Fast Planar Detection System Using a GPU-Based 3D Hough Transform for LiDAR Point Clouds

Round 1

Reviewer 1 Report

This manuscript describes a method to evaluate planar features from LIDAR measurements.

The following points summarize my observations about the manuscript and possible revisions:

1. In section 1, “GPU–GPU hybrid system” should be replaced by “CPU–GPU hybrid system”.
2. At the end of section 1, when the authors present the manuscript organization, the section “2. Related Works” is missing.
3. In section 2, the following sentence seems more suitable for section 1 (introduction): “Plane detection is a significant function to support efficient information for many applications on model reconstruction and environment perception for mobile robots and unmanned vehicles. Because of the extensive usage of plane surfaces in these applications, plane extraction algorithms from a large number of 3D point clouds become a useful point in many research domains.”
4. In the introduction, the authors say that the “two main plane detection algorithms” are the RANSAC and 3DHT approaches, while in section 2 several related works are based on PCA and region-growing techniques. The analysis of prior art presented in sections 1 and 2 is thorough, but the authors should present a clearer classification of the known methods, possibly highlighting the advantages and drawbacks of each defined class. In the present form, it is difficult to have an organic and systematic opinion of the known approaches from the reading of section 2.
5. In section 2, the authors say that “statistics-based methods have better computation efficiency than the distribution-based methods”, without a previous clear definition of the statistics- and distribution-based classes.
6. In section 2, the authors should better highlight the differences and the advantages of the proposed approach with reference to both their previous works [35] and [36].
7. In section 3.1, the authors say:” raw point clouds of a whole scene P are divided into individual m subsets Pm based on the storage order”; some more details about the definition and creation of the subsets Pm are needed.
8. In section 3.1, the terms “voxels”, “voxel clusters”, “voxel blocks” should be defined and explained for the proposed application, e.g. are voxels the 3D points yielded by the lidar without any pre-processing, except for the m subsets division? Are “voxel clusters” and “voxel blocks” defined in the cartesian space (x,y,z) or in the Hough space (r,theta, phi) ? Are they the “set S” or the “descriptor Nc” defined in section 3.3 ? Or the “voxels owing the same flag value” defined on line 293 in section 3.3?
9. Figure 1, and particularly its middle part, is not self-explicative and deserves a more detailed explanation.
10. In section 3.2, the matrix A in equation 1 has size 3x1, while the same matrix A in equation has a different size. The authors should clarify.
11. In section 3.3, figure 2 is not very informative and not very useful.
12. In section 3.3, line 269, what is the “the count of voxels at three directions”, i.e. which voxels are counted? Due to equation 7, I think that the “three directions” are along r, theta and phi in the Hough space; is it correct? If yes, it should be clearly declared when the counts I,J,K are defined.
13. In section 3.3, which are the reasons behind the initialization selection defined by equation 3? The authors should provide an explanation.
14. In section 3.3, figure 4 and its associated description (“high value voxels in flag map are extracted out, which voxel coordinates are considered as targeting planes’ parameters in polar system”) are not clear. After a connected block of voxels with the same flag value is identified, how are the “coordinates (i, j, k) in Hough space (θ, φ, r)” selected? Inside the connected block, whose voxels should have the same flag value, is the voxel with the highest Hough value selected?
15. In section 3.4, the authors say: “The resolution, representing the defined size of each voxel, of the Hough space is determined by the environmental condition during the initialisation process”; more details about the resolution selection in the Hough space would be interesting.
16. In section 4, figure 4 is not clear. The author could depict the Hough and flag values only for one or two examples of sub-spaces.
17. In section 4.2, the authors say “The sub planes, whose absolute parameter difference is less than a predefined threshold, are considered to belong to a same plane surface”; more details about the threshold selection and how it affects the final error on the evaluated planes could be interesting.
18. I know that it is an unpleasant observation, but a result comparison with a known approach would considerably increase the value of the manuscript.

Recommendation

The presented experimental results are promising and the proposed procedure is interesting. The manuscript is well-written, but it could be improved. I recommend to require major revisions according to the list of observations.

Author Response

Thank you very much for your patient reviewing. Please see the attachment of our reply letter. Before the publication, we will revise the English writting by the MDPI English editting service. 

Author Response File: Author Response.pdf

Reviewer 2 Report

Brief summary

The Authors presented a way to improve the efficiency of 3D Hugh transform algorithm in application to detection planes from 3D LIDAR point cloud. The improvement consists of two parts: dividing point cloud into parts (fraction-to-fraction method) and application of advanced hardware function of CPU and GPU. Experimental data are used to test performance of the proposed method.

 

Broad comments

Description of the methodology is generally clear but it should be extended in some places. At the beginning of the section 3.2 definition of θ, r and φ parameters should be introduced - including graphical illustration.

Assuming such a definition of the plane parameters we can meet an ambiguity for φ = 0º or φ = 180º. In such a case any value of θ refers to the same (horizontal) plane. Authors are asked for discussing this case and answering how does their algorithm deal with this case.

Conclusion section is to broad. Please concentrate on your method advantages and experiment results.

 

Specific comments and minor mistakes

 

- Line 70: "this study proposed a parallel 3D Hough transform algorithm... " present tense will be better: "this study proposes a parallel 3D Hough transform algorithm... "

 

- Line 196: Sentence "For example, in building a reconstruction ... were the focus of researchers." should be rebuilt.

 

- Line 242 and Equation (1): Symbol r (radius) is defined but never used in equations and symbol ψ used in equation (1) is not defined before.

- Equation (2): The last column of A'' matrix seem to be wrong (see the 1,j element of the result matrix)

- Line 269: Is J not necessary in equation (3)?

 

- Line 288: Where is t variable used?

 

- Line 307: "...raw point clouds and require plane parameters..." it should be "...raw point clouds and required plane parameters..." I suppose.

 

- Lines 358-359: Defining Hough space authors assumed 1º step for horizontal angle θ and 5º for vertical angle φ. What is justification for such values? Why are they different?

 

- Figure 8b: The colours for points in Hough space are selected wrongly. There is almost nothing visible in the pictures.

 

- Figures 8b and 8d: A sine waves are visible in the pictures corresponding to the steep vertical angles (first and last rectangles). Authors are asked for explanation such a view.

 

- Lines 360-361: "the resolutions of axis θ is set lower than that of axis φ" - this sentence seem to be wrong. Refer lines 358-359.

 

- Lines 416-417: The first sentence of 4.3. section is no clear. Please extend your idea description to be understandible.

 

- Figure 10: Resolution of the picture is to weak. Please try to improve the pictures quality.

 

- Table 1: What is the interpretation of the numbers (1.0 .. 10.0) describing resolution?

 

Conclusion:

Authors should extend description of methodology in some places, comment some details of the results, rebuild conclusions section and correct minor mistakes.

Author Response

Thank you very much for your patient reviewing. Please see the attachment of our reply letter. Before the publication, we will revise the English writting by the MDPI English editting service. 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The improvements introduced by the authors satisfy all my comments.

I recommend to accept the manuscript.

Reviewer 2 Report

Tank you for detailed explanations for all my comments. I recomend accepting the article for publication.