Next Article in Journal
Influence and Optimization of Geometric Parameters for Small-Radius Curved Tunnel Based on Response Surface Method
Next Article in Special Issue
Segmentation Agreement and AI-Based Feature Extraction of Cutaneous Infrared Images of the Obese Abdomen after Caesarean Section: Results from a Single Training Session
Previous Article in Journal
The Influence of Axial Compressive Stress and Internal Pressure on a Pipeline Network: A Review
 
 
Article
Peer-Review Record

Classification of Micromobility Vehicles in Thermal-Infrared Images Based on Combined Image and Contour Features Using Neuromorphic Processing

Appl. Sci. 2023, 13(6), 3795; https://doi.org/10.3390/app13063795
by Bastian Stahl *, Jürgen Apfelbeck and Robert Lange
Reviewer 1:
Reviewer 2: Anonymous
Reviewer 3:
Appl. Sci. 2023, 13(6), 3795; https://doi.org/10.3390/app13063795
Submission received: 20 February 2023 / Revised: 10 March 2023 / Accepted: 14 March 2023 / Published: 16 March 2023
(This article belongs to the Special Issue Recent Progress in Infrared Thermography)

Round 1

Reviewer 1 Report

Photos b) and c) of Figure 1 could be better detailed, containing arrows pointing to the points of interest in figure b). In figure c), the photo is very tight, cutting precisely the upper part, where the sensor and the equipment are mounted.

Author Response

Hello,

Thank you for your positiv feedback.

I added numbers to figure 1b) and the caption to clarify what is shown. I'm not sure what you men in regard to figure 2c). I'm aware that the picture is not ideal, my goal with this image was to show the whole setup including the mounting.

Author Response File: Author Response.pdf

Reviewer 2 Report

1.By reading the full text, the reviewer learned about the main work and contributions of the article. But still hope that the authors can summarize them in the introduction.

2.It is mentioned in the manuscript that through field experiments, the optimum height for LWIR equipment installation was determined to be 3m. The reviewers are interested in processing data from other locations and hope that these data will be presented in future editions.

3.In the manuscript, the feature vectors are actually extracted by the difference in temperature (driver vs bike vs background). If there is a compact vehicle (similar to I-road), and the driver is completely inside the vehicle, how to extract the feature vector through the temperature difference? Can the method classify such vehicles?

4.The reviewer found that the training dataset in the manuscript is huge, which is good. But does this data contain some datasets where vehicles overlap/occlude each other? This directly determines whether the proposed system can be applied in practice. Therefore, the reviewer hope that the processing results of similar datasets can be presented separately in future manuscripts.

5.Due to the inconsistent range of the horizontal axis, there may be some ambiguity in Figure 8. Hopefully some adjustments can be made for better readability

6.There are multiple tables used in the manuscript to represent the results of the test series. The parameter "median" is used to represent the performance. The reviewers noticed that the unit "%" was added to this parameter in some tables, while others did not. Hope to unify the way of writing.

7.From the results, Carg has the highest classification accuracy. Why is this? This should be discussed.

8. In the manuscript, the method of using the scaled ROI and the image moments to process data successively to obtain combined features may shows great advantages. Hopefully this can be highlighted in the Abstract and Introduction.

Author Response

Hello,

Thank you for your extensive review. I tried to address all your points in the marked changes.

Regarding your comments 1 and 8, I added a short summary to the Introduction.

Regarding your second comment, I extended the section about the positioning by a few remarks. When we continue the research, we will keep your comment in mind and address it in future publications.

Small E-Mobility vehicles such as the I-Road were not part of the scope of our research. We focused on bike like vehicles that can use bike lanes and walkways. Extracting the silhouette of the driver in a closed vehicle using thermal images is, as far as I’m aware, not possible, however recognition should be possible using the silhouette of the vehicle.

Our complete dataset includes overlapping objects, and we labeled them accordingly, however we decided to exclude these cases to reduce the complexity. I added a paragraph to describe how our dataset was modified and what data was excluded. We plan to address the problem of obstruction and overlapping in future research.

Regarding your fifth comment, the change in the horizontal axis is the goal of the mapping functions. I modified the layout of the figure und checked my explanation and hope that it is now easier to understand.

I corrected the table you mentioned in commend number 6, thank you for brining it to my attention.

While contemplating your seventh comment, I found an error in some of my visualizations. I fixed these and extended the discussion as you suggested.

Thank you again for your feedback.

Reviewer 3 Report

this article introduce a cost and energy efficient way to do classification of subtype of micromobility transports. for a paper which involves machine learning, i think it is important to show some results of the training/cross validation/testing process e.g how the test error/loss function changes as iteration increases. Another important aspect to address is how the model architecture as well as hyperparameters affect the model performance. it would be a more comprehensive article if these two parts are addressed as well.

Author Response

Thank you for your feedback.

The hardware bound nature of the NM500 made it difficult for us to apply certain kind of tests and prohibited us from adjusting many parameters. We tried using common machine learning practices, but describing/visualizing these were not on our scope for this work, as we were focused on the general application and getting the system running. Better utilization ,understanding and benchmarking of the NM500 will be the scope of future projects.

However, we performed a few tests that might go into the direction you mentioned. I included one into section  4.5.

Round 2

Reviewer 3 Report

thanks for adding the training performance behavior plot, it is helpful to gain a better understanding to the system's performance

Back to TopTop