Multi-Vehicle Tracking Based on Monocular Camera in Driver View
Round 1
Reviewer 1 Report
The authors' proposed multi-vehicle tracking algorithm is based on a driver-facing monocular camera. It incorporates appearance and detection descriptors into a single network and adheres to the tracking-by-detection paradigm. Three prediction heads, a modified BiFPN as a neck layer, and a backbone make up the one-stage detection method. A two-step matching strategy and a Kalman filter make up the data association. The authors claim based on their results that the suggested approach outperforms cutting-edge algorithms. The work is useful because advanced driver assistance systems track obstacles with the help of multiple vehicles, which is essential for complex tasks, and dealing with changes in object illumination and deformations requires real-time performance. It is useful in that it can also resolve the tracking issue in driving scenarios while maintaining FPS. All in all, it is good and useful work and can be accepted.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
it is a good paper
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
This paper proposes a multi-vehicle tracking algorithm based on a monocular camera in driver view which is designed basically on the tracking-by-detection paradigm and integrates detection and appearance descriptors into a single network. Simulation results verify the proposed method. There are some comments as follows:
1) What are the motivations for considering the multi-vehicle tracking problem? The Kalman filter, Hungarian algorithm, and CSPDarkNet have been studied widely. It seems that the authors just combined these techniques for the multi-vehicle tracking problem, which questions the technical contributions of the paper.
2) In the proposed architecture, do the classification and regression problems use the same convolutional neural network? How many convolutional layers to achieve the best detection performance?
3) The proposed architecture is a combination of many processing techniques such as Kalman filter, Hungarian algorithm, and deep convolutional neural network with multi-scale feature fusion block, what is the complexity of the proposed architecture as compared to several baseline schemes in Table 1?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 4 Report
In this paper authors propose a novel multi-vehicle tracking algorithm based on a monocular camera in driver view. It follows the tracking-by-detection paradigm and integrates detection and appearance descriptors into a single network.
The reviewer comments and concerns:
• Some sentences are too long to make readers confused, and there are also some typos and grammar errors in this paper.
• The introduction, the rest organization of the paper should be described in the end of section.
• The related work section must be added. Related work must provide a description of more existing works from literature. This will allow you stating the contribution of this paper.
• The quality of all figures should be improved.
• Future works as an integral part should be included in the conclusions.
• In order to highlight the innovation of this work, it is better to cite more up-to-date references.
• Please improve the reference format and verify the number of each reference cited in the paper.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 5 Report
The idea is novel, and the mathematical model is well-written. The concern I currently have is related to the applicability at the application level; for instance, the experiments have been done on the RTX3090 Nvidia GPU, which is considered an HPC GPU. The computer used has a reasonably high RAM size, which is concerning for those who work in the Internet of Things (IoT) and embedded systems, as such an application suits IoT and the embedded world well. I want to see in this paper a clear message showing what needs to be done for such an algorithm to be hosted in embedded/IoT systems. Moreover, what type of accuracy will be sacrificed to reach the level of low-power applications?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report
I have no further comments.
