Rationale for the Development of an Intelligent Digital Level Crossing Protection System Based on AI and Machine Vision: A Safety Analysis of Railway Crossings in the Republic of Kazakhstan

Response to Reviewer X Comments

1. Summary

We would like to express our sincere gratitude for your thorough review and valuable comments. Your suggestions have significantly contributed to improving the quality, clarity, and scientific rigor of the manuscript.

We have carefully considered all your remarks and revised the manuscript accordingly. The corresponding changes and responses to your comments have been incorporated into the text and are highlighted in yellow for ease of reference.

2. Questions for General Evaluation

Reviewer’s Evaluation

Response and Revisions

Does the introduction provide sufficient background and include all relevant references?

Yes/Can be improved/Must be improved/Not applicable

Detailed responses are provided below in Section 3. Point-by-point Response to Comments and Suggestions for Authors.

Is the research design appropriate?

Yes/Can be improved/Must be improved/Not applicable

Are the methods adequately described?

Yes/Can be improved/Must be improved/Not applicable

Are the results clearly presented?

Yes/Can be improved/Must be improved/Not applicable

Are the conclusions supported by the results?

Yes/Can be improved/Must be improved/Not applicable

Are all figures and tables clear and well-presented?

Yes/Can be improved/Must be improved/Not applicable

3. Point-by-point response to Comments and Suggestions for Authors

Comments 1: The authors should include quantitative performance metrics (such as detection accuracy, precision, recall, and processing latency) for the YOLOv8 + ByteTrack obstacle detection module, supported by validation on a real or simulated dataset. The authors could strengthen the Results section by presenting simulation or field-test outcomes of the KZ-DALCS-AI system, including a direct comparison of closure times and accident-risk reduction against the current relay-based ALCS

Response 1: Thank you for this valuable comment. We fully agree that the inclusion of quantitative performance metrics, such as detection accuracy, precision, recall, and processing latency, as well as validation based on real or simulated datasets, would significantly strengthen the Results section.

At the current stage of the research, the proposed KZ-DALCS-AI system represents a conceptual and architectural development, and full-scale experimental validation requires deployment on an operational railway level crossing equipped with the necessary hardware infrastructure. Such validation involves significant financial and organizational resources, including access to certified railway facilities and field equipment, which are not available within the current phase of the project.

Nevertheless, we acknowledge the importance of quantitative evaluation and have explicitly addressed this aspect in the revised manuscript.

Comprehensive experimental studies, including both field testing and simulation-based validation, are planned for the next phase of the project, with expected implementation in late 2026 or early 2027.

Comments 2: The authors should add a dedicated Related Work section to properly situate their proposed KZ-DALCS-AI system within the broader landscape of intelligent transportation and railway automation research, thereby demonstrating the novelty of their AI and machine-vision based approach to obstacle detection and dynamic signaling while acknowledging prior contributions that address similar challenges in level-crossing safety, traffic throughput optimization, and the integration of autonomous technologies. Such a section would strengthen the paper’s academic rigor by contrasting the authors’ Kazakhstan-specific, multi-sensor fusion solution (combining YOLOv8, ByteTrack, microwave sensors, and inductive loops) with existing studies on AI-driven railway safety systems and the evolving competition between rail and road infrastructure.

Response 2: Thank you for this insightful comment. We fully agree that a dedicated Related Work section is essential to properly position the proposed KZ-DALCS-AI system within the broader context of intelligent transportation systems and railway automation research.

Accordingly, we have added a new Section 2. Related Work, in which we review and analyze existing approaches related to AI-based obstacle detection, multi-sensor data fusion, intelligent level crossing protection systems, and the integration of autonomous technologies in transport infrastructure. The revised section highlights the limitations of existing solutions and emphasizes the novelty of the proposed Kazakhstan-specific approach based on multi-sensor fusion (YOLOv8, ByteTrack, microwave sensors, and inductive loops).

The added content is presented in Section 2. Related Work (lines 120–184) of the revised manuscript. In addition, relevant references [7–29] have been incorporated to support the discussion, which are listed in the reference section (lines 915–963).

These revisions strengthen the academic rigor of the manuscript and clearly demonstrate the contribution and originality of the proposed system in comparison with existing studies.

Comments 3: The authors should explain in the Discussion or a new Related Work section why their KZ-DALCS-AI system remains fully justified and essential for Kazakhstan despite directly opposing Y. Wiseman, “Autonomous vehicles will spur moving budget from railroads to roads”, International Journal of Intelligent Unmanned Systems, Vol. 12(1), pp. 19-31, 2024, who argues that autonomous vehicles will increase road capacity by up to 187% and shift investments away from railways. They should cite this paper and should acknowledge this global trend while emphasizing Kazakhstan’s distinctive realities (vast distances, extreme weather, heavy freight reliance, and over 1,000 mostly unguarded level crossings) that road automation alone cannot resolve. By demonstrating how the AI-driven multi-sensor fusion (YOLOv8, ByteTrack, microwave radars, and inductive loops) reduces annual human-factor fatalities and enables dynamic real-time signaling, the authors can show that their solution future-proofs rail’s competitive advantage in long-haul freight and passenger corridors, transforming the apparent contradiction into a compelling argument for continued intelligent rail investment and technological sovereignty.

Response 3: We appreciate the reviewer’s insightful comment and fully acknowledge the growing global trend toward autonomous vehicles and road-oriented intelligent transportation systems, as highlighted in Autonomous vehicles will spur moving budget from railroads to roads. Several studies confirm that autonomous driving technologies and cooperative traffic systems can significantly improve road capacity and efficiency.

However, the applicability of this trend is strongly dependent on regional and infrastructural conditions. The majority of these studies focus on highly urbanized regions with dense road networks and advanced digital ecosystems. In contrast, railway transport remains a backbone of freight and long-distance mobility in countries with large territories and low population density (UIC, 2021; Rodrigue, 2020).

Kazakhstan represents a distinctive case where railway infrastructure plays a critical strategic role due to:

•            vast geographical distances,

•            high dependence on bulk freight transportation,

•            extreme climatic conditions,

•            and the presence of a large number of level crossings, many of which require modernization.

Under such conditions, a purely road-centric approach based on autonomous vehicles cannot fully address safety and efficiency challenges. Instead, the modernization of railway infrastructure through intelligent systems is essential.

In particular, the implementation of AI-driven multi-sensor safety systems at level crossings enables:

•            reduction of human-factor-related accidents,

•            real-time obstacle detection,

•            and dynamic control of signaling systems.

Therefore, rather than contradicting the global trend identified by Wiseman (2024), the proposed approach complements it by addressing region-specific challenges and ensuring the sustainable development of railway transport in Kazakhstan.

The added content is presented in Section 5. Discussion / Subsection 5.1. Autonomous Transport Trends and Implications for Railways (lines 744-778) of the revised manuscript.

Comments 4: The authors could also consider the model suggested in Amin, A. L., Chimba, D., & Hasan, K. (2025). Integrating AI and edge computing for advanced safety at railroad grade crossings. Journal of Rail Transport Planning & Management, 33, 100501. that presents an integration of AI and edge computing in Railroad High-Grade Crossing.

Response 4: We thank the reviewer for highlighting the relevant study by Integrating AI and edge computing for advanced safety at railroad grade crossings, which presents a state-of-the-art approach integrating artificial intelligence and edge computing for railway grade crossing safety.

We agree that the integration of AI with edge computing is a promising direction for real-time, safety-critical railway applications. The cited study proposes an edge–cloud architecture with decentralized data processing, enabling reduced latency and efficient bandwidth utilization. It also demonstrates high-performance results using an ensemble of YOLOv8 models combined with a UNet segmentation model, achieving detection accuracy up to 97–98% and real-time processing at approximately 10 frames per second.

At the same time, the proposed KZ-DALCS-AI system shares several conceptual similarities with this approach, particularly in terms of local (edge-level) data processing and real-time object detection. However, the developed system extends this paradigm in several key aspects:

•            Multi-sensor fusion vs. vision-only approach:

While Amin et al. primarily rely on camera-based perception, the KZ-DALCS-AI system integrates heterogeneous data sources (video, microwave sensors, inductive loops, axle counters), significantly improving reliability under adverse environmental conditions (fog, snow, low visibility).

•            Three-level decision-making logic:

The proposed system introduces a formalized multi-level obstacle classification and decision framework, including behavioral analysis and probability estimation of object clearance from the danger zone, which is not addressed in the cited work.

•            Safety-critical integration (SIL4):

Unlike the primarily experimental framework of Amin et al., the KZ-DALCS-AI system is designed in accordance with CENELEC standards and ensures Safety Integrity Level 4 (SIL4), enabling direct integration with railway signaling systems.

•            Dynamic signaling and traffic optimization:

The system not only detects obstacles but also dynamically controls signaling logic, reducing unnecessary emergency stops and improving traffic throughput.

•            Adaptation to regional conditions:

The proposed architecture is specifically tailored to Kazakhstan’s operational environment, characterized by extreme climate condition, long distances, and a large number of unguarded crossings.

Thus, while the approach proposed by Amin et al. (2025) represents an important contribution to AI-based railway safety systems, the KZ-DALCS-AI system advances this direction by providing a fully integrated, safety-certified, multi-sensor solution suitable for real-world deployment in large-scale railway networks.

The added content is presented in Section 2. Related Work (lines 147-184) of the revised manuscript.

Comments 5: The authors should add a dedicated subsection in the Discussion that explicitly addresses system limitations, potential failure modes of the AI models under adverse weather or lighting conditions, and planned mitigation strategies.

Response 5: Thank you for this important comment. We fully agree that explicitly addressing system limitations and potential failure modes is essential for a comprehensive evaluation of the proposed approach.

Accordingly, we have added a dedicated subsection entitled “5.2. System Limitations and Future Work” to the Discussion section. In this subsection, we describe the main limitations of the proposed system, including the impact of adverse weather and lighting conditions on video-based detection, potential failure modes of AI models (e.g., false-positive and false-negative detections), and challenges related to deployment under varying operational conditions.

In addition, we outline the mitigation strategies implemented in the system, such as multi-sensor data fusion (combining video, microwave sensors, and inductive loops), semantic threat classification, and fail-safe mechanisms compliant with SIL4 principles.

These additions can be found in the 5.Discussion section (lines 796-818, 826-835) of the revised manuscript: 5.2. System Limitations and Future Work”

Comments 6: The authors could improve Figure 3 (system architecture) and Figure 4 (decision logic) by using higher-resolution diagrams with consistent symbols, clearer labels, and a legend explaining all components and data flows.

Response 6: Thank you for this helpful suggestion. We have made every effort to improve the clarity and visual consistency of the figures in accordance with your recommendations.

Figure 4 (decision logic) has been reviewed and updated where possible. In particular, higher-resolution versions have been provided, symbols have been standardized, labels have been clarified, and legends have been added to improve the interpretation of components and data flows.

At the same time, we note that for Figure 3, the scope for modification was limited, as the original diagram already reflects the established system architecture and follows standardized engineering conventions. Therefore, only minor refinements were introduced to preserve its technical accuracy and consistency.

Additionally, the diagrams were developed using Microsoft Visio in accordance with the requirements of the national standard GOST 2.749-84 “Unified System for Design Documentation. Conventional Graphic Symbols in Diagrams. Railway Automation, Telemechanics, and Communication Devices,” ensuring compliance with established engineering documentation practices.

Comments 7: The authors should provide a more detailed explanation of the three-level obstacle detection fusion logic, including the mathematical formulation or pseudo-code for how the KZ-ODC-AI controller combines camera, microwave, and inductive-loop data.

Response 7: Thank you for this valuable comment. We agree that a more detailed explanation of the multi-sensor fusion logic improves the clarity and technical transparency of the proposed approach.

Accordingly, we have expanded the manuscript by adding a pseudocode description of the three-level obstacle detection fusion algorithm implemented in the KZ-ODC-AI controller. The pseudocode illustrates how data from heterogeneous sources (camera-based detection, microwave sensors, and inductive loops) are combined, how object classification and tracking are performed, and how safety-related decisions are generated.

This addition can be found in Section 4. Results / 4.3. Multi-Sensor Fusion Decision Algorithm (Pseudocode Implementation) (system description) (lines 580-717) of the revised manuscript.

Comments 8: The authors would enhance credibility by verifying and clearly labeling all accident statistics (especially the 2016–2025 projections) and citing official sources with direct links or DOIs where possible.

Response 8: Thank you for this important comment. We agree that clearly verified and properly referenced statistical data are essential for ensuring the credibility of the study.

Accordingly, we have revised the manuscript to improve the presentation of accident statistics, including the data for the 2016–2025 period. The description of the corresponding figure has been expanded to provide clearer interpretation of the trends and underlying causes. In addition, the statistical data have been verified and supplemented with references to official sources where available.

These revisions can be found in the Section 3. Materials and Methods / 3.1. Justification for Development (Figure 2 and the corresponding description) (lines 249-268), where the accident statistics are now clearly labeled and supported by appropriate references.

Comments 9: The authors could expand the cybersecurity discussion (mentioned in keywords and abstract) with concrete technical measures, such as specific encryption protocols, OPC UA security features, or compliance details beyond the general CENELEC SIL4 reference.

Response 9: Thank you for this valuable comment. We agree that a more detailed discussion of cybersecurity measures strengthens the technical completeness of the manuscript.

Accordingly, we have expanded the system description section to include concrete cybersecurity mechanisms implemented in the proposed KZ-DALCS-AI system. In particular, we describe the use of the OPC UA protocol with its built-in multi-layered security model (including message signing and encryption modes), certificate-based authentication, and user access control. In addition, secure communication via HTTPS is specified for data transmission from IP cameras, and SafeEthernet is used for communication with field devices.

Furthermore, compliance with relevant industrial cybersecurity standards (e.g., IEC 62541 and IEC 62443) has been explicitly indicated, along with network-level protection measures such as VLAN-based segmentation and access control policies.

These additions can be found in Section 3. Materials and Methods / 3.2. Architecture of the System Under Development (lines 408-437) of the revised manuscript.

4. Response to Comments on the Quality of English Language

Point 1: The English is fine and does not require any improvement

Response 1: Thank you for your positive assessment. We appreciate your feedback.

5. Additional clarifications

-

Response to Reviewer X Comments

1. Summary

We would like to thank you for your insightful comments and constructive suggestions. Your feedback has been highly valuable in strengthening the manuscript and enhancing its overall quality.

All comments have been carefully addressed, and the manuscript has been revised accordingly. The corresponding changes and responses are highlighted in green in the revised version of the manuscript for clarity.

2. Questions for General Evaluation

Reviewer’s Evaluation

Response and Revisions

Does the introduction provide sufficient background and include all relevant references?

Yes/Can be improved/Must be improved/Not applicable

Detailed responses are provided below in Section 3. Point-by-point Response to Comments and Suggestions for Authors.

Is the research design appropriate?

Yes/Can be improved/Must be improved/Not applicable

Are the methods adequately described?

Yes/Can be improved/Must be improved/Not applicable

Are the results clearly presented?

Yes/Can be improved/Must be improved/Not applicable

Are the conclusions supported by the results?

Yes/Can be improved/Must be improved/Not applicable

Are all figures and tables clear and well-presented?

Yes/Can be improved/Must be improved/Not applicable

3. Point-by-point response to Comments and Suggestions for Authors

Comments 1: You state that accidents persist due to the "human factor" and stagnant mid-20th-century technologies. How does the proposed AI system specifically mitigate deliberate risk-taking by drivers (e.g., bypassing closed barriers), which you identify as a major cause of fatalities?

Response 1:

Thank you for this important comment. We agree that deliberate risk-taking behavior by drivers, such as bypassing closed barriers, represents a critical safety challenge and cannot be entirely prevented by any technical system if interpreted strictly as a human behavioral factor.

In this context, the proposed KZ-DALCS-AI system is not intended to eliminate intentional violations themselves but rather to mitigate their consequences. Specifically, the system enables real-time detection of hazardous situations within the level crossing zone and can promptly generate emergency signals, including alerts to the train driver and activation of restrictive signaling. This allows for timely initiation of emergency braking, thereby preventing collisions or significantly reducing their severity.

Therefore, within the scope of this study, risk reduction is considered not only in terms of preventing hazardous events but also in terms of minimizing their potential consequences. This approach aligns with modern safety engineering principles, where mitigating the impact of unavoidable or unpredictable human actions is a key objective.

The corresponding clarification has been added to the revised manuscript in Section 5. Discussion/Subsection 5.3. System Limitations and Future Work (lines 819-825).

Comments 2: You claim the system achieves Safety Integrity Level 4 (SIL 4). Since SIL 4 requires rigorous formal proof and specific hardware redundancy , could you provide more detail on how the AI/Machine Vision software (which is inherently probabilistic) is validated to meet the deterministic safety standards required for SIL 4?

Response 2:

Thank you for this important and insightful comment. We fully agree that achieving Safety Integrity Level 4 (SIL 4) requires rigorous adherence to deterministic safety principles, including formal validation methods and hardware redundancy.

In the revised manuscript, we have clarified that the AI/Machine Vision module is not responsible for executing safety-critical functions. Instead, it operates at the level of diagnostics and situational awareness, providing supplementary information for hazard detection and classification. Due to its probabilistic nature, the AI module is not used to directly generate safety-critical commands.

All safety-critical decisions, including the generation of “STOP” commands, are performed exclusively by a certified safety controller designed in accordance with SIL4 requirements. This controller implements deterministic logic, redundancy mechanisms (e.g., dual-channel architecture), and fail-safe principles to ensure compliance with the required safety standards.

This architectural separation ensures that the probabilistic behavior of AI-based components does not compromise system safety, as all critical actions are validated and executed within a deterministic and safety-certified control layer.

These clarifications have been added to Section 5. Discussion/Subsection 5.2. Safety Architecture and AI Role (lines 779-795) of the revised manuscript.

Comments 3: You use YOLOv8 and ByteTrack for identification and trajectory estimation. What is the measured latency from object detection to the issuance of a "Stop" command, and is this latency factored into the safety margin of the notification time?

Response 3:

Thank you for this important comment. We agree that quantitative evaluation of system latency is essential for assessing real-time performance and safety margins. At the current stage of the study, the exact end-to-end latency—from object detection to the issuance of a “STOP” command—has not yet been experimentally measured, as this requires full-scale deployment and testing under real operating conditions.

Comprehensive latency measurements, including their incorporation into the safety margin of the notification time, are planned as part of future experimental validation. These studies are scheduled for late 2026 or early 2027 and will be included in subsequent work.

Comments 4: Could you provide a comparative result showing the time saved when a train approaches at a lower-than-maximum speed using the AI-driven dynamic adjustment?

Response 4:

Thank you for this valuable suggestion. We agree that a comparative analysis of time savings achieved through AI-driven dynamic adjustment would strengthen the evaluation of the proposed system.

At the current stage, a precise quantitative assessment of time savings requires experimental validation under real operating conditions. Such measurements depend on field deployment and controlled testing scenarios, which are planned for the next phase of the project.

These experimental studies, including comparative analysis with conventional systems, are scheduled for late 2026 or early 2027.

4. Response to Comments on the Quality of English Language

Point 1: The English is fine and does not require any improvement

Response 1: Thank you for your positive assessment. We appreciate your feedback.

5. Additional clarifications

-