Insulator Extraction from UAV LiDAR Point Cloud Based on Multi-Type and Multi-Scale Feature Histogram

Thank you very much for taking the time to review this manuscript and providing positive feedback on our research. We have seriously considered your suggestions and have made the revisions to the manuscript. Please find the detailed responses below and the corresponding revisions in track changes in the re-submitted files.

Comments 1: Parameter width in the figures (figure 3, figure 4, ...). Add the units, centimeters.

Response 1: Thanks for suggestion. The 'width' in Figures 3 and 4 and 'height' in Figures 7 and 9 are a relative concept, corresponding to the image, indicating the width or height of the current position from the starting position in the horizontal or vertical direction. So the units may not need to be added. If there are any further questions, please let us know.

Comments 2: Parameter wg, in line 189. Explain, although it is explained later.

Response 2: We are sorry that we made an editing error on line 157. The phrase 'with grid width wi' should be corrected to 'with grid width wg,' as revised in the new manuscript. And the wg in line 189 is same as the wg in line 157.

Comments 3: Figure 15. Add coordinates to be able to locate the areas of the experiment

Response 3: Thank you for pointing out this issue. We are glad to add coordinates in Fig. 15. However, we are very sorry that the experimental area may contain some secret areas. We have been instructed not to reveal too much detailed information. In fact, the method we propose mainly focuses on the extraction of insulator point clouds, and the experimental area has little impact on our method. We hope this response can alleviate your concerns.

Comments 4: References. We have not been able to find 28, which is also considered relevant.

Response 4: Thank you for your comments. Reference [28] is a doctoral dissertation in Chinese, and the author has published an English version of some of the research in Reference [16], which includes the portion we cited. Therefore, we replaced [28] with [16], and the Reference [28] is removed.

Thank you very much for taking the time to review this manuscript and providing insightful comments on our research. We have seriously considered your suggestions and have made the revisions to the manuscript. Please find the detailed responses below and the corresponding revisions in track changes in the re-submitted files.

Comments 1: Although 36 references are included in this work, but more than 13 references are published more than 5 years, even 10 years, which definitely could not reflect the edge cutting development of this research direction of Insulator extraction from UAV LiDAR point cloud.

Response 1: We made a recheck on the related works, and to the best of our knowledge, we have covered most related works. The most relevant reference, [18], published in 2023, reviews 6 papers utilizing point clouds for insulator extraction. Among these, 3 papers were published in 2013, 2015, and 2018, respectively and did not use UAV (Maybe there is no mature UAV Lidar product at this time). Additionally, 2 papers ([15] and [22]) using ALS point clouds are already discussed in our manuscript, and the remaining one ([17]) does not detail specific insulator extraction methods.

The reason for the lack of relevant research is that UAV technology was not very mature in the early years, especially the use of UAVs equipped with LiDAR sensors. Another reason is that compared with other larger objects in transmission corridors, such as towers and power lines, the point cloud density of insulators is smaller and more difficult to extract. Although airborne laser scanning (ALS) capture point cloud from similar view as UAV and has been widely used for more than 20 years, its flight height is much larger than UAV, and the point density is much smaller on insulator because of the occlusion of crossarm. Sometimes, even no data is captured on insulator. Thus, researches using ALS point cloud focus more on pylon and powerline extraction other than insulator extraction. Besides, early researches mainly use TLS for insulator measuring.

Furthermore, the identification of insulators using UAV image data and deep learning methods has been a research hotspot in recent years, as evidenced by publications from 2023 ([8-10]). We also cover those references. Given that UAV LiDAR point clouds offer distinct advantages over image data, our study is both relevant and significant in advancing this field.

We rechecked the references published more than 5 years ago. Some references ([5-7]) include early attempts to identify insulators using traditional image processing methods, which have since been largely replaced by deep learning methods. Some references ([27, 31, 32, 34]) indicate that we used widely recognized point cloud processing methods.

Comments 2: Many parameters are concerned in this work. But only part parameters are included in Table I with the absence of setting reference to each parameter. The table of all modules, components and parameters must be included in the modification of this work and the reference and setting support of all parameters and the datasheets must be labelled to prove the physical implement ability and repeatability.

Response 2: Thank you for your pertinent comment. There is a mistake you may have made: Table 1 is not about the parameters of the proposed method; it is about the information regarding the experimental data.

The set of parameters is definitely important in the proposed method. Most parameters in this work are used to achieve rough positioning or segmentation. Some parameters are set according to the grid width Wg.

For the most important parameter Wg, we designed an adaptive method in Section 2.2.4.

For other parameters, since the design of transmission lines usually follows a consistent standard, the settings of these parameters are relatively flexible and suitable for most transmission corridors. We think that it is unnecessary to design experiments on the impact of different parameter values on the changes in experimental results. The further expansion of these parameters is below:

(1) 10m in line 134: As we explain” since the insulator is connected to the pylon and its length will not exceed a certain limit”, we think 10 m is a suitable range and applicable to most transmission corridors. In fact, replacing 10 with any value from 8-20 has little effect on the results

(2) H0 in line 149: Since the horizontal orientation of pylon depends on the upper structure of pylon [16], we select points 3m above the tower to calculate the reorientate angle in our work. Similarly, replacing 3 with any value from 1 to 10 has little effect on the result.

(3) 0.5m in line 163: Theoretically, the bm value of a suspension tower is much smaller than that of a tension tower, so there is a large gap between the bm values of the two types. The judgment threshold of 0.5m is flexible

(4) wd in line 190: Since the location of crossarm under a grid with a width of wg, the maximum error will not exceed wg theoretically. We set wd = 2wg ensures that all points at the end of the crossarm are selected and there are enough points to fit the crossarm boundary.

(5) The other parameters such as “upper half in line 171”, “half of maximum width in 256” is consider as fix value according to the structure of towers and power lines.

The above is our detailed response to the parameter issue. If there are any other parameters that confuse you, please feel free to let us know.

Comments 3: The all concerned point cloud dataset should be added to the reference of this work for readers to identify the repeatability of this work.

Response 3: Thank you for your comment. We are glad to add original point cloud and classification results of each step of our work. However, our data collection work is authorized by enterprise of “China Southern Power Grid”, and we are not allowed to public all data due to some confidentiality agreements. In the open-source code link, we provide point clouds for each tower type used in the experiment and integrate interactive applications. Our insulator extraction experiments were all run in the code. So the repeatability of this work is easy to achieve. We hope this response can alleviate your concerns.

Comments 4: Only limited absolute performance of metric but no numerical performance gains compared the conventional schemes are included in the abstract to illustrate the superiority of the discussed Insulator extraction scheme from UAV LiDAR point cloud.

Response 4: Thanks for your suggestion. We calculated the mean F1 score of the two methods and summarized the running time. The modifications in new manuscript are as follows:

Compared with point feature-based method,the mean F1 score of the proposed method improved by 0.3, and the runtime for each tower is within 2 seconds. the proposed method has higher efficiency and accuracy.

Comments 5: The work focus on the Insulator extraction method, but the limited mathematical induction from (1) to (3) is not consistent enough. Note that the authors must provide the complete formula architecture to readers. After all, the proposed scheme should be novel and different to the existent ones.

Response 5: Thanks for your suggestion. After reviewing the corresponding content of formulas (1)-(3), we feel that the text matching the formulas has clearly explained the point cloud processing process. We guessed that there are two locations where the text description may not be enough, so we have made corresponding changes and marked the new manuscript in red:

(1) In lines 148-153: We removed the unused symbol P1 and changed the parameter H0 to the value we set 3m which is explained in response 2.

(2) In lines 148-153: We added descriptions about cross arm coordinates and wd

If something is confusing to you, we would like to get more detailed suggestions for further modifications.

Comment 6: The flow chart is included in fig 1 in the work, and the input and output signal of each module must be labeled. And the expression of these input and output signals and expression must be presented.

Response 6: Thanks for your suggestion. In the proposed method, all input and output signals are point cloud data which represent different segmentation results. They are a set containing many points in math and stored in data structures called vectors in computer. We are confused by “expression of these input and output signals expression must be presented”. If there are more detailed suggestions, we are glad to make further modifications.

Thank you again for your thoughts on improving the quality of this article

Comment 7: The UAV is included in the title of this work, the authors must provide comparative discussion and numerical illustration that what is the difference between the UAV lidar point cloud and the non-UAV lidar point cloud, at the meantime, the authors must provide the customized design in this work for the UAV point cloud.

Response 7: Thank you for your inspiring suggestion. The specific explanation is as follows:

In transmission corridor inspection using point cloud data, in addition to UAV lidar point cloud, the acquisition methods of LiDAR point cloud include: terrestrial laser scanning (TLS), mobile laser scanning (MLS), cable inspection robot (CLR) LiDAR and Airborne laser scanning (ALS) of helicopter platforms.

(1) For TLS and MLS, limited by the factors of terrain and forest environment in the transmission corridor, these two methods are difficult to apply. There are a few applications in electrical substation. As shown in the figure below, the installation method of its insulator is different from the application scenario of this article, so we did not discuss it too much. (line 87)

(2) CLR LiDAR acquires point cloud data by moving on wires or fixed to pylons (As shown in the figure below). The point cloud density is more than three orders of magnitude relative to the ALS point cloud. This makes the proposed method have certain application potential in this type of data. However, this data collection method will also collect other features in the power transmission scene. How to effectively remove these objects from massive point clouds is a new challenge, which leads to the need to reorganize the insulator extraction framework proposed in this article.

For now, such methods are limited by cost and flexibility, and existing research is still very limited. Multi-scale histogram methods based on the overall characteristics of towers and power lines have not been developed.

(3) ALS point cloud data acquired using non-UVA is generally less dense. As shown in the figure below, the suspension insulator points number usually only have a dozen or even none. The proposed method has great limitations this type of data. Existing methods mainly extract insulators through the verticality features of points and are accompanied by many parameter tests.

As shown in the figure below, the geometry of the tension tower is almost impossible to extract in non-UVA ALS point clouds as the low points density. To the best of our knowledge there is no research on the extraction of this type of insulators.

In summary, the proposed method of insulator extraction can achieve the most ideal results for UVA point clouds. Due to the gradual maturity and popularization of related technologies, the proposed method has great application prospects.

We have summarized the above analysis and added it to the Discussion section of the new manuscript:

“Data source is also an important factor for insulator extraction. Early studies mainly used terrestrial laser scanning (TLS) for insulator measuring[36,37]. But it is limited in wide application because of the complex environment along transmission corridor and the low scanning efficiency. Cable inspection robot (CIR) LiDAR is another way to capture enough insulator points[38]. However, due to its high cost, low flexibility, and more com-plex data processing, there is little research on this type of data for now. Airborne laser scanning (ALS) has similar scanning view as UAV Lidar. But as the flight height of ALS is much larger than UAV, the density of ALS point cloud is much smaller, e.g., several points/m2 (ALS) versus hundreds of points/m2 (UAV). Under some cases, there is no point captured on the insulator. Thus it is difficult to extract insulator from ALS point cloud and recognition methods based on ALS mainly focuses on the large targets, such as pylon and powerline[15,39].

In summary, the proposed method of insulator extraction can achieve the most ideal results for UVA point clouds. Due to the gradual maturity and popularization of related technologies, the proposed method has great application prospects.”

The other comment is about the customized design in this work for the UAV point cloud. In our understanding, the three-dimensional coordinates of the LiDAR point cloud can be obtained directly from the LiDAR sensor. If there is more information about 'customized design', we are glad to provide further details.

The above is our detailed response to the parameter issue. We hope this answer can relieve your confusion.