to calculate the covariance matrix of the point cloud segment. The first principal component vector corresponding to the largest eigenvalue in the covariance matrix is ​​extracted, and the direction pointed to by the first principal component vector is identified as the main extension direction of the bridge, which serves as the potential main axis of symmetry of the bridge.

[0018] For each point feature in the original feature matrix, using the potential principal axis of symmetry as a reference, the mirror position of the point feature on the other side of the axis of symmetry is calculated by the mirror mapping algorithm.

[0019] If there is a directly corresponding point cloud data at the mirror position, then the features of the points at that mirror position are extracted as symmetrical region features.

[0020] If there is no direct point cloud data at the mirror location, then select a set of neighboring points within a preset range around the mirror location, and calculate the symmetrical region features of the point based on the mean of the neighborhood features of the neighboring point set.

[0021] The features of each point in the original feature matrix are concatenated with the features of its corresponding symmetrical region to obtain the enhanced semantic feature matrix.

[0022] In one possible implementation, the bridge classification module is used for:

[0023] Map the feature vector of each point in the enhanced semantic feature matrix to an unnormalized initial attention score;

[0024] LeakyReLU is used as a non-linear activation function to process the initial attention score;

[0025] The initial attention scores are normalized using the Softmax function to obtain the final semantic importance score for each point. The semantic importance score of each point represents its contribution to the bridge classification task.

[0026] In one possible implementation, the bridge classification module is used for:

[0027] For each point in the preprocessed original point cloud, its local neighborhood is constructed, and the local geometric features of the local neighborhood are calculated; where the local geometric features include local point density, normal consistency coefficient, and local curvature estimation.

[0028] The calculated local geometric features are normalized and then weighted and fused into a reliability score.

[0029] Finally, the importance score is fused by weighting the semantic importance score and the reliability score to obtain the attention score, which is used as the final sampling weight for each point.

[0030] The final sampling weights are converted into a probability distribution using the Softmax function, and random sampling is performed based on this probability distribution to select key feature points.

[0031] In one possible implementation, the feature extraction module is configured with a PointNet++ network; the bridge classification module is used for:

[0032] Calculate the low-dimensional features of each point in the original point cloud data; the low-dimensional features include normal vector, curvature, and point density.

[0033] The preprocessed raw point cloud data is concatenated with the low-dimensional features to obtain the original feature matrix.

[0034] In one possible implementation, the bridge classification module is used for:

[0035] The selected key feature points are grouped within their Euclidean space neighborhood, and local features are extracted.

[0036] Hierarchical feature aggregation is performed based on the grouping results and local features to obtain multi-scale local features;

[0037] Global max pooling is performed on the multi-scale local features to obtain a global feature vector describing the entire bridge, which is then fed into a fully connected layer and a Softmax classifier to obtain the bridge classification result.

[0038] In one possible implementation, the bridge classification module is used for:

[0039] The raw point cloud data is denoised, ground points are removed, and downsampling is performed to obtain preliminarily cleaned point cloud data.

[0040] The cleaned point cloud data is labeled and augmented to complete the data preprocessing.

[0041] A second aspect of the present invention provides a method for rapid bridge type classification for autonomous path planning of unmanned aerial vehicles, comprising:

[0042] Obtain the three-dimensional point cloud data of the bridge;

[0043] Clean and standardize the raw point cloud data;

[0044] Based on the preprocessed raw point cloud data and the bridge type classification model, the bridge classification results are obtained; wherein, the bridge type classification model determines the attention score of dynamic anti-occlusion importance sampling based on a semantic importance scoring mechanism based on symmetry perception enhancement.

[0045] Based on the bridge classification results, the planned inspection route for drones is determined.

[0046] In one possible implementation, the bridge classification model includes a feature extraction module, a symmetry-aware feature enhancement module, a semantic importance scoring module, a dynamic anti-occlusion importance sampling module, and a bridge classification module; based on the preprocessed raw point cloud data and the bridge classification model, the bridge classification result is obtained, including:

[0047] The preprocessed raw point cloud data is input into the feature extraction module to obtain the raw feature matrix;

[0048] The original feature matrix is ​​input into the symmetry-aware feature enhancement module to obtain the enhanced semantic feature matrix;

[0049] The enhanced semantic feature matrix is ​​input into the semantic importance scoring module to obtain a semantic importance score;

[0050] The semantic importance score and the local geometric features based on the preprocessed raw point cloud computing are input into the dynamic anti-occlusion importance sampling module to obtain the filtered key feature points;

[0051] The filtered key feature points are input into the bridge type classification module to obtain the bridge classification results.

[0052] Compared to traditional technologies, this invention provides a rapid bridge type classification system and method for autonomous path planning of unmanned aerial vehicles (UAVs). The system includes a data acquisition module for acquiring three-dimensional spatial point cloud data of bridges; a preprocessing module for cleaning and standardizing the raw point cloud data; a bridge classification module for obtaining bridge classification results based on the preprocessed raw point cloud data and a bridge type classification model; wherein the bridge type classification model determines an attention score for dynamic anti-occlusion importance sampling based on a semantic importance scoring mechanism enhanced by symmetry perception; and a path planning module for determining the planned inspection path for the UAV based on the bridge classification results. This invention overcomes the problems of poor generalization and weak anti-occlusion in traditional bridge type classification technologies by embedding symmetry perception feature enhancement and a dynamic anti-occlusion importance sampling mechanism, enabling real-time inspection of UAVs in complex environments.

[0053] Specifically, the present invention has the following technical effects:

[0054] 1. Improved training efficiency and generalization ability:

[0055] This patent employs Batch Normalization (BN) after each linear layer of the MLP in the feature extraction network. BN forces the input distribution of each layer to stabilize, eliminating the need for the network to constantly adjust parameters to adapt to drastic changes in data distribution. This significantly accelerates training convergence and saves valuable parameter tuning time and computational costs. Furthermore, the random noise introduced by BN when calculating batch statistics can slightly suppress overfitting, thereby improving the model's performance on novel, unseen bridges. This is crucial for engineering applications requiring high stability.

[0056] 2. A leap in data utilization efficiency and decoupling from labeled data:

[0057] This invention introduces a symmetry enhancement mechanism, essentially providing the model with a powerful structured inductive bias. This allows the model's learning process to move beyond blindly memorizing massive amounts of data variations and instead directly grasp the geometric essence of the bridge structure. This injection of prior knowledge significantly improves data utilization efficiency and model convergence speed, specifically by significantly reducing the number of training iterations required to achieve the same classification accuracy. Quantitative tests show that, using only 50% of the training data, the model with symmetry enhancement achieves 5% higher accuracy than the baseline model using 100% of the data. This means that in bridge inspection scenarios where data annotation is extremely costly, this invention greatly reduces the reliance on expensive labeled data.

[0058] 3. Enhanced model training stability:

[0059] This invention employs LeakyReLU as the activation function for attention scores in the semantic importance scoring module. This ensures that even when a linear transformation produces a negative input, the neuron still receives a small gradient, effectively avoiding the "neuron death" problem and enhancing the stability of the entire training process. This means that even if some features are initially judged as unimportant (negative values) during training, they retain the opportunity to be re-evaluated and activated in subsequent training processes, thereby comprehensively improving the model's learning ability.

[0060] 4. Fundamental improvements to the sampling mechanism and intelligent focusing:

[0061] This patent introduces a reliability score and probability sampling strategy, fundamentally improving the traditional point cloud sampling mechanism. By embedding a dynamic anti-occlusion importance sampling module, the model is endowed with an expert-like "attention" mechanism, enabling it to actively focus on key structural regions with high discriminative power and reliable data, rather than sampling uniformly without discrimination.

[0062] 5. System-level anti-shading effect linked by "symmetry and reliability":

[0063] The core advantage of this invention lies in the synergistic operation of the "symmetry enhancement" and "reliability assessment" modules, resulting in system-level intelligence that cannot be achieved by a single technology. The following test scenarios and data chains clearly demonstrate this:

[0064] Test scenario: Simulates a single cable tower being completely obscured (structural occlusion).

[0065] Data observation: On the occluded side, due to the lack of point cloud data, the reliability score of points is generally lower than 0.3; however, due to the symmetry enhancement mechanism, the model can still infer the importance of the region, so the semantic importance score of points in the occluded side of the pylon region remains higher than 0.7.

[0066] Fusion Results: Under the fusion strategy, the final sampling weights are significantly suppressed by the reliability score, and the sampler automatically avoids unreliable occluded regions. Meanwhile, on the intact side, the Sota region possesses both high semantic importance (>0.8) and high reliability (>0.9), and is therefore selected by the sampler with a high probability.

[0067] Quantitative Results and Conclusions: Under this extreme occlusion, the model of this invention still maintains an accuracy of 85%, while the accuracy of the baseline model (without symmetry enhancement) plummets to <40%. This data strongly demonstrates that the linkage mechanism of this invention can intelligently guide sampling attention and make correct judgments even under severe occlusion, achieving excellent robustness. Attached Figure Description

[0068] Figure 1This is a structural schematic diagram of a bridge-type rapid classification system for autonomous path planning of unmanned aerial vehicles provided in an embodiment of the present invention. Detailed Implementation

[0069] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0070] Figure 1 This is a schematic diagram of the bridge-type rapid classification system for autonomous path planning of unmanned aerial vehicles provided in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0071] Data acquisition module 11 is used to acquire three-dimensional spatial point cloud data of the bridge;

[0072] Preprocessing module 12 is used to clean and standardize the raw point cloud data;

[0073] The bridge classification module 13 is used to obtain the bridge classification result based on the preprocessed raw point cloud data and the bridge type classification model; wherein, the bridge type classification model determines the attention score of dynamic anti-occlusion importance sampling based on the semantic importance scoring mechanism based on symmetry perception enhancement.

[0074] The path planning module 14 is used to determine the planned inspection path of the UAV based on the bridge classification results.

[0075] In this embodiment of the invention, the data acquisition module 11 acquires three-dimensional spatial point cloud data of the bridge, providing basic data support for subsequent bridge type classification. The preprocessing module 12 cleans and standardizes the acquired raw point cloud data, specifically including operations such as noise reduction, ground point removal, and downsampling, to reduce the amount of data and improve data quality. The bridge classification module 13 derives the bridge classification result based on the preprocessed raw point cloud data and the bridge type classification model; the bridge type classification model determines the attention score required for dynamic anti-occlusion importance sampling through a semantic importance scoring mechanism based on symmetry perception enhancement. First, features are extracted from the preprocessed point cloud data to obtain the original feature matrix, then the feature semantic information is enhanced by the symmetry perception feature enhancement module, and then the semantic importance scoring module outputs the semantic importance score, which serves as the key basis for dynamic anti-occlusion importance sampling. After selecting key feature points, the bridge type classification is completed. The function of the path planning module 14 is to plan an appropriate inspection path for the UAV based on the bridge classification results obtained by the bridge classification module 13, combined with the structural characteristics and inspection requirements of different bridge types, to ensure that the inspection covers key parts and is adapted to the inspection difficulty of different bridge types.

[0076] The purpose of this invention is to provide a rapid bridge type classification system and method for autonomous path planning of unmanned aerial vehicles (UAVs). This system first efficiently processes large-scale bridge point cloud data, then accurately and quickly identifies and classifies bridge types. Finally, based on the classification results, it automatically calls a preset inspection parameter library to plan precise inspection paths for the UAV.

[0077] The data acquisition module is responsible for acquiring raw 3D spatial data. It consists of a LiDAR, a Global Navigation Satellite System (GNSS) / Real-Time Kinematics (RTK), and an Inertial Measurement Unit (IMU), and outputs raw point cloud data with spatial coordinates and intensity information.

[0078] The preprocessing module cleans and standardizes the raw point cloud data, and prepares the dataset for subsequent algorithms. It includes a point cloud preprocessing submodule and a data annotation and management submodule.

[0079] In addition, the core algorithm model library supports and improves deep learning algorithms to achieve fast and accurate bridge-type classification. This includes submodules for feature extraction, feature importance prediction, adaptive sampling, and classifiers. The model training and evaluation platform provides offline training, tuning, and evaluation of the core algorithm models.

[0080] The path rule knowledge base stores optimal inspection path templates for different bridge types. It consists of an expert experience base and a path template management system. The expert experience base pre-defines standardized, refined inspection paths defined by industry experts for various bridge types, including beam bridges, arch bridges, cable-stayed bridges, and suspension bridges. The path template management system includes key parameters for each path template, such as the drone's flight path, sensor angles, and shooting positions.

[0081] In some embodiments, the bridge classification model includes a feature extraction module, a symmetry-aware feature enhancement module, a semantic importance scoring module, a dynamic anti-occlusion importance sampling module, and a bridge classification module. The bridge classification module is used to: input preprocessed raw point cloud data into the feature extraction module to obtain an original feature matrix; input the original feature matrix into the symmetry-aware feature enhancement module to obtain an enhanced semantic feature matrix; input the enhanced semantic feature matrix into the semantic importance scoring module to obtain a semantic importance score; input the semantic importance score and local geometric features based on the preprocessed raw point cloud data into the dynamic anti-occlusion importance sampling module to obtain filtered key feature points; and input the filtered key feature points into the bridge classification module to obtain the bridge classification result.

[0082] In this embodiment of the invention, the preprocessed raw point cloud data, which has undergone denoising, ground point removal, and downsampling, is first input into the feature extraction module. This module first supplements each point in the point cloud with low-dimensional handcrafted features (including the normal vector describing the orientation of the surface where the point is located, the curvature describing the curvature of the surface, and the point density describing the sparsity of the point cloud region), and concatenates them with the three-dimensional coordinate features of the point to form input features. Then, the input features are encoded through a 3-layer structure (each layer contains 64 neurons) of SharedMLP, and each MLP layer is followed by batch normalization (BN) and ReLU activation functions to stabilize the feature distribution and enhance nonlinear expression. Finally, the original feature matrix containing the bridge's geometric structure and preliminary semantic information is output.

[0083] Next, the original feature matrix is ​​input into the symmetry-aware feature enhancement module. This module first performs principal component analysis (PCA) on the point cloud fragments corresponding to the original feature matrix, extracts the first principal component vector corresponding to the largest eigenvalue in the covariance matrix, and determines it as the main extension direction of the bridge (usually the bridge span direction) and uses it as the potential principal axis of symmetry. Then, for each point feature in the original feature matrix, based on the potential principal axis of symmetry, the mirror position on the other side of the axis of symmetry is calculated using a mirror mapping algorithm. If there is a directly corresponding point cloud data for the mirror position, the feature of that point is extracted as the symmetry region feature. If there is no direct point cloud data, the neighboring point set within a preset range around the mirror position is selected, and the symmetry region feature is calculated based on the mean of the neighborhood features. Finally, the original point features and their corresponding symmetry region features are concatenated, and projection compression fusion is used to ensure that the feature dimension is consistent with the original feature dimension, generating an enhanced semantic feature matrix.

[0084] Next, the enhanced semantic feature matrix is ​​input into the semantic importance scoring module. This module uses a mapping function with learnable parameters (attention vector a, weight matrix W, bias vector b) to map the feature vector of each point in the enhanced semantic feature matrix to an unnormalized initial attention score, and then applies the LeakyReLU activation function (satisfying... f ( x )=max( αx , x ),in α The initial attention score is nonlinearly processed to a value of 0.01 to avoid neuron "death" and learn complex nonlinear relationships. Finally, the processed initial attention score is normalized by the Softmax function to obtain the semantic importance score of each point's contribution to the bridge classification task.

[0085] Then, the semantic importance score and the local geometric features (including local point density, normal consistency coefficient, and local curvature estimation) based on the preprocessed original point cloud are input into the dynamic anti-occlusion importance sampling module: First, a spherical neighborhood with a radius of 5-10 times the average point spacing of the point cloud is constructed with each point as the center. The local point density (the ratio of the number of points in the neighborhood to the volume of the spherical neighborhood), normal consistency coefficient (the mean of the cosine of the angle between the normal of each point in the neighborhood and the average normal), and local curvature estimation (calculated by the eigenvalue of the PCA covariance matrix of the neighborhood point set) are calculated, and the three types of features are normalized to the [0,1] interval. Then, the reliability score is obtained by weighted summation according to the weight coefficients (local point density 0.4, normal consistency coefficient 0.4, normalized local curvature estimation 0.2). Subsequently, the semantic importance score and the reliability score are weighted and fused to obtain the final sampling weight (i.e., attention score), and converted into a probability distribution through the Softmax function. During the training phase, random sampling is performed according to the probability distribution, and during the inference phase, the K points with the highest scores are selected to screen out key feature points containing high-quality features.

[0086] Finally, the selected key feature points are input into the bridge classification module. This module first groups the key feature points into Euclidean space neighborhoods and extracts local fine features through the PointNet module. Then, it repeatedly performs the "sampling-grouping-feature extraction" operation through multiple SetAbstraction (SA) layers to gradually expand the receptive field and fuse multi-scale local features. Finally, it performs global max pooling on the fused features to obtain a global feature vector describing the overall bridge. This vector is then input into a fully connected layer and combined with a Softmax classifier to output the classification probability distribution of various bridge types such as beam bridges, arch bridges, cable-stayed bridges, and suspension bridges. The category with the highest probability is taken as the final bridge classification result.

[0087] In some embodiments, the bridge classification module is configured to: perform principal component analysis on the original feature matrix, calculate the covariance matrix of the point cloud segment, extract the first principal component vector corresponding to the largest eigenvalue in the covariance matrix, identify the direction pointed to by the first principal component vector as the main extension direction of the bridge, and use it as the potential principal axis of symmetry of the bridge; for each point feature in the original feature matrix, calculate the mirror position of the point feature on the other side of the axis of symmetry using a mirror mapping algorithm based on the potential principal axis of symmetry; if there is a directly corresponding point cloud data for the mirror position, extract the feature of the point at the mirror position as the symmetric region feature; if there is no direct point cloud data for the mirror position, select a set of neighboring points within a preset range around the mirror position, and calculate the symmetric region feature of the point feature based on the mean of the neighborhood features of the neighboring point set; and concatenate the feature of each point in the original feature matrix with its corresponding symmetric region feature to obtain an enhanced semantic feature matrix.

[0088] In this embodiment of the invention, the core design of the embedded feature importance sampling module is to replace the original farthest point sampling (FPS) with importance sampling in the Set Abstraction (SA) layer of PointNet++; at the same time, a lightweight feature encoding network (such as a small PointNet or a few layers of MLP) is added before the first SA layer. This feature encoding network predicts an importance score Si for each point in the point cloud. This importance score Si can clearly indicate the contribution of the corresponding point to the final bridge classification task, providing a key basis for subsequent accurate sampling.

[0089] The Shared MLP component in this module is designed as follows: First, input feature optimization. Since using only the 3D coordinates (x, y, z) of bridge points as input would limit the information and slow down the recognition of different bridge types such as suspension bridges, cable-stayed bridges, and beam bridges, a series of low-dimensional handcrafted features are calculated and concatenated for each point before inputting the data into the Shared MLP. This significantly reduces the learning difficulty of the model. These low-dimensional handcrafted features include the normal vector (nx, ny, nz) describing the orientation of the surface where the point is located, the curvature describing the degree of curvature of the surface, and the point density describing the sparsity of the point cloud in that area. For example, in practical applications, the input dimension can be transformed from (100000, 3) to (100000, 6) (i.e., combining coordinates and normal vectors). This input feature design allows the model to capture the geometric information of the bridge more quickly, thus significantly improving the accuracy of subsequent importance sampling. This optimization is crucial for distinguishing between key scenarios such as flat bridge decks and complex anchorage areas.

[0090] In balancing the depth and width of Shared MLP, a 3-layer network structure with 64 neurons in each layer was ultimately adopted to ensure feature extraction capabilities while avoiding computational inefficiency caused by excessive network complexity.

[0091] In terms of learning strategies, an optimized design of Batch Normalization (BN) and ReLU is adopted after each MLP layer. This choice stems from the data characteristics in actual inspection scenarios: on the one hand, the size, scale, and orientation of the bridges being inspected vary significantly, and different drone flight missions also lead to different scanning angles, distances, and resolutions for each scan. Furthermore, environmental factors such as tree and vehicle occlusion, as well as changes in lighting and weather, cause the point cloud data distribution to fluctuate dramatically. On the other hand, the process of collecting bridge data is difficult and the annotation cost is extremely high, directly resulting in a limited training dataset. Based on these characteristics, introducing BN after each linear layer of the MLP forces the input distribution of each layer to stabilize, eliminating the need for frequent parameter adjustments to adapt to drastic changes in data distribution. This not only significantly accelerates the model's training convergence speed but also saves valuable parameter tuning time and computational costs. Simultaneously, BN introduces random noise when calculating batch statistics. This random noise can slightly suppress overfitting, effectively improving the model's performance on unfamiliar bridges, which is a core requirement in engineering applications. Given the infinite diversity of testing environments, the bridges the model will need to deal with in the future may differ from the training data in terms of size ratio, scanning angle and distance, surrounding environment (such as occlusion and weather), point cloud density and quality. Therefore, the model must be able to ignore these "irrelevant noises" and firmly grasp the essential characteristics of the "bridge type". The batch-by-batch variation noise introduced by Batch Normalization can precisely add "difficulty" and "uncertainty" to the model's learning process, preventing the model from fitting those overly subtle and specific patterns in the training set. It forces the model to learn features that are less sensitive to noise and more robust. These features are the real essential characteristics that distinguish "beam bridges" from "cable-stayed bridges", rather than a specific scanning angle of a specific point cloud. The core function of Batch Normalization (BN) is to stabilize the data within a distribution range suitable for nonlinear transformation. The data processed by BN is then input into the ReLU activation function. After processing through multiple layers of "BN->ReLU" modules, the semantic feature matrix H is finally obtained. Placing BN before ReLU ensures that ReLU always receives input with a stable distribution (approximately 0 mean, 1 variance), thus guaranteeing the effective functioning of the activation function.

[0092] Symmetry-aware feature enhancement is based on the structural prior of bridge symmetry, used to further enhance the feature representation of each point: First, principal component analysis (PCA) is performed on the current point cloud segment to identify the main extension direction of the bridge (usually the bridge span direction) as the potential axis of symmetry. Specifically, the first principal component vector corresponding to the largest eigenvalue is determined as the direction of the bridge's main axis of symmetry. Next, for each point feature in the semantic feature matrix H, its corresponding mirror symmetric point is found according to the determined axis of symmetry (if there is no direct point cloud data for the mirror position, a neighboring region is selected), and the features of the symmetric region are obtained. Finally, the original point features and its symmetric point features are fused by concatenation to generate the enhanced features. The technical effect of this design is very significant. Even if the tower on one side of the bridge is occluded, the enhanced features will still contain information about the intact tower on the other side, so that the points in that region still have high discriminative features, thereby significantly improving the robustness of the model in occlusion scenarios. For example, in the anti-occlusion test with an occlusion rate >50%, the accuracy of the baseline model will plummet to below 60%, while the model of this solution can still maintain an accuracy of over 85%. It should be clarified that this is a feature-endogenous enhancement mechanism oriented towards the symmetry of bridge structures. This mechanism is only effective for macroscopic structures with strong symmetry, such as bridges, and is not applicable to asymmetrical or small objects.

[0093] Finally, through the projection compression fusion stage, it is ensured that the feature dimensions after the above feature fusion are consistent with the original feature dimensions (e.g., 64 dimensions). Therefore, the subsequent network structure does not need to be adjusted, effectively ensuring the continuity and adaptability of the entire model process.

[0094] In some embodiments, the bridge classification module is configured to: map the feature vector of each point in the enhanced semantic feature matrix to an unnormalized initial attention score; process the initial attention score using LeakyReLU as a non-linear activation function; normalize the processed initial attention score according to the Softmax function to obtain the final semantic importance score of each point, wherein the semantic importance score of each point represents the contribution of that point to the bridge classification task.

[0095] In this embodiment of the invention, the semantic importance score is calculated using an attention mechanism. In this process, the calculation of the semantic importance factor strongly depends on the aforementioned "symmetry enhancement feature". If this feature is lacking, the importance score of the occluded key component will drop sharply, thereby affecting the accuracy of subsequent sampling and classification.

[0096] The input to this step is a semantic feature matrix H, which is processed by a network consisting of Shared MLP, BN, and ReLU, and then output after symmetry-aware enhancement and compression fusion. The shape of H is [N, D'], where N represents the number of points in the point cloud and D' represents the feature dimension (e.g., 64 dimensions). This matrix fully contains the deep semantic information of each point on the bridge, providing a high-quality data foundation for subsequent score calculation.

[0097] Initial attention score ( e i The calculation method is as follows: through a learnable function g( ), and the D'-dimensional feature vector of each point h i Mapped to a scalar fraction e i This scalar score is the raw, unnormalized score. To ensure the model can learn complex nonlinear relationships, LeakyReLU is used as the nonlinear activation function, which is expressed as:

[0098] f(x) = max(αx, x)

[0099] in, It is a very small constant (usually with a value of 0.01), and the specific calculation formula is as follows:

[0100]

[0101] In the formula, Represents the attention vector. Represents the weight matrix. This represents the bias vector. Before training begins, and The parameters are randomly initialized, typically by sampling initial values ​​from a normal distribution with a mean of 0 and a small standard deviation. During training, the model calculates the gradient of the loss function with respect to the attention vector using the backpropagation algorithm, and then updates the model based on these gradients using an optimizer (such as Adam). and The attention vector will eventually converge to a set of stable values. This set of stable values ​​enables the model to distinguish different bridge types most accurately and allows the model to learn to allocate high weights to the most discriminative feature combinations.

[0102] The adoption of LeakyReLU in this scheme is significantly necessary for three main reasons: First, the bridge point cloud data collected by UAVs is complex and diverse. Some points may be deemed "useless" (i.e., the linear transformation output is negative) in the early stages of training. Standard ReLU would directly set these outputs to zero, causing the gradient to also become zero, potentially leading to the "death" of related neurons that cannot be recovered. LeakyReLU, however, allows a very small gradient (e.g., α=0.01) to flow towards negative inputs, ensuring that all neurons have a chance to be updated in the early stages of training, greatly enhancing the stability and robustness of the training. Second, the feature patterns of some key parts of the bridge (such as the cable anchorage area and the initiation point of minor cracks) are quite subtle. LeakyReLU's small leakage in the negative region acts like a highly sensitive listener; even if the signal is weak and temporarily judged as "unimportant," it can retain a noise floor, preventing it from being detected in subsequent Softmax operations. In the global competition, being set to zero completely loses its voice, providing the possibility for the model to discover more refined features; thirdly, the point cloud acquired by LiDAR itself is sparse, with a large number of points located in information-poor areas (such as flat bridge surfaces and sky backgrounds). LeakyReLU's sparsity control is more gentle than ReLU, and it can achieve a better balance between suppressing noise and preserving potentially effective information, making it very suitable for point cloud data processing scenarios.

[0103] Due to the original score obtained in the previous step e i Since the value range is uncertain and cannot be directly used for subsequent judgments, it needs to be normalized using Softmax to transform it into a final semantic importance score in the form of a probability distribution. S i The specific calculation formula is as follows:

[0104] = exp( ) /

[0105] This calculation makes the semantic importance scores of each point more comparable, clearly reflecting the contribution of each point to the bridge classification task.

[0106] In some embodiments, the bridge classification module is configured to: construct a local neighborhood for each point in the preprocessed original point cloud, and calculate the local geometric features of the local neighborhood; wherein the local geometric features include local point density, normal consistency coefficient, and local curvature estimation; normalize the calculated local geometric features and fuse them into a reliability score through weighted summation; finally fuse the importance scores, weighted summation of the semantic importance score and the reliability score to obtain an attention score, which serves as the final sampling weight for each point; convert the final sampling weight into a probability distribution through a Softmax function, and randomly sample according to the probability distribution to select key feature points.

[0107] In this embodiment of the invention, for each point in the point cloud A spherical neighborhood with radius R is constructed with the point cloud as the center. The radius R is set according to the average density of the point cloud, and is usually 5-10 times the average spacing between points in the point cloud.

[0108] Local geometric feature extraction is achieved through the following formula:

[0109]

[0110] in, For local curvature estimation, For local point density, The normal uniformity coefficient is... This indicates the number of points in the neighborhood. Higher density indicates a more complete scan of the area and higher reliability.

[0111]

[0112] Among them, the average normal value of all points in the neighborhood. The cosine of the angle between the normal of each neighboring point and the average normal. The more consistent the direction of the normals ( The closer to 1, the more continuous and intact the surface of the region is, and the higher the reliability. M For point The number of points within the local neighborhood;

[0113]

[0114] Principal component analysis (PCA) is performed on the neighborhood point set to calculate the eigenvalues ​​of the covariance matrix. , The smaller the value, the flatter the area, and the more reliable the data is usually.

[0115] , , The three features are normalized to a uniform [0,1].

[0116] The three normalized features are weighted and fused to obtain the final reliability score:

[0117]

[0118] Wherein, the weight coefficients satisfy + + =1. Based on experience, it can be set to... , , It emphasizes the importance of point density and normal consistency.

[0119] This reliability score can effectively identify:

[0120] High reliability areas: regions with dense point clouds, continuous surfaces, and geometrically flat surfaces (such as complete bridge decks or towers).

[0121] Low reliability areas: regions with sparse point clouds, disordered normals, and complex geometry (such as occlusion boundaries, the junction of trees and bridges).

[0122] By combining it with semantic importance scores, the sampler is guided to actively avoid unreliable regions, significantly improving the classification robustness in occluded environments.

[0123] The semantic importance score and the reliability score are weighted and fused to obtain the final sampling weight for each point.

[0124] The weighted fusion refers to a technical solution that fuses semantic importance scores and reliability scores based on preset weight coefficients, including but not limited to the following two implementation methods:

[0125] The product-weighted fusion expression for Implementation Method 1 (preferred, adapted for occlusion-resistant scenarios) is:

[0126] S final =( w s × S i )×( w r × S reliability )

[0127] in, w s The weighting coefficients for semantic importance scores. w r This is the weighting coefficient for the reliability score, with values ​​ranging from [0.8, 1.2] (the default value is 1.0, which is equivalent to the original product operation).

[0128] The weighting coefficients can be adjusted according to the actual scenario: for example, in scenarios with severe occlusion, they can be set to... w r =1.2、 w s =0.9, prioritizing the weighting of reliable region features; in scenarios with good point cloud quality, it can be set to... w s =1.2、 w r =0.9, prioritizing features with high semantic value.

[0129] Implementation Method Two (Adapted for Bridges with Complex Geometric Structures). The linear weighted fusion expression is:

[0130] S final = w s × S i + w r × S reliability

[0131] in, w s and w r For normalized weighted coefficients, satisfying w s + w r =1, and the value range is [0.3, 0.7].

[0132] Typical value examples: w s =0.6、 w r =0.4 (balancing semantic value and data reliability); for bridges with strong symmetry such as arch bridges and suspension bridges, a setting can be made. w s =0.7、 w r =0.3, which strengthens the dominant role of semantic features.

[0133] This invention, through a flexible design of weighted fusion, can adapt to classification requirements of different point cloud qualities and different bridge structures. Its core is to adjust the weight coefficients so that the final sampling weights simultaneously take into account "semantic discriminability" and "data reliability", avoiding sampling bias caused by a single score.

[0134] This strategy ensures that the sampled points must simultaneously meet two conditions: (1) they have high semantic value for bridge type identification; and (2) they are located in a complete and reliable area of ​​the point cloud. This AND logic effectively avoids selecting points that are important but severely occluded or polluted by noise as key points.

[0135] The final importance scores are then converted into a probability distribution using the Softmax function.

[0136]

[0137] Based on the probability distribution P, K points are randomly selected using multinomial sampling.

[0138] This invention employs a probability-based random sampling strategy, rather than a simple Top-K selection. First, the final importance scores are converted into a probability distribution, and then random sampling is performed based on this distribution. This strategy introduces beneficial randomness into model training, enhancing the model's generalization ability and robustness, preventing it from overly relying on a few 'seemingly perfect' key points.

[0139] Based on the index obtained from sampling, the corresponding points and features are extracted from the original data. After sampling, the features of these points (not just coordinates) must be passed, because the subsequent local feature extraction module needs to group and aggregate them based on these features.

[0140] To optimize performance, this invention employs a differentiated sampling strategy during the training and inference phases: probabilistic sampling is used during training to enhance exploratory and regularization effects; deterministic Top-K sampling is switched during deployment and inference to ensure stable and repeatable output, meeting the requirements of engineering applications.

[0141] Use the topk function to find the indices of the K elements with the largest importance scores S.

[0142] This module fundamentally improves the traditional point cloud sampling mechanism by introducing reliability scores and probabilistic sampling strategies:

[0143] (1) From geometry-driven to semantic and reliability-driven: The fixed geometry-based FPS algorithm in the SA layer is replaced with a learnable intelligent sampling module that considers both semantic importance and data reliability.

[0144] (2) The unity of efficiency and accuracy: Semantic sampling can actively focus on discriminative regions such as towers and arch ribs, achieving higher accuracy with fewer points (smaller K value), greatly reducing computational overhead and providing the possibility for airborne real-time processing.

[0145] (3) Excellent anti-occlusion ability: Even if the bridge body is severely occluded, as long as the remaining key parts (such as half of the tower or a section of the arch rib) are scanned, their high semantic importance will be combined with the reliability assessment, so that they still have a high probability of being selected, ensuring that the model can make reliable judgments.

[0146] In some embodiments, the feature extraction module is equipped with a PointNet++ network; the bridge classification module is used to: calculate the low-dimensional features of each point in the original point cloud data; wherein the low-dimensional features include normal vector, curvature, and point density; and concatenate the preprocessed original point cloud data with the low-dimensional features to obtain the original feature matrix.

[0147] In this embodiment of the invention, the feature extraction module uses the PointNet++ network as the core feature extraction structure. This network has multi-level feature extraction capabilities and can effectively extract local and global features of point clouds, adapting to the processing needs of complex scenarios such as bridge point clouds.

[0148] The bridge classification module primarily undertakes two tasks: First, it calculates the low-dimensional features of each point in the original point cloud data. The extracted low-dimensional features specifically include three categories: normal vector, curvature, and point density. The normal vector describes the orientation of the surface where the point is located, the curvature describes the degree of curvature of the surface, and the point density describes the sparsity of the point cloud in the region where the point is located. Second, it performs feature stitching. The preprocessed original point cloud data, after cleaning and standardization, is integrated with the calculated low-dimensional features corresponding to each point. This stitching fuses the coordinate and geometric attribute information of the point cloud, ultimately forming the original feature matrix, which provides the basic input for subsequent feature enhancement modules.

[0149] In some embodiments, the bridge classification module is used to: group the selected key feature points into their Euclidean space neighborhoods and extract local features; perform hierarchical feature aggregation based on the grouping results and local features to obtain multi-scale local features; perform global max pooling on the multi-scale local features to obtain a global feature vector describing the entire bridge, and pass it into a fully connected layer and a Softmax classifier to obtain the bridge classification result.

[0150] In this embodiment of the invention, the keypoint set obtained by the above-mentioned dynamic anti-occlusion importance sampling is used as the center, and it is grouped within its Euclidean space neighborhood. The PointNet module is then used to extract local fine features. Since the sampled keypoints have both high semantic information and high reliability, the quality of the subsequently extracted local features is significantly improved, with less noise. Multiple "sampling-grouping-extraction" operations are performed through multiple Set Abstraction (SA) layers to gradually expand the receptive field and fuse multi-scale local features. In each SA layer, the above-mentioned dynamic anti-occlusion importance sampling mechanism can be applied to achieve hierarchical and intelligent feature aggregation. Finally, global max pooling is used to obtain the global feature vector describing the entire bridge, which is then fed into a fully connected layer and a Softmax classifier to output the final bridge type classification probability distribution.

[0151] In some embodiments, the preprocessing module is used to: perform denoising, ground point removal and downsampling on the original point cloud data to obtain pre-cleaned point cloud data; and perform annotation and data augmentation on the cleaned point cloud data to complete data preprocessing.

[0152] In this embodiment of the invention, the goal of data acquisition and preprocessing is to prepare high-quality, standardized bridge point cloud data for model training and inference. The input is the original LiDAR scan point cloud (in formats such as .las, pcd, bin), and the output is a cleaned, labeled, and enhanced point cloud dataset (for training) or a single-frame standardized point cloud (for inference). The specific process is as follows:

[0153] 1. Data Acquisition: Using drones equipped with LiDAR, multi-angle scans are performed on various types of bridges, including beam bridges, arch bridges, cable-stayed bridges, and suspension bridges, to ensure coverage of key structures;

[0154] 2. Data cleaning: First, noise reduction, using algorithms such as statistical filtering and radius filtering to remove outliers generated by birds, vehicles, tree branches and leaves; second, ground point removal, using terrain filtering algorithms such as cloth simulation filtering (CSF) to separate bridge point clouds from ground point clouds; third, downsampling, using voxel mesh downsampling to reduce the number of point clouds while maintaining the geometry of the bridge, thereby improving the efficiency of subsequent processing.

[0155] 3. Data labeling: Each frame of point cloud data after cleaning is manually or semi-automatically labeled and tagged with bridge type (e.g., 0: Beam Bridge, 1: Arch Bridge, 2: Cable-stayed Bridge).

[0156] 4. Data Augmentation: This includes geometric transformations (randomly rotating, translating, and scaling the point cloud to increase data diversity), point perturbation (adding small random noise to the position of points to improve model robustness), and random discarding (randomly discarding a certain proportion of points to simulate point cloud gaps that may occur in actual scanning). For example, if only the front of a cable-stayed bridge is scanned, data augmentation can generate "new" data from the side (rotation), "new" data at a larger scale (scaling), "new" data partially obscured by trees (randomly discarding points), and "new" data with slightly different lighting (adding noise). It can also actively generate point clouds with severe occlusion (simulating large vehicles passing under the bridge), allowing the model to learn to infer that it is a cable-stayed bridge even if only a part of the bridge is seen (such as a pylon).

[0157] 5. Dataset partitioning: Divide the processed data into training set, validation set and test set (e.g. 70% / 15% / 15%).

[0158] In addition, the purpose of model training and optimization is to quantify the gap between the model's predicted values ​​and the true labels. The method is to compare the predicted probabilities of the model output with the true labels of the data (one-hot encoding, such as [0,0,1,0] for cable-stayed bridges).

[0159] The choice of loss function is crucial; cross-entropy loss is commonly used. If the number of samples in each class is uneven in the dataset (e.g., there is very little data on suspension bridges), weighted cross-entropy or Focal Loss is employed to make the model focus more on the harder-to-classify samples. The standard cross-entropy loss formula for a single sample is:

[0160] Loss = - (true_label[i] * log(pred_probs[i]))

[0161] In this model, `true_label[i]` is a one-hot encoded value, `pred_probs[i]` is the model's predicted probability, and the final calculation is the average cross-entropy loss. A high loss value forces the model to make significant parameter updates during backpropagation, thus striving to avoid repeating the same mistakes in the future. The optimizer (such as Adam) uses this huge gradient to make a very large adjustment to all parameters in the model (including weights in the attention mechanism, weights in fully connected layers, etc.). When the data is imbalanced, weighted cross-entropy is used, giving higher penalty weights to the class with fewer samples. If the model misclassifies a rare suspension bridge sample, it is severely penalized, leaving a lasting impression.

[0162] Specifically, in the standard cross-entropy loss formula, a weight coefficient weight[c] is introduced for each class c. The formula is loss = -weight[c] * log(predicted probability), where weight[c] is calculated as total_samples / (num_classes * samples_per_class[c]). Here, total_samples refers to the total number of samples in the entire training dataset, i.e., the sum of all bridge type samples; num_classes refers to the total number of bridge types, i.e., the number of bridge types classified in this invention; samples_per_class[c] refers to the number of samples in the c-th bridge type, where "c" is the class index.

[0163] For example, suppose there are 4 types of bridges (beam bridges, arch bridges, cable-stayed bridges, and suspension bridges), with sample sizes of 10,000, 2,000, 1,500, and 200 respectively. Then their weights can be set as [(10,000+2,000+1,500+200) / (4*10,000), … ]≈ [0.3425, 1.7125, 2.2833, 17.125]. This means that if a beam bridge is misclassified, the penalty weight is 0.34 (a relatively light penalty because there is a lot of data), while if a suspension bridge is misclassified, the penalty weight is 17.13 (a very heavy penalty because there is a lot of data). Furthermore, Focal Loss aims to "focus on difficult examples." It not only focuses on class balance but also on the difficulty of the samples themselves. Its purpose is to reduce the contribution of "too easy to classify samples" (such as bridge samples that the model already has 99% confidence in) to the loss, allowing the model to concentrate on learning those "hard nuts to crack" (samples that are difficult to classify and may have sparse classes).

[0164] Focal Loss adds a modulation factor (1 - p_t)^γ to the standard cross-entropy loss, and its formula is as follows:

[0165] Focal Loss = - α * (1 - p_t)^γ * log(p_t)

[0166] Where p_t is the model's predicted probability for the true class; a higher p_t indicates that the sample is classified more easily and accurately. (1 - p_t) is a difficulty factor; the higher the predicted probability (easier), the closer this factor is to 0, and the lower the predicted probability (more difficult), the closer this factor is to 1. γ is a modulating factor, usually set to 2; the larger γ is, the stronger the modulation ability, and the more the loss of "easy samples" is attenuated. α is a class balancing factor, similar to the weights in weighted cross-entropy, used to give a global weight boost to rare classes. For optimizer selection, the Adam optimizer is commonly used, as its adaptive learning rate characteristic usually achieves good results. Simultaneously, it is necessary to adjust training hyperparameters such as learning rate, batch size, and training epochs, and a learning rate decay strategy can be used. During training, it is necessary to monitor the loss and accuracy on the training and validation sets in real time to prevent overfitting; if the validation set metrics no longer improve, training should be stopped.

[0167] The model evaluation phase requires a comprehensive assessment of the model's performance. First, the overall accuracy is evaluated. Then, the confusion matrix clearly shows which bridge types the model is prone to confusing (e.g., whether it easily misclassifies a suspension bridge as a cable-stayed bridge), guiding subsequent model improvements and data collection. Furthermore, precision, recall, and F1 score for each class need to be calculated. To verify the accuracy of this patent's attention mechanism, a keypoint hit rate metric is introduced: in the test set, point clouds of truly critical structural regions (such as towers, arch ribs, and anchorage areas) are manually labeled. The proportion of highly important points predicted by the model falling within these manually labeled areas is then calculated. A higher proportion indicates a more accurate attention mechanism. To highlight the superiority of this invention, comparative experiments are conducted, providing comparative data on accuracy, F1-Score, and efficiency (inference speed) on the same test set compared to baseline models (such as standard PointNet++ using FPS sampling) and other advanced point cloud classification models (such as PointCNN). Simultaneously, anti-interference tests were conducted. Different intensities of noise were actively added to the test set point cloud, some points were randomly discarded (simulating occlusion), and rotation and scaling transformations were performed. The change curves of various model indicators were observed to demonstrate that the system can cope with the complex and uncontrollable environment in actual UAV operations. Considering practical deployment requirements, the model's lightweight indicators also need to be reported, including the model's file size, computational cost for inference, and memory usage, and compared with the baseline model. Finally, a visual evaluation was provided for intuitive demonstration, including an attention point cloud map, which maps the predicted importance scores back to the original bridge point cloud, using color depth to represent importance, to visually demonstrate whether the model truly focuses on discriminative regions such as pylons, arch ribs, and suspenders; and training curves, showing the change curves of loss and accuracy on the training and validation sets with training epochs, to demonstrate the stability and convergence of model training.

[0168] In the model deployment phase, the trained model (usually a .pt or .onnx file) is first integrated into the UAV's ground station control system or onboard computer. Edge deployment better meets the requirements of real-time and offline operations. After the UAV scans a new bridge, it first uses the classifier to perform real-time or near-real-time inference to quickly identify the bridge type. Once the bridge type is identified, the system can automatically call or generate the optimal inspection path scheme preset for that type of bridge, achieving true "adaptive" path planning. To ensure system robustness and safety, a low-confidence handling mechanism is set, i.e., a confidence threshold (e.g., 0.9) is set. When the model's highest prediction probability for all categories is lower than this threshold, it is considered "uncertain." The system should not blindly execute path planning but should trigger safety mechanisms, such as hovering, automatic return, or transmitting point cloud data and raw data back to the ground station to request manual intervention.

[0169] In addition, a specific assessment of occlusion resistance and reliability is required. By plotting occlusion robustness curves, different proportions (0%-80%) of random and structural occlusion (such as occlusion of a single cable tower) are systematically simulated on the test set, and the model accuracy is plotted as a function of occlusion rate to quantitatively demonstrate the occlusion resistance capability of the invention. Simultaneously, the effectiveness of the reliability score is verified by calculating the average reliability score of the key points finally sampled by the model in the heavily occluded samples and comparing it with the average reliability score of randomly sampled points. If the former is significantly higher than the latter, it proves that the reliability score module effectively guides the sampling process and avoids unreliable areas.

[0170] In some embodiments, a bridge type rapid classification method for autonomous path planning of unmanned aerial vehicles is characterized by comprising:

[0171] Obtain the three-dimensional point cloud data of the bridge;

[0172] Clean and standardize the raw point cloud data;

[0173] Based on the preprocessed raw point cloud data and the bridge type classification model, the bridge classification results are obtained; wherein, the bridge type classification model determines the attention score of dynamic anti-occlusion importance sampling based on a semantic importance scoring mechanism based on symmetry perception enhancement.

[0174] Based on the bridge classification results, the planned inspection route for drones is determined.

[0175] In some embodiments, the bridge classification model includes a feature extraction module, a symmetry-aware feature enhancement module, a semantic importance scoring module, a dynamic anti-occlusion importance sampling module, and a bridge classification module; based on the preprocessed raw point cloud data and the bridge classification model, the bridge classification result is obtained, including:

[0176] The preprocessed raw point cloud data is input into the feature extraction module to obtain the raw feature matrix;

[0177] The original feature matrix is ​​input into the symmetry-aware feature enhancement module to obtain the enhanced semantic feature matrix;

[0178] The enhanced semantic feature matrix is ​​input into the semantic importance scoring module to obtain a semantic importance score;

[0179] The semantic importance score and the local geometric features based on the preprocessed raw point cloud computing are input into the dynamic anti-occlusion importance sampling module to obtain the filtered key feature points;

[0180] The filtered key feature points are input into the bridge type classification module to obtain the bridge classification results.

[0181] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0182] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A bridge-type rapid classification system for autonomous path planning of unmanned aerial vehicles (UAVs), characterized in that, include: The data acquisition module is used to acquire the three-dimensional spatial point cloud data of the bridge. The preprocessing module is used to clean and standardize the raw point cloud data; The bridge classification module is used to obtain bridge classification results based on the preprocessed raw point cloud data and the bridge type classification model; wherein, the bridge type classification model determines the attention score of dynamic anti-occlusion importance sampling based on a semantic importance scoring mechanism based on symmetry perception enhancement. The path planning module is used to determine the planned inspection path for the drone based on the bridge classification results.

2. The bridge-type rapid classification system for autonomous path planning of unmanned aerial vehicles according to claim 1, characterized in that, The bridge classification model includes a feature extraction module, a symmetry-aware feature enhancement module, a semantic importance scoring module, a dynamic anti-occlusion importance sampling module, and a bridge classification module; The bridge classification module is used for: The preprocessed raw point cloud data is input into the feature extraction module to obtain the raw feature matrix; The original feature matrix is ​​input into the symmetry-aware feature enhancement module to obtain the enhanced semantic feature matrix; The enhanced semantic feature matrix is ​​input into the semantic importance scoring module to obtain a semantic importance score; The semantic importance score and the local geometric features based on the preprocessed raw point cloud computing are input into the dynamic anti-occlusion importance sampling module to obtain the filtered key feature points; The filtered key feature points are input into the bridge type classification module to obtain the bridge classification results.

3. The bridge-type rapid classification system for autonomous path planning of unmanned aerial vehicles according to claim 2, characterized in that, The bridge classification module is used for: Principal component analysis is performed on the original feature matrix to calculate the covariance matrix of the point cloud segment. The first principal component vector corresponding to the largest eigenvalue in the covariance matrix is ​​extracted, and the direction pointed to by the first principal component vector is identified as the main extension direction of the bridge, which serves as the potential main axis of symmetry of the bridge. For each point feature in the original feature matrix, using the potential principal axis of symmetry as a reference, the mirror position of the point feature on the other side of the axis of symmetry is calculated by the mirror mapping algorithm. If there is a directly corresponding point cloud data at the mirror position, then the features of the points at that mirror position are extracted as symmetrical region features. If there is no direct point cloud data at the mirror location, then select a set of neighboring points within a preset range around the mirror location, and calculate the symmetrical region features of the point based on the mean of the neighborhood features of the neighboring point set. The features of each point in the original feature matrix are concatenated with the features of its corresponding symmetrical region to obtain the enhanced semantic feature matrix.

4. The bridge-type rapid classification system for autonomous path planning of unmanned aerial vehicles according to claim 3, characterized in that, The bridge classification module is used for: Map the feature vector of each point in the enhanced semantic feature matrix to an unnormalized initial attention score; LeakyReLU is used as a non-linear activation function to process the initial attention score; The initial attention scores are normalized using the Softmax function to obtain the final semantic importance score for each point. The semantic importance score of each point represents its contribution to the bridge classification task.

5. The bridge-type rapid classification system for autonomous path planning of unmanned aerial vehicles according to claim 4, characterized in that, The bridge classification module is used for: For each point in the preprocessed original point cloud, its local neighborhood is constructed, and the local geometric features of the local neighborhood are calculated; wherein, the local geometric features include local point density, normal consistency coefficient, and local curvature estimation; The calculated local geometric features are normalized and then weighted and fused into a reliability score. Finally, the importance score is fused by weighting the semantic importance score and the reliability score to obtain the attention score, which is used as the final sampling weight for each point. The final sampling weights are converted into a probability distribution using the Softmax function, and random sampling is performed based on this probability distribution to select key feature points.

6. The bridge-type rapid classification system for autonomous path planning of unmanned aerial vehicles according to claim 2, characterized in that, The feature extraction module is equipped with a PointNet++ network; the bridge classification module is used for: Calculate the low-dimensional features of each point in the original point cloud data; wherein, the low-dimensional features include normal vector, curvature, and point density; The preprocessed raw point cloud data is concatenated with the low-dimensional features to obtain the original feature matrix.

7. The bridge-type rapid classification system for autonomous path planning of unmanned aerial vehicles according to claim 2, characterized in that, The bridge classification module is used for: Using the selected key feature points as the center, group them within their Euclidean space neighborhood and extract local features; Based on the grouping results and the local features, hierarchical feature aggregation is performed to obtain multi-scale local features; The multi-scale local features are subjected to global max pooling to obtain a global feature vector describing the entire bridge, which is then fed into a fully connected layer and a Softmax classifier to obtain the bridge classification result.

8. The bridge-type rapid classification system for autonomous path planning of unmanned aerial vehicles according to claim 7, characterized in that, The preprocessing module is used for: The raw point cloud data is denoised, ground points are removed, and downsampled to obtain preliminarily cleaned point cloud data. The cleaned point cloud data is labeled and augmented to complete the data preprocessing.

9. A method for rapid bridge type classification for autonomous path planning of unmanned aerial vehicles, characterized in that, include: Obtain the three-dimensional point cloud data of the bridge; Clean and standardize the raw point cloud data; Based on the preprocessed raw point cloud data and the bridge type classification model, the bridge classification results are obtained; wherein, the bridge type classification model determines the attention score of dynamic anti-occlusion importance sampling based on a semantic importance scoring mechanism based on symmetry perception enhancement. Based on the bridge classification results, the planned inspection route for drones is determined.

10. The bridge type rapid classification method for UAV autonomous path planning according to claim 9, characterized in that, The bridge classification model includes a feature extraction module, a symmetry-aware feature enhancement module, a semantic importance scoring module, a dynamic anti-occlusion importance sampling module, and a bridge classification module; Based on the preprocessed raw point cloud data and the bridge type classification model, the bridge classification results are obtained, including: The preprocessed raw point cloud data is input into the feature extraction module to obtain the raw feature matrix; The original feature matrix is ​​input into the symmetry-aware feature enhancement module to obtain the enhanced semantic feature matrix; The enhanced semantic feature matrix is ​​input into the semantic importance scoring module to obtain a semantic importance score; The semantic importance score and the local geometric features based on the preprocessed raw point cloud computing are input into the dynamic anti-occlusion importance sampling module to obtain the filtered key feature points; The filtered key feature points are input into the bridge type classification module to obtain the bridge classification results.

Images