Fixed-Wing Unmanned Aerial Vehicle 3D-Model-Based Tracking for Autonomous Landing
Round 1
Reviewer 1 Report
This paper introduces a 3D model-based autonomous landing tracking system for UAVs. The system uses camera information for pose tracking to control the trajectory of the UAV. The paper also describes the design and implementation of the system and evaluates its performance. In addition, the paper mentions the differences between the system and other UAV landing systems and discusses possible future directions for improvement. Authors have performed adequate experiments to examine the method’s effectiveness and feasibility, also compared the system with alternative methods available, thereby the superiority of the solution is well proven.
The following comments should be addressed:
1. The “Introduction” seems to be too long and discussed in too much depth comparing with other papers of Drones journal. Whether it is possible to split this section into two parts?
2. It is unclear that how particle filters make a difference in the tracking process. According to the paper, at each time step, the particle set is updated by resampling the particles from the previous state to the new state. Then, the particle with the highest similarity is selected as the estimate of the current state. Compared with previous methods, what changes have been made to the system, then how and why does that works? How could the performance be further optimized in the future?
3. Now that the system is hoped to work on fixed-wings UAVs, discussing in more depth the computational volume of the system would be better, such as FLOPs, number of parameters, and FPS.
4. A great deal of space is often spent on ablation studies in papers discussing on pose estimation. This one only focuses on performance with and without pose estimation. Whether ablation studies are adequate here or not? If so, why? If not, such experiments should be added.
Author Response
Many thanks for the comments made during your review.
1. The “Introduction” seems to be too long and discussed in too much depth comparing with other papers of Drones journal. Whether it is possible to split this section into two parts?
R: The introduction section seems too long, mainly due to the figures used. I have included some context, objectives, innovation, and article organization without excessive description.
2. It is unclear that how particle filters make a difference in the tracking process. According to the paper, at each time step, the particle set is updated by resampling the particles from the previous state to the new state. Then, the particle with the highest similarity is selected as the estimate of the current state. Compared with previous methods, what changes have been made to the system, then how and why does that works? How could the performance be further optimized in the future?
R: According to what is described in line 243, only the Hight-weight particles are replicated to the current frame. Combining these particles with the pose boosting stage (Section 3.1) generates our proposal. Changes (resume) when compared with previous particle filter applications, including a pose boosting stage to use current frame information, including directional statistics, and developing a new similarity metric, among others. The main innovations are described in line 94.
3. Now that the system is hoped to work on fixed-wing UAVs, discussing in more depth the computational volume of the system would be better, such as FLOPs, number of parameters, and FPS.
R: I have not included this in my initial analysis because we use a ground-based system. We do not have power limitations and can easily access CPU and GPU with high processing capabilities. The manuscript was updated according to this comment.
4. A great deal of space is often spent on ablation studies in papers discussing on pose estimation. This one only focuses on performance with and without pose estimation. Whether ablation studies are adequate here or not? If so, why? If not, such experiments should be added.
R: Without pose estimation, we do not have tracking. If we do not have tracking, we cannot perform the landing maneuver since we use small UAVs without processing capabilities apart from a simple autopilot. I am not using any external information from the UAV. The pose estimation is based only on a single monocular camera and the UAV CAD model.
Reviewer 2 Report
The paper presents estimation filtering techniques to carry out a ground-based autonomous landing. The article is sound and results are thoroughly explained. I have few minor comments:
- There is a lack of detail on the mathematical implementation of the filters. Beside the solved-for variables, nothing is said regarding the dynamics for transaltional motion, error modeling, etc. Please insert these details.
- You say that using a ground-based system you can access more processing power: can you provide details on the navigation frequency? How often a solution is provided to the estimation?
Author Response
Many thanks for the comments made on the article review.
The paper presents estimation filtering techniques to carry out a ground-based autonomous landing. The article is sound and results are thoroughly explained. I have few minor comments:
R: Many thanks for your comment.
- There is a lack of detail on the mathematical implementation of the filters. Beside the solved-for variables, nothing is said regarding the dynamics for transaltional motion, error modeling, etc. Please insert these details.
R: Many thanks for your comment. These details are inserted in the cited references (a previous article).
- You say that using a ground-based system you can access more processing power: can you provide details on the navigation frequency? How often a solution is provided to the estimation?
R: Many thanks for your comment. The pose estimation and tracking only uses the 3D model of the UAV and a monocular camera. Since we use GPU for all these computations, we depend only on the GPU processing capability. Since we use a Ground-based system, we can easily upgrade it and have easy power available. If we use a UAV-based algorithm, we need to have an onboard processor with higher processing capabilities that are not easily available in small-size UAVs. I have updated the article according to this comment.
Reviewer 3 Report
Manuscript Fixed-wing Unmanned Aerial Vehicle 3D Model-based Tracking for Autonomous Landing describes interesting application of UAV tracking for automatic landing. It is a valuable presentation of long-going research for which authors, for the most part of results, clearly defined the improvements and novelties given in this manuscript. Pose estimation and pose tracking, or to be clear in aeronautical terms, an attitude estimation and attitude tracking for UAV, is a topic worth researching.
Title and abstract are appropriate, structure of manuscript is adequate. English language is acceptable. Paper as a whole is acceptable but there are several points, several recommendations, that would improve the quality paper.
It would be valuable to include description of elements of experimental setup, for example specifics of camera, size of the net, data about the UAV (size, mass, ...).
Angles alpha, beta and gamma from Fig.11 (and used also later in text) are not defined.
It is not clear how the data from Fig.15 were obtained.
Fig.20, 21, 22 - are not referred to in the text, what is their context and relevance to the section Experiment results.
For Figs. 21. and 22 it is not clear what they present, and what is their mutual relevance, and relevance to the Fig. 14.
In addition to the last comment that there is also a matter of Fig. 3 - it is not clear if this figure presents results from this work or from previous work (in which case it should be noted/cited).
Author Response
Many thanks for the comments made during the article review.
Title and abstract are appropriate, structure of manuscript is adequate. English language is acceptable. Paper as a whole is acceptable but there are several points, several recommendations, that would improve the quality paper.
R: Many thanks for your comment.
It would be valuable to include description of elements of experimental setup, for example specifics of camera, size of the net, data about the UAV (size, mass, ...).
R: Many thanks for your comment. I have updated the article according to your comment.
Angles alpha, beta and gamma from Fig.11 (and used also later in text) are not defined.
R: Many thanks for your comment. I have updated the article according to your comment.
It is not clear how the data from Fig.15 were obtained.
R: Many thanks for your comment. I have updated the article according to your comment.
Fig.20, 21, 22 - are not referred to in the text, what is their context and relevance to the section Experiment results.
R: Many thanks for your comment. I have updated the article according to your comment.
For Figs. 21. and 22 it is not clear what they present, and what is their mutual relevance, and relevance to the Fig. 14.
R: Many thanks for your comment. I have updated the article according to your comment.
In addition to the last comment that there is also a matter of Fig. 3 - it is not clear if this figure presents results from this work or from previous work (in which case it should be noted/cited).
R: Many thanks for your comment. Results from this work. I have updated the article according to your comment.
