A classification model for power corridors based on the improved PointNet++ network

Abstract

Aiming at the existing deep learning classification model for power corridor point cloud still need to improve the classification efficiency and the robustness of the classification model to meet the requirements of practical applications. An improved classification model based on PointNet++ is proposed. Based on the fact that the main features of the power corridor scene are power lines, poles, and vegetation, the initial data are first optimally filtered, and then the ensemble abstraction module of the classical PointNet++ is modified to better adapt to the power corridor scene. Finally, h-Swish is used as the activation function to realize the accurate classification of the features of the power corridor scene, and the training time of deep learning is also greatly reduced. The experimental results show that the improved algorithm achieves an average F1 value of 97.58%, which is 3.62 percentage points higher than the classical PointNet++. Therefore, the algorithm has great potential in point cloud classification.

Keywords:

• Deep learning
• PointNet++
• Power corridor
• Activation function

1. Introduction

UAV inspection is an efficient, accurate, and safe means of transmission line inspection, which has gradually replaced the traditional inspection methods. Using various types of equipment loaded by drones, it can realize all-around monitoring and rapid analysis of transmission lines, effectively improving inspection efficiency and safety (Wu, Ding, et al. 2022). The power industry is continuously increasing its investment in drone inspection technology and promoting the intelligent and autonomous process of drone inspection to meet the refined and efficient needs of transmission line inspection (Hou et al. 2020; Liu et al. 2020). The three-dimensional point cloud high-precision model of transmission lines is an indispensable key link in the refined and autonomous inspection of UAVs. Using equipment, such as LIDAR and high-definition cameras carried by UAVs, high-precision three-dimensional point cloud data of various parts of transmission lines can be quickly obtained to achieve refined monitoring and analysis of equipment status (Wang 2023). This technology not only improves the accuracy and coverage of UAV inspections but also effectively reduces inspection costs and risks. At present, the establishment of high-precision models of three-dimensional point clouds of transmission lines has become one of the important links of UAV inspection and is one of the key technologies for realizing autonomous inspection by UAV (Zhang et al. 2023).

At present, many scholars have conducted related research on point cloud data classification of transmission lines. A comprehensive comparison of three state-of-the-art deep learning networks: PointNet++, SparseCNN, and KPConv is presented in the literature (Li et al. 2021), which compares the performance of these three deep learning networks in terms of classification accuracy, computation time, generalization ability, and sensitivity to hyperparameter selection. Overall, we find that PointNet++, SparseCNN, and KPConv outperform random forests in classification. A comprehensive review of the development of UAV technology for monitoring and identifying faults during overhead power line inspections is presented in the literature (Foudeh et al. 2021), exploring its advantages and recognizing the potential of the said method and how it can be utilized to avoid obstacles, especially when compared to state-of-the-art mechanical methods. In the literature (Zhou et al. 2019), a JointBoost-based point cloud classification method for high-voltage transmission corridors is proposed, which can effectively reduce the number of point cloud processing, improve the classification efficiency, and have a high classification accuracy. The method includes converting point clouds into images, detecting transmission corridor candidate regions on the images, defining and computing multi-scale local feature vectors, and using JointBoost classifier to classify the point clouds to be classified into four categories: ground, vegetation, power lines, and power towers. The literature (Yu et al. 2022) uses the LightGBM algorithm to implement the transmission line point cloud classification, defines two classes of features, elevation, and eigenvalue, and classifies the point clouds into three classes: ground points, tower points, and power line points. The analysis shows that the elevation class features contribute the most and the eigenvalue class features refine the classification results. The literature (You et al. 2019) designs a fully automated airborne LiDAR power patrol 3D point cloud classification method to reduce the reliance on local features by using global statistical features. The point clouds are classified into three categories: power lines, electrical towers, and ground. The literature (Dong 2022) uses PointNet in deep learning technology to finely classify the UAV laser scanning point cloud and divide the power tower point cloud into seven parts with a total accuracy of 0.91. It provides a feasible solution for point cloud classification in power scenes. The literature (Peng et al. 2019) proposes an autonomous inspection system of UAV based on laser point cloud precise positioning, which realizes unmanned inspection operation flight with spatial collision detection and automatic wall barrier capability, improves inspection efficiency and operational reliability, and provides a new development direction for transmission line inspection exploration. The literature (Wu, Ding, et al. 2022) proposes a sliding map convolutional neural network-based transmission line point cloud classification model, which effectively considers the overall geometric relationship and topological information of point cloud data, extracts local map features by iterative subsampling, multi-scale K-nearest neighbor, and edge convolution, and achieves a higher classification accuracy than conventional methods. In the literature (Wang et al. 2020), an automatic classification method of laser point cloud of transmission lines based on elevation projection is proposed, and the three-dimensional space of power lines is reconstructed by point cloud extraction thinning and least squares method, which achieves a more efficient and accurate transmission line safety detection with feasibility. In the literature (Yang and Kang 2018), an airborne point cloud power line scene classification method based on random forest posterior probability and Markov random field model is proposed, which can effectively classify electric towers, power lines, and vegetation.

Research on the PointNet++ deep learning model, which is used as a backbone network for transporting 3D point clouds in corridors. PointNet is a pioneer in this area. However, by design, PointNet is unable to capture the local structure induced by the metric space in which the points are located, thus limiting its ability to recognize fine-grained patterns as well as its ability to generalize to complex scenes. In particular, the PointNet++ network was thus introduced in challenging 3D point cloud benchmarking (Qi et al. 2017). Point Net++ network adds a layer-by-layer abstraction structure to extract local spatial features of the target point cloud by calculating distances in metric space and increasing the context ratio to achieve better results. Li et al. (2020) better captured 3D hand gestures by introducing jump connections in Point Net and extracting features at different levels. Yax and Mingwu (2021) added top-down network branching to Point Net++ and incorporated the intermediate features of the original network into the final prediction network to obtain better classification results. To enhance the ability of acquiring local features, Wu, Cui, et al. (2022) introduced the self-attention mechanism into Point Net++, which improved the accuracy of scene semantic segmentation.

2. Point cloud classification algorithm

2.1. Fabric analog filter preprocessing

To maximize the application effect of the airborne LiDAR scanned point cloud data, the point cloud filtering preprocessing is especially critical, based on the high-precision point cloud filtering can effectively differentiate the ground points and feature points, to highlight the important information and simplify the calculation. When extracting ground points, to reduce the effect of roughness, the existing tools in CloudCompare software are used to detect and reject the roughness of the original airborne LiDAR point cloud data. To further remove the near-ground points from the coarseness-rejected airborne LiDAR point cloud data, the fabric simulation filtering algorithm is applied. The algorithm assumes that there exists a piece of fabric sticking to the surface of the object, and the shape of the fabric surface can be approximated to the ground representation of the digital surface model. If the fabric has a certain degree of hardness, the point cloud data can be divided into two parts, ground and non-ground, based on the up and down flip of the fabric. Before and after filtering are shown in Figures 1 and 2.

Figure 1. Point cloud information before processing.

Figure 2. Filtered point cloud information.

2.2. Classical PointNet++ classification algorithm

PointNet++ is a deep learning model based on point clouds, which is an improvement of the original PointNet model and focuses on local invariance and global consistency in point cloud data. PointNet is a deep learning algorithm for point cloud processing, which can achieve tasks, such as point cloud classification, segmentation, and detection. Compared with traditional point cloud processing methods, PointNet can better handle the irregularity and disorder of point clouds, and has certain rotation invariance and translation invariance. Specifically, the feature extraction part of PointNet employs a combination of MLP (multilayer perceptron) and maximum pooling. First, for each point P, PointNet obtains its local feature vector f(P) through an MLP network and then performs maximum pooling on all the local feature vectors to obtain the global feature vector of the whole point cloud. Finally, the global feature vectors are classified by two fully connected layers. During the training process, PointNet employs a T-Net-based spatial transformation network (STN) to transform the point cloud by rotation and translation, which makes the model have certain rotation and translation invariance. Meanwhile, to alleviate the overfitting problem, PointNet also adopts regularization methods, such as Dropout and Batch Normalization. Of course, PointNet also has its drawbacks. PointNet is to project points into a high-dimensional space and then maximizes pooling, so it either processes one point or all points, leading to its notion of no localization, so it is difficult to learn fine features, making it limited. So it is improved into PointNet++ deep learning model based on point cloud.

The PointNet++ classification model structure consists of two main parts: the Set Abstraction module and the Fully connected layers. The Set Abstraction module is used to extract local features from the point cloud, and it contains several SA (Set Abstraction) layers, each SA layer consists of Each SA layer consists of three steps: Sampling of points In each SA layer, the input point cloud is first sampled in dimension C, and the cloud is divided into several groups, each containing a fixed number of points. This operation can be realized by the FPS (Farthest Point Sampling) algorithm, which selects the point farthest from the initial point as the next point, and so on, to finally form a uniformly distributed set of points as the group’s center point. Feature aggregation (Grouping) aggregates the points within each group with their corresponding centroids to form a new set of point clouds after sampling. This operation can be implemented by the k-NN (k-Nearest Neighbor) algorithm, i.e. for each centroid, find its k nearest neighbors and use these points as the point cloud representation of the group. Feature Extraction is performed on the point cloud representation of each group after aggregation, and the extracted features are represented as local features of the group. This operation can be implemented by an MLP (Multi-Layer Perceptron) model, which consists of multiple fully connected layers, and takes the coordinates and feature vectors of each point in the group as input and outputs the local feature vectors of the group. With the multi-layer SA module, higher-level local features can be gradually extracted, transforming the point cloud into a more abstract representation. As shown in Figure 3.

Figure 3. Single SA layer.

Fully Connected Layers (FCLs) are used for feature extraction and classification. Fully connected layers are the most common type of neural network layers that connect all nodes of the input data to each node of the output layer, enabling an end-to-end mapping from input to output. In PointNet++, the role of the fully connected layer is to combine and compress the feature vectors extracted from the previous convolution and pooling layers to generate higher-level feature representations for classification and semantic segmentation. The fully connected layer in PointNet++ is implemented with two 1 × 1 convolutional layers, where the first convolutional layer is used to downscale the feature vectors and the second convolutional layer is used to The first convolutional layer is used to downscale the feature vector and the second convolutional layer is used to upscale the downscaled feature vector. This has the advantage of reducing the number of parameters in the fully connected layer and improving the efficiency of the model. In addition, after the fully connected layer, PointNet++ uses a maximum pooling layer to achieve spatial invariance, which further improves the robustness and generalization of the model. Overall, fully connected layers play an important role in PointNet++ as they transform point cloud data into high-level feature representations, enabling the network to better understand and process point cloud data.

2.3. Improved point cloud classification model for PointNet++ power corridors

The classical PointNet++ classification model is trained on the ModelNet40 dataset, which is designed for indoor scenes with small sizes of targets to be segmented. In the classic PointNet++ classification model, the number of random samples of point clouds in the local area is set too small, and the number of points of the tower and tree point clouds in the power corridor is much larger than the number of point clouds in the indoor scene. The density distribution of each point cloud is different, the objects are relatively large, and the local neighborhood radius of the classical model is difficult to capture more geometric structure information and local feature information. Therefore, the original PointNet++ classification model structure cannot meet the classification task of power corridors.

The improved PointNet++ model slices the scene into 1 × 1 m blocks based on the classical model and randomly samples 4096 points from the blocks as training data, using five SA modules where the radius of the local neighborhood of the SA modules is 0.1/0.5/2.5/12.5/32.5 in that order. The output dimensions of each layer of the MLP network are in that order 128/256/512/1024/2048.

In this paper, the improvement of the classical PointNet++ is mainly in three aspects: (1) increasing the depth of the model, the deeper model has stronger feature extraction ability, and it is also easy to smoothly increase the radius of the receptive field; (2) increasing the radius of the model receptive field, so that the model can efficiently extract the global features of the transmission towers; (3) changing the number of features extracted in each layer, and replacing the original PointNet++ in which the feature splicing operation (concat) (Huang et al. 2017) is replaced with residual operation (residual) (He et al. 2016). The concat connection splices two features, which causes a sharp increase in the feature dimension, while the residual computation mainly calculates the difference between the two features. This model facilitates the model to extract more features while reducing the amount of model computation but does not change the dimension of the features.

After the analysis of point cloud data, such as pole towers, the model structure of the classical PointNet++ classification model is adjusted in this paper, and the model parameters before and after the adjustment are shown in Tables 1 and 2. The improved PointNet++ network structure is shown in Figure 2.

Table 1. Before adjustment.

	Sampling rate	Local neighborhood radius	MLP output feature dimension	Characteristic number
SA1	2	0.1	64	1024
SA2	2	0.2	128	256
SA3	4	0.4	256	64
SA4	4	0.8	512	16

Table 2. After adjustment.

	Sampling rate	Local neighborhood radius	MLP output feature dimension	Characteristic number
SA1	2	0.1	128	4096
SA2	2	0.5	256	1024
SA3	4	2.5	512	256
SA4	4	12.5	1024	64
SA5	4	32.5	2048	16

From Table 2, it can be seen that changing the parameters of the classical PointNet++ model transforms the originally dense point cloud data into fewer local areas, which in turn reduces the amount of data and computation and improves the efficiency and speed of the model. At the same time, the radius of each local region is increased, which makes the model have better generalization performance. It can also reduce the information loss caused by feature discontinuity, enhance the correlation between features, and improve the stability of features. In this way, the point cloud data can be better processed and the performance of the model can be improved.

The input point cloud shown in Figure 4 has N0 points and each point has C0 dimension, which is input to four SA layers to get N1 × C1, N2 × C2, N3 × C3, N4 × C4 features, respectively, and then the obtained features are input to the PointNet network for feature aggregation to get 1 × C5 dimensional features. Generally, the number of point clouds decreases with each SA layer, but the dimensionality gradually increases, which means that the abstract feature information of the point clouds is getting more and more. Finally, the points are aggregated to a point with C5 dimensionality by PointNet and it is input to the fully connected network for classification, and the score of k classes is obtained, and the class with the highest score is the final classification result.

Figure 4. Improving the PointNet++ network structure.

The PointNet++ network, which takes into account neighbors and neighborhood relations, lacks spatial consistency constraints by not taking into account the neighboring block information during point cloud processing. To enhance the spatial consistency of the predicted points, conditional random field (CRF) (Qin et al. 2020) is used to globally optimize the predicted labels. Conditional random fields often minimize an energy function to achieve global optimality. The function contains two important parts: a unitary energy term and a binary energy term. (1) $$\text{E(x)}={\sum }_i\phi u(x_i)+{\sum }_i\phi \mathrm{p}(x_i,x_j)$$(1) where φ_u (x_i) is the unitary energy component of the point classification label, which represents the probability that point i is classified into label x_i; φ_p (x_i, x_j) is the binary energy component, which represents the joint probability distribution of the variables with correlation under the influence of multiple variables, i.e. the loss value of classifying points i and j into labels x_i and x_j. The binary energy term is a smoothing term that constrains similar attribute points to be classified into the same label. CRF global processing encourages similar attribute features to be classified into the same category, which enhances the spatial consistency of the point cloud and reduces misclassification.

2.4. Data processing based on h-Swish activation function in MLP network model

Rectified linear unit (ReLU) is one of the most commonly used activation functions in neural networks because it is simple and effective to train deep neural networks quickly. In the classical PointNet++ model, ReLU has also achieved good results as an activation function for MLP (multilayer perceptron), but there are some drawbacks:

1. Dead ReLU problem: When the input value is ≤0, the output value of ReLU is 0, which is called Dead ReLU. In this case, the derivative of the activation function of ReLU is 0, and the gradient of the backpropagation is also 0, so the neurons cannot be updated and thus cannot learn the effective features. Although Batch normalization and skip connection are used in PointNet++ to alleviate the Dead ReLU problem to some extent, it does not completely solve the problem.

2. Lack of smoothness: The ReLU function is a segmented function with discontinuity, which leads to its derivative not existing or being 0 at some points. this makes the neural network unable to use these points for optimization, thus limiting the performance of the model. In the power corridor point cloud data, the point cloud data is irregular and the amount of data varies greatly between different categories. Different densities and distributions exist, so a smoother activation function is needed to better capture the information in the point clouds.

The equation of the activation function is: (2) $$\begin{array}{c}\mathrm{\text{Re}}{\text{LU = {}}_{\text{0 x}<0}^{\text{x x}\ge 0}\\ \text{=}\mathrm{\text{max}}(0,x)\end{array}$$(2)

Therefore, to address the above drawbacks, the Swish activation function, as a nonlinear mapping of the MLP network, has better smoothing and nonlinear representation to better capture the information in the point cloud data. In PointNet++, adopting this activation function can improve the performance and generalization ability of the model.

The equation for the Swish activation function is: (3) $$\mathrm{f}(x)=x·\mathit{\text{sigmoid}}(\beta x)$$(3)

Where β is a constant or trainable parameter, the sigmoid function is shown below (4) $$\sigma =\frac{1}{1+{\mathrm{e}}^{-x}}$$(4)

The sigmoid function converges to a smooth state when it tends to positive or negative infinity. When x tends to negative infinity, y tends to 0; tends to positive infinity, y tends to 1.

As shown in Figure 5, as β increases, the function becomes more and more close to the ReLU function as it tends to infinity, and when β is 0, the function becomes a linear function f(x)=. So the Swish function can also be seen as a smooth function between the linear function and the ReLU function. It is also easy to understand that when x > 0, again there is no gradient disappearance. In the case of x < 0, the neurons do not die as in the case of ReLU. It is also an advantage that the Swish derivative is not constant compared to the ReLU derivative. Moreover, Swish can be conducted everywhere and is continuously smooth. Another feature is that Swish is not a monotonic function. But swish is not without any disadvantage, the biggest disadvantage is that it is computationally intensive, because the sigmoid function is not easy to compute. So on top of Swish, this paper uses an activation function called h-Swish.

Figure 5. Swish function when β takes different values.

The formula for the h-Swish activation function is: (5) $$h-\mathit{\text{swish}}\left[x\right]=x\frac{\mathrm{\text{Re}}\mathit{\text{LU}}6(x+3)}{6}$$(5) where ReLU indicates an unbounded function, and ReLU6 indicates that the function is given an upper bound of 6, beyond which the function value is equal to 6.

The principle of the h-Swish function is the sigmoid function we use the h-sigmoid function to construct and relu6 and then divide by 6 and then translate three units to the left, you will get a similar image to the sigmoid function. As shown in Figure 6.

Figure 6. h-Sigmoid function and sigmoid function image.

In summary, the h-Swish activation function has the advantages of better nonlinear expression, fewer dead neurons, smoother derivatives and better generalization than the ReLU function. The h-Swish activation function has better performance and computational efficiency than the Swish activation function.

3. Experimental results and analysis

3.1. Experimental data

Experimental data We used Cloud Compare software to segment and label three types of objects of power corridors, transmission lines, towers, and trees, and the obtained point cloud example data are shown in the following table, in addition, the point cloud visualization of towers, transmission lines, and trees is shown in Figure 7. To ensure the average of the dataset we fine-tune the dataset using down sampling (Table 3).

Figure 7. Point cloud view of towers, transmission lines, and trees.

Table 3. Point cloud object categories and number of objects.

Point cloud object categories	Number of point cloud instances	Number of training sets	Number of test sets
Tower	162	112	50
Tree	215	165	50
Power line	184	134	50

To improve the generalization performance of the model, data improvement is performed on the pole tower point cloud data. The data improvement methods include normalizing the point clouds and performing random rotation operations on the point clouds during the training process. This increases the diversity of the data and thus improves the adaptability of the model to new samples.

The point cloud normalization process is to improve the convergence speed of the algorithm: after normalization, the data range is within a certain interval, which can make the gradient descent algorithm converge more easily and thus accelerate the training of the model. Reduce the differences between features: Different point cloud data may have different scales and distributions, which can lead to large differences between features and affect the training effect of the model. Normalization can map the data into the same scale and reduce the differences between features, thus improving the training effect of the model.

Random rotation of point clouds increases the diversity of data: Random rotation of point clouds can rotate point clouds in 3D space arbitrarily, thus increasing the diversity of data and making the model learn the features of point cloud data more comprehensively. Improving the robustness of the model: Random rotation allows the model to better adapt to the performance of the point cloud data in different poses, thus improving the robustness of the model. The point cloud model with normalization and random rotation is shown in Figure 8.

Figure 8. Perform normalization with random rotation of the point cloud model.

3.2. Evaluation criteria

The evaluation metrics used in this paper are calculated based on a confusion matrix, and the single-category binary confusion matrix is shown in Table 4.

Table 4. Confusion matrix.

		Projected value
		True	False
Actual value	True	True example (Tp)	True negative (Tn)
	False	False example (Fp)	False negative (Fn)

In Table 4, Tp denotes that an instance is a positive sample and is predicted to be a positive sample; Fn denotes that an instance is a positive sample but is predicted to be a negative sample, underreporting; Fp denotes that an instance is a negative sample but is predicted to be a positive sample, misreporting; and Tn denotes that an instance is a negative sample and is predicted to be a negative sample. In order to evaluate the experimental results, accuracy (p_re), recall (r_ec), and F1 value (F₁) are introduced as model evaluation metrics for each evaluation metric. (6) $$\mathrm{\text{Pr}}\text{e=}\frac{\text{Tp}}{\text{Tp}+\text{Fp}}$$(6) (7) $$\text{rec=}\frac{\text{Tp}}{\text{Tp}+\text{Fn}}$$(7) (8) $$\mathrm{F}1\text{=2}\times \frac{\text{prerec}}{\text{pre}+\text{rec}}$$(8)

Accuracy denotes the proportion of actual positive samples out of all samples that were predicted to be positive; recall denotes the proportion of samples that were correctly predicted to be positive out of all positive samples; and the F1 value is the reconciled mean of accuracy and recall. The final experimental results are shown in Table 5. The comparison results are shown in Figure 9, from which it can be seen that the improved PointNet++ model proposed in this paper outperforms the original PointNet++ model in all the classifications of power corridors, with an average improvement of 3.62 percentage points in the F1 value.

Figure 9. Comparison of F1 value between this model and PointNet++ model.

Table 5. Results of semantic classification model for transmission line point cloud based on improved PointNet++.

%	Tree	Tower	Power line	Average value
Accuracy	99.82	97.90	97.64	98.45
Recall	99.58	98.21	96.56	98.12
F1 value	99.70	97.99	97.05	97.58

3.3. Experimental analysis and comparison

The experimental environment of this paper is: the CPU is i7-9700K, the GPU is NVIDIAGeForceGTX965M, and the RAM is 96GB, which is equipped with Pytorch framework. The optimizer during the training of the model uses Adam, the learning rate is set to 0.001, the L2 regularization parameter is 0.1, the learning rate decay period is 15, and the learning rate decay factor is 0.5. The total loss and accurate results of the three categories and the classification results of each category under the classical PointNet++ classification model are shown in Figures 10 and 11.

Figure 10. Training time and total loss rate and total accuracy before improvement.

Figure 11. Training time and total loss rate and total accuracy after improvement.

Experiments were conducted to compare the performance of this paper’s improved PointNet++ segmentation model with the original PointNet++ segmentation model for semantic segmentation of transmission tower point clouds. The model in this paper adjusts the model parameters and structure, and uses h-Swish as the loss function in the training phase. The original PointNet++ only changes the final number of categories, the model structure remains unchanged, uses MLP + maxpooling as the feature extraction method, and uses CE as the loss function. The experimental data and training strategy are the same. As can be seen from the figure, the proposed model in this paper outperforms the original PointNet++ model in classifying all categories of transmission tower point cloud, and the F1 value is improved by 3.62 percentage points on average. The training time of the model is also 13 min higher than the classical model.

The degree of intelligence, inspection efficiency, and coverage of hand-controlled UAV inspection is low, and its application benefits have not been brought into full play. To enhance the degree of intelligence of UAV inspection, power grid enterprises are actively exploring the research and large-scale application of UAV refined autonomous inspection technology, in which the three-dimensional point cloud high-precision model of transmission lines is an essential part of UAV refined autonomous inspection. The modification of the parameters and structure of the classical PointNet++ network, and the choice of the loss function in the training phase are the difficulties of this thesis. This paper presents an improved PointNet++ algorithm for the classification task of power corridor point cloud data. In this paper, the experimental dataset is constructed using Cloud Compare for point cloud classification and annotation of three types of objects in power corridors, including transmission lines, poles, and trees. Firstly, the Set Abstraction module of classical PointNet++ is modified to make it better adapted to the power corridor scenario and to improve the computational accuracy and computational efficiency of the model. After that, because of the irregularity of cloud data in the power corridor scene and the large difference in data volume between different categories, the activation function in the MLP network model was modified and h-Swish was adopted as the activation function to achieve accurate classification of features in the power corridor scene, and the training time in deep learning was also greatly shortened. To improve the generalization performance of the model, this paper performs data improvement on the pole tower point cloud data, including point cloud normalization and random rotation operation. The experimental results show that the algorithm in this paper improves the accuracy and training time compared with the classical PointNet++ algorithm, with F1 average accuracy improvement of 3.62 percentage points and 13 min less training time than the classical model. Overall, the algorithm in this paper shows better performance and robustness in handling point cloud classification tasks. However, there are some limitations to the study, such as a finer-grained segmentation of the transmission corridor, including buildings and other smaller electrical equipment. In addition, the generalization of the method should be explored when dealing with variable point volumes and densities. These aspects can be considered to extend the study when semantic segmentation of 3D point clouds of transmission corridors is performed.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Correction Statement

This article has been corrected with minor changes. These changes do not impact the academic content of the article.

Additional information

Funding

This work was supported by Liaoning Applied Basic Research Plan (2022JH2/101300251) and Liaoning Education Department Scientific Research Project (LJKZ1101).