Estimation of Human Body Height Using Consumer-Level UAVs

Round 1

Reviewer 1 Report

The paper presents a methodology for assessing object dimensions from tilted UAV images, based on some assumptions, and avoiding knowledge of some data about image pose. The paper is well written and describes the method, simulations of several scenarios and actual experiments with real UAV images. The paper is very similar to a paper already published in Remote Sensing, by the same authors and does not clarify what is clearly what is new.

There are a few minor corrections to do in the description of the camera model, and a reference: 

Line 97: Is there any reason for target be written with capital letter? 

Line 97: “… Target height estimation can be estimated…” (repetition)

Line 100: “… the vertical target-to-camera distance using.” (incomplete sentence)

Line 121: “?μ and ?ν represent the focal length”. Not very rigorous language because there is only one focal length. It should be something like: “represent, respectively, the focal length and the focal length affected by the compression effect in the ν axis”.

Line 121: “μ0 and ν0 are the coordinates of the image center expressed in pixels”. Not exactly: it is the principal point, which normally is a few pixels displaced from the image center. 

Line 132: There is a missing point before matrix R.

Line 135: Letters for the world coordinate system were lower case in line 112.

Line: 166: Detorsion?

Line 206: I think the roll angle will in general be small. Otherwise, with a large roll angle, it would be needed.

Reference 11 is incomplete. It should be:

Vass, G., & Perlaki, T. (2003, June). Applying and removing lens distortion in post production. In Proceedings of the 2nd Hungarian Conference on Computer Graphics and Geometry (pp. 9-16).

My main concern is with several assumptions about the data of the UAV used, which I think are not correct. Some may have impact on the results obtained.

Authors refer the Phantom 4 manual, which states that the vertical accuracy is of 0.5 m with GPS positioning (page 62 of the manual). This was considered by the authors too large. In fact, even the 0.5 m for the vertical accuracy is too optimistic. Typically, a low cost navigation receiver, such as the one in the Phantom, will have elevation errors of a few meters, sometimes as much as 5 meters, or even more. The 0.5 m accuracy is for the “hover accuracy”, not for the determination of an absolute height. Authors should just say that a navigation grade receiver has errors of a few meters and avoid the reference to the manual.

Also with respect to information obtained from the Phantom 4 manual, it is referred that the angular uncertainty is of 0.02 degrees (line 383, line 488). Such a high accuracy is given only by top quality IMU units (e.g. https://www.oxts.com/products/xnav650/). The Phantom manual refers the “Angular control accuracy”, which respects to the accuracy with which the gimbal can be pointed, in relative terms. The actual accuracy of the pitch angle measured between the optical axis and the vertical, is much lower. From my experience with aerial triangulations of Phantom 4 Pro and Phantom 4 RTK images, I think the RMS error of the pitch angle is close to 0.3 degrees, and maximum error can reach 1 degree. It is quite likely that even these errors have small impact in the results, but the accuracy is not at all 0.02 degrees. Authors should remove the reference to the angular accuracy from the manual and assess what is the influence of larger pitch errors in the approach followed. 

Finally, there is another concern about the “distortion coefficient”. What is the distortion coefficient? Radial distortion is normally expressed by 3 coefficients of a polynomial, k1, k2, k3. Why only one value? Is it k1? In fact, for the Phantom 4 Pro the radial distortion is negligible. The camera that equips the Phantom 4 camera (both the Pro and the RTK models) has, in the original geometry, a severe radial distortion (more than 400 pixels in the corners). However, in the case of the Phantom 4 Pro, users only access images that were corrected of the radial distortion (at least to my knowledge, I never found a control software that allows to collect images in the original sensor geometry). Only a residual radial distortion (smaller than 2 pixels at 2000 pixels distance from the center) remains. So, I find very strange that some correction was needed. Please clarify what “distortion coefficient” has that value of 0.0275, and what is the size of the corrections that it is doing. If it is k1, I think it should be much smaller for a Phantom 4 Pro unit. 

Author Response

See attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper proposes a method to estimate people height using drone images.

The general idea seems interesting. However, my concern is regarding the robustness of the approach. A single measurement of the scene elements can be very noisy. I suggest using an average of the scene elements to estimate the ground length and the estimated height as an average of different drone positions. The authors should evaluate this as a third field test.

Also, Field Test 2 would be nicer to have more people in the image. This result will improve the applicability of the proposal. Also, test with different angles.

The person is always in the center. Also, it is essential to show how it will behave in other parts of the images and evaluate the caveats of the proposal. 

Author Response

See attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Summary

The paper proposes a method to estimate the height of human target. With the known horizontal distance between two points on the ground (same level), formulation is devised based on trigonometry, and intrinsic and extrinsic camera parameters. The authors also performed uncertainty analysis to evaluate the error of the estimated target height.

Comments

The authors published a paper (Ref. [7]) on the same topic two years ago.

In the previous study, they calculated the target height based on the projected horizontal distance and orientation of the camera. In the current study, they calculate the target height based on the known horizontal distance between two ground points.

The formulation of the uncertainty analysis is very similar to that in the previous study – just with fewer error terms.

The authors should compare the current method, from the theoretical point of view, with the previous one in what way it is better.

Most of the references quoted are not recent.

The novelty of the paper is moderate. The principal idea is the same as before – estimate target height based on camera parameters and trigonometry.

Mistakes

Line 87, “c)” should be “d)”.

Line 166, typo “detorsion”.

Change of symbol – should use “u” instead of “µ”.

Author Response

See attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Responses are satisfactory

Author Response

The Authors would like to thank the Reviewer 1 for the valuable revision and constructive comments. 

Reviewer 3 Report

The revised manuscript is substantially improved with additional uncertainty analysis, new field test, new references, and thorough explanations.

Some minor corrections needed:

Line 569, reference error?

Line 729, “o”.

Author Response

The Authors would like to thank the Reviewer 3 for the valuable revision and constructive comments. 

The two typos have been corrected.

Regards