Vision-Based Mid-Air Object Detection and Avoidance Approach for Small Unmanned Aerial Vehicles with Deep Learning and Risk Assessment

Round 1

Reviewer 1 Report

In this paper, the target detection and obstacle avoidance of small and medium-sized fixed-wing UAVs in airspace are studied, and four processes are analyzed in depth: long-distance target detection, target area estimation, collision risk assessment and collision avoidance. The advantages and disadvantages of the existing sensors are pointed out, and the object detection based on monocular vision is further proposed, which reduces the cost, adopts reaction avoidance control, firstly extracts the corners evenly, uses the background subtraction method based on the optical flow method to detect the target area that may move, then post-processes the image, uses two morphological operations to further filter out the noise, enhances the brightness and darkness respectively, and then tracks the target in combination with the two-dimensional Kalman filter. The processed image area containing the intrusive object was input into the Mask-R-CNN to further estimate the target area size, which was a two-stage detector, the extra Mask part was analyzed, the neural network was trained, and a risk assessment method based on the two-dimensional image was proposed to evaluate the possible risk of horizontal steering direction.

Suggestions for improvement:

1. The model pre-training uses the MS COCO dataset, and it is recommended to try other datasets.

2. Only the horizontal rotation direction of the avoidance analysis and experiment, in the actual application scenario, the attitude of the fixed-wing UAV is ever-changing, and the lack of analysis of other possible collision scenarios may cause the application scenario of the method to be limited.

3. The collision risk of approaching UAV is composed of three parts: (1) area expansion coefficient, (2) position coefficient, (3) azimuth coefficient, although the article mentions the degree of influence of the three influencing factors on the results, but the corresponding experimental analysis is missing, as well as the corresponding ablation experiment, how to influence the method, it is recommended to improve the corresponding analysis and experimental comparison.

4. The topic is recommended to add fixed-wing restrictions.

5. The flight speed size range was not taken into account, only the time was taken into account.

Minor editing of English language.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

A very interesting and well presented work, verified by experimental & simulated data.

As mentioned in the conclusions, it could be improved by a more extensive training of the R-CNN and by further optimising the risk calculation weights, but this can be put in a next study, combined with real colision avoidance experiments.

Some details should be added though, in 2.1.2, regarding the Kalman filter tracking e.g., the filter type (linear, extended, adaptive, ...), the estimated state, etc. (see also [11] section III.C.2), and also in 2.2.1 regarding the R-CNN layers' parameters.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors present an original solution to the problem of collision avoidance with small unmanned aerial vehicles through long-distance object detection, object region estimation, and collision risk assessment and collision avoidance.

Comments:

1. The manuscript in its current form does not constitute a scientific article, it is an extensive report on the research conducted.

2. In the Introduction, you should formulate the thesis and the resulting 2-3 objectives of the work, which will become its chapters.

3. Equation (1), (10) and (11), provide source or justification.

4. Figure 3, provide the source and graphically adapt it to the applicable template.

5. Equation (3), give units of quantities ao and at.

6. Conclusions, is a repetition of the summary.

7. The conclusion should present a quantitative and qualitative assessment of the research conducted and show progress regarding previous work on this topic.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf